ZHANG AND ZOHREN

This Element provides a comprehensive guide to deep learning

in quantitative trading, merging foundational theory with
hands-on applications. It is organized into two parts. The first
part introduces the fundamentals of financial time-series and
supervised learning, exploring various network architectures,
from feedforward to state-of-the-art. To ensure robustness Quantitative
and mitigate overfitting on complex real-world data, a
complete workflow is presented, from initial data analysis to Finance
cross-validation techniques tailored to financial data. Building
on this, the second part applies deep learning methods to a
range of financial tasks. The authors demonstrate how deep
learning models can enhance both time-series and cross-

Deep Learning in Quantitative Trading

Deep Learning
sectional momentum trading strategies, generate predictive
signals, and be formulated as an end-to-end framework for
portfolio optimization. Applications include a mixture of data
from daily data to high-frequency microstructure data for a
variety of asset classes. Throughout, they include illustrative
code examples and provide a dedicated GitHub repository with
in Quantitative
detailed implementations.

About the Series Series Editor

Trading
Cambridge Elements in Quantitative Riccardo Rebonato
Finance aims for broad coverage of all EDHEC Business
major topics within the field. Written School
at a level appropriate for advanced
undergraduate or graduate students
and practitioners, Elements combines
Zihao Zhang and
reports on original research covering
an author’s personal area of expertise, Stefan Zohren
tutorials and masterclasses on emerging
methodologies, and reviews of the most
important literature.

Cover image: NASA, ESA, M. Robberto

(Space Telescope Science
Institute/ESA) and the
Hubble Space Telescope
Orion Treasury
Project Team ISSN 2631-8571 (online)
ISSN Core
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge 2631-8563
terms(print)
of use,
available at [Link] [Link]
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Elements in Quantitative Finance
edited by
Riccardo Rebonato
EDHEC Business School

DEEP LEARNING IN
QUANTITATIVE TRADING

Zihao Zhang
University of Oxford
Stefan Zohren
University of Oxford

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Shaftesbury Road, Cambridge CB2 8EA, United Kingdom
One Liberty Plaza, 20th Floor, New York, NY 10006, USA
477 Williamstown Road, Port Melbourne, VIC 3207, Australia
314–321, 3rd Floor, Plot 3, Splendor Forum, Jasola District Centre,
New Delhi – 110025, India
103 Penang Road, #05–06/07, Visioncrest Commercial, Singapore 238467

Cambridge University Press is part of Cambridge University Press & Assessment,

a department of the University of Cambridge.
We share the University’s mission to contribute to society through the pursuit of
education, learning and research at the highest international levels of excellence.

[Link]
Information on this title: [Link]/9781009707121
DOI: 10.1017/9781009707091
© Zihao Zhang and Stefan Zohren 2025
This publication is in copyright. Subject to statutory exception and to the provisions
of relevant collective licensing agreements, no reproduction of any part may take
place without the written permission of Cambridge University Press & Assessment.
When citing this work, please include a reference to the DOI 10.1017/9781009707091
First published 2025
A catalogue record for this publication is available from the British Library
ISBN 978-1-009-70712-1 Hardback
ISBN 978-1-009-70711-4 Paperback
ISSN 2631-8571 (online)
ISSN 2631-8563 (print)
Additional resources for this publication at [Link]/deeplearningquant
Cambridge University Press & Assessment has no responsibility for the persistence
or accuracy of URLs for external or third-party internet websites referred to in this
publication and does not guarantee that any content on such websites is, or will
remain, accurate or appropriate.
For EU product safety concerns, contact us at Calle de José Abascal, 56, 1◦ , 28003
Madrid, Spain, or email eugpsr@[Link]

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading

Elements in Quantitative Finance

DOI: 10.1017/9781009707091
First published online: September 2025

Zihao Zhang
University of Oxford
Stefan Zohren
University of Oxford

Author for correspondence: Zihao Zhang, zhangzihao@[Link]

Abstract: This Element provides a comprehensive guide to deep learning

in quantitative trading, merging foundational theory with hands-on
applications. It is organized into two parts. The first part introduces the
fundamentals of financial time-series and supervised learning, exploring
various network architectures, from feedforward to state-of-the-art. To
ensure robustness and mitigate overfitting on complex real-world data, a
complete workflow is presented, from initial data analysis to
cross-validation techniques tailored to financial data. Building on this, the
second part applies deep learning methods to a range of financial tasks.
The authors demonstrate how deep learning models can enhance both
time-series and cross-sectional momentum trading strategies, generate
predictive signals, and be formulated as an end-to-end framework for
portfolio optimization. Applications include a mixture of data from daily
data to high-frequency microstructure data for a variety of asset classes.
Throughout, they include illustrative code examples and provide a
dedicated GitHub repository with detailed implementations.

Keywords: deep learning, machine learning, reinforcement learning,

time-series, neural networks, quantitative trading, portfolio optimization,
market microstructure, momentum trading, volatility scalings
JEL classifications: A12, B34, C56, D78, E90

© Zihao Zhang and Stefan Zohren 2025

ISBNs: 9781009707121 (HB), 9781009707114 (PB), 9781009707091 (OC)
ISSNs: 2631-8571 (online), 2631-8563 (print)

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Contents

Preface 1

1 Introduction 2

PART I: FOUNDATIONS 6

2 Fundamentals of Financial Time-Series 6

3 Supervised Learning and Canonical Networks 21

4 The Model Training Workflow 65

PART II: APPLICATIONS 75

5 Enhancing Classical Quantitative Trading Strategies

with Deep Learning 75

6 Deep Learning for Risk Management

and Portfolio Optimization 101

7 Applications to Market Microstructure

and High - Frequency Data 123

8 Conclusions 143

Acronyms 147

Appendix A: Different Asset Classes 149

Appendix B: Access to Market Data 159

Appendix C: Investment Performance Metrics 162

Appendix D: Code Scripts 164

References 167

Additional resources for this publication at

[Link]/deeplearningquant
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 1

Preface
Over the past decade, deep learning has attracted considerable interest, prima-
rily due to its exceptional performance across a range of application domains,
with image recognition and natural language processing standing out as two of
the most notable examples. Deep learning algorithms possess the ability to learn
complex, nonlinear relationships from large volumes of data. Unlike traditional
mathematical or statistical models, which often struggle in such environments,
deep learning models excel at uncovering complex patterns and making pre-
dictions. The capacity to manage and learn from large volumes of data has
made deep learning models a transformative technology across industries like
healthcare, finance, entertainment, and many others.
Given its successful applications in other fields, deep learning has also
become a natural candidate for applications to quantitative trading, as trading
firms and investment managers continuously seek innovative ways to uncover
“alpha,” or excess returns. With the rise of electronic trading, exchanges now
process billions of messages daily, generating vast amounts of data well suited
for deep learning algorithms. Additionally, investors also have access to a
growing range of alternative data sources, such as mobile app downloads,
social media trends, and search engine activity (e.g., Google Trends), which can
be used to further improve decision-making. As a result, deep learning tech-
niques are increasingly becoming powerful tools for quantitative researchers
and traders, enabling more sophisticated strategies and potentially higher
returns.
A significant body of research has explored the diverse financial applica-
tions of deep learning, including areas such as alpha generation, time-series
forecasting and portfolio optimization. The goal of this Element is to weave
these disparate threads together, placing a particular emphasis on how deep
learning algorithms can be leveraged to develop quantitative trading strategies
and systems. Whether an experienced quantitative trader aiming to enhance
strategies, a data scientist exploring opportunities within the financial sector,
or a student eager to delve into cutting-edge financial technology, the reader of
this Element should come away with a comprehensive understanding of how
deep learning is transforming the landscape of quantitative trading. By com-
bining theoretical foundations with practical applications, we seek to equip
readers with the insights and tools necessary to excel in this rapidly evolv-
ing domain. Our objective is to navigate the complexities of the field while
inspiring innovation in the integration of deep learning within quantitative
finance.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 2 Quantitative Finance

To promote reproducibility and enhance readers’ understanding of the algo-

rithms discussed in this Element, we have created a dedicated GitHub reposi-
tory.1 This repository contains many of the experiments presented in the book,
and it includes everything from fundamental data processing pipelines to imple-
mentations of cutting-edge deep neural networks. By providing these resources,
we aim to empower readers to apply the concepts and techniques in practi-
cal, real-world settings. This repository is designed to be user-friendly and
accessible, and it includes step-by-step examples and demonstrations. All deep
learning models are built using PyTorch, a widely used and flexible deep learn-
ing framework. Accordingly, readers can easily experiment with and extend
these implementations. Whether readers are looking to replicate the included
experiments, refine the models, or use the provided pipelines as foundations for
their own projects, the repository offers a hands-on platform to bridge theory
and practice. Our commitment to transparency and accessibility ensures that
readers can not only learn but also actively engage with and contribute to the
evolving field of quantitative finance powered by deep learning.

1 Introduction
Quantitative trading boasts a rich and fascinating history, with its origins dat-
ing back to the groundbreaking work of Louis Bachelier in 1900. In his seminal
thesis, Bachelier introduced the concept of Brownian motion as a framework
for modeling the stochastic behavior of financial price series. This pioneering
work established the basis for the mathematical modeling of financial mar-
kets and set the stage for modern quantitative finance (Bachelier, 1900). Over
the years, the field has undergone remarkable evolution, propelled by progress
in mathematics, statistics, and computational advancements. From the intro-
duction of fundamental theories like the Black-Scholes model in the 1970s to
the emergence of algorithmic trading in the late twentieth century, quantitative
trading has consistently been at the forefront of financial innovation. Key devel-
opments have been documented in works such as Cesa (2017), which offers a
detailed exploration of quantitative finance’s historical trajectory and major
milestones.
As computational power and data availability have both increased, the
field has expanded further, incorporating machine learning and deep learning
techniques into its toolkit. Today, quantitative trading represents a dynamic
intersection of finance, mathematics, and computer science, continuing to
evolve as new methods and technologies emerge. Experts from diverse fields

1 See [Link] or [Link]

Trading.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 3

have collaborated with a common goal: to optimize financial returns while

minimizing the inherent risks of trading. This shared ambition has fueled
the evolution of quantitative trading strategies, which harness the power of
mathematical and computational models to analyze and interpret financial data.
Traditionally, statistical time-series models have served as the cornerstone
of predictive signal generation in quantitative trading. These models, such as
ARIMA and GARCH, have proven effective in capturing trends and volatil-
ity in financial time-series data. However, such models are often constrained
by their linear nature and the stringent assumptions, such as stationarity and
normality, that they impose upon the data. Given the inherently complex and
nonlinear behavior of financial markets, these limitations can lead to subopti-
mal performance, particularly in dynamic and unpredictable market conditions.
To address these challenges, practitioners have historically relied upon man-
ually crafted features to enhance the predictive power of their models. By
engineering features that capture specific market dynamics, such as momen-
tum, mean reversion, and volatility clusters, researchers aim to approximate
the underlying complexity of financial systems. However, this process is labor-
intensive, requiring significant domain expertise and time. Moreover, manual
feature engineering is susceptible to human bias, potentially introducing or
overlooking critical patterns or relationships in the data.
The increasing demand for more robust and scalable solutions has under-
scored the need for advanced methodologies capable of identifying and lever-
aging nonlinear relationships within financial data. Deep learning, a specialized
branch of machine learning, utilizes multi-layered neural networks to auton-
omously learn and uncover meaningful patterns within large and complex
datasets. The core advantage of deep learning is its capacity to learn hierar-
chical representations of data. By progressively extracting features from raw
inputs, deep learning models are capable of capturing complex relationships
and subtle patterns that traditional statistical methods often fail to detect. These
capabilities make them especially well suited for addressing the complexities
of financial markets, which are characterized by high volatility, intricate inter-
dependencies, and noisy data. Specifically, deep learning offers several distinct
advantages: It can handle both structured and unstructured data, such as news
articles and social media sentiment; it can adapt to changing market conditions
and regimes; it can uncover complex patterns as more complex data increas-
ingly requires more complex modeling techniques; it can be used for a range of
strategy types, from high-frequency execution problems to long-term portfolios
optimization.
This Element delves deeply into the transformative role of deep learn-
ing in modern quantitative trading, offering a thorough examination of how

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 4 Quantitative Finance

this advanced technology is transforming the landscape of financial markets.

Through this exploration, we aim to showcase how deep learning models excel
at automating complex feature extraction processes and uncovering patterns
within vast volumes of financial data. Through their ability to do so, deep
learning models drive more informed, precise, and effective trading strate-
gies. Our objective is to guide readers, whether researchers, data scientists,
or traders, through the practical applications and theoretical underpinnings of
deep learning in quantitative trading. This Element seeks to demonstrate how
the unmatched computational power and adaptability of deep learning can be
leveraged to develop applications for real-world, high-stakes financial trad-
ing environments. Readers will obtain a meaningful understanding of how
these models can be applied to automate decision-making, enhance predic-
tive accuracy, and optimize trading performance in the ever-evolving financial
markets.
The Element is split into two parts: Foundations and Applications. In the
first part, we cover the fundamentals of financial time-series including statis-
tics and hypothesis testing. Financial data, like any other type of data, has its
own characteristics. Accordingly, a good understanding of a financial dataset’s
underlying statistics is the basis for any financial analysis. We then introduce
the concept of supervised learning and deep learning models. These concepts
range from basic fully connected layers to the attention mechanism and trans-
former architectures, which excel at capturing long-range dependencies in
structured datasets. Despite the significant advancements in deep learning, deep
networks frequently encounter challenges like overfitting, when models excel
on training data but struggle to generalize to new, unseen data. To address this,
we present a complete workflow for developing deep learning algorithms for
quantitative trading. This process includes essential steps like data collection,
exploratory data analysis (evaluating characteristics of the data, such as distri-
bution and stationarity), and cross-validation techniques tailored specifically
for financial data. These steps are critical for building models that are robust
and reliable.
In the second part of this Element, we focus on applying deep learning
algorithms to various financial problems. One of the most fundamental tasks
in quantitative trading is generating predictive signals. We explore various
deep learning architectures for this purpose, showcasing how these networks
can be leveraged to predict market movements. Building on this foundation,
we delve into more advanced frameworks where deep networks are adopted
to enhance time-series momentum and cross-sectional momentum trading
strategies. Further, we discuss portfolio optimization and present methods to
optimize portfolio weights from market inputs that form an end-to-end frame-
work. This bypasses the intermediate requirements of estimating returns and
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 5

constructing a covariance matrix of returns, processes that are often difficult to

implement in practice.
Alongside our exploration of deep learning techniques, this Element dis-
cusses the nature and intricacies of financial data itself. To provide a detailed
perspective, we introduce the operational mechanisms of modern securities
exchanges, illustrating how financial transactions occur and the ways in which
high-frequency microstructure data, such as order book updates and trade
executions, are generated. Additionally, we analyze the unique characteris-
tics of several main asset classes, including equities, bonds, commodities, and
cryptocurrencies, shedding light on the distinct challenges and opportunities
they each present for deep learning applications. Throughout this Element,
we include code scripts to highlight important concepts, and we provide a
dedicated GitHub repository2 to further demonstrate these ideas.

An Outline of the Element This Element contains two parts: Foundations

and Applications. The Foundations part contains Sections 2, 3 and 4, in
which we introduce the fundamentals of financial time-series and deep learn-
ing algorithms. The Applications contains Sections 5, 6, and 7, in which
we discuss prediction, portfolio optimization, trade execution and real-world
applications.

• Section 2 discusses the statistics frequently used in the analysis of financial

time-series, including returns, data distributions, hypothesis testing, sta-
tistical moments, serial covariance, correlation, and statistical time-series
models such as AR and ARMA. This section also introduces the notions of
“alpha” and “beta” and examines the phenomenon of volatility clustering.
• Section 3 introduces supervised learning and its primary components,
including loss functions and evaluation metrics. We then introduce neural
networks, starting with the canonical fully connected layers, convolutional
and recurrent layers. Finally, we explore some state-of-the-art networks,
including WaveNet, encoder-decoders, and transformers.
• Section 4 presents a complete training workflow from the very first step of
data collection through the final model deployment. We discuss the problem
of overfitting and introduce cross-validation for hyperparameter tuning. We
also include a discussion of various popular model pipelines so that readers
can choose the most appropriate platform for their respective applications.
• Section 5 introduces classical quantitative strategies such as time-series
momentum and cross-sectional momentum strategies, and shows how they

2 See [Link] or [Link]

Trading.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 6 Quantitative Finance

can be enhanced with deep learning methods. In particular, we explore net-

works that directly output trade positions and are end-to-end optimized for
Sharpe ratio or other performance metrics.
• Section 6 focuses on risk management and portfolio optimization. We dem-
onstrate how deep learning models can help better forecast risk measures
such as volatility. We also look into end-to-end deep learning frame-
works for portfolio optimization, bypassing the need to estimate returns
or construct a covariance matrix for classical mean-variance problems.
• Section 7 introduces high-frequency microstructure data. We demonstrate
how bespoke hybrid-networks can serve to forecast future price trends and
exploit additional structure in limit order books. Additionally, we discuss
various promising applications including the adoption of reinforcement
learning for trade execution and generative modeling for financial data.
• Section 8 brings together the insights and knowledge presented through-
out this Element, summarizing the key takeaways from our exploration of
deep learning and quantitative trading. Looking ahead, we discuss emerg-
ing trends and explore future possibilities where deep learning might bring
innovative transformations to financial markets.

PART I: FOUNDATIONS

2 Fundamentals of Financial Time-Series

Financial time-series analysis is an indispensable tool in understanding the
ever-changing nature of financial markets. It involves the study of certain data
points collected or recorded at specific time intervals, such as daily stock prices.
This analysis is crucial for identifying trends, modeling market behaviors, and
making informed decisions in trading, risk management, and investment. This
section explores the fundamental concepts of statistics used in such analyses,
including returns, distributions, moments, hypothesis testing, serial covariance,
various time-series models, and more. These concepts form the basis of finan-
cial time-series modeling and provide the foundation to move to more complex
models later in the Element.

2.1 Returns
Returns are a key metric in the field of finance, playing an important role
in evaluating investment performance over time. They reflect the profit or
loss achieved relative to the initial value of an investment, demonstrating
insights into the potential profitability and risks associated with different traded
financial assets, including stocks, bonds, mutual funds, and other instruments.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 7

By calculating and analyzing returns, investors, analysts, and portfolio man-

agers can assess the effectiveness of their strategies, compare diverse invest-
ment options, and make data-driven decisions to enhance trading performance.
There are several different ways to calculate financial returns, with simple
returns and logarithmic (log) returns being the most common. Simple returns
calculate the percentage change in an asset’s price between two consecutive
periods. They are straightforward to calculate and understand, which makes
them a widely used tool for routine financial analysis. In addition, simple
returns are often used to calculate portfolio returns (which will be defined
in later sections). However, simple returns have limitations, particularly when
dealing with long-term investments or compounding returns.
Logarithmic returns, on the other hand, calculate the natural logarithm of
the ratio of consecutive prices. This method provides a time-additive measure,
meaning that the returns over multiple periods can be summed up to obtain the
total return, which is particularly useful for continuous compounding contexts.
Logarithmic returns are often more statistically desirable due to their properties,
such as normality and symmetry, which make them more suitable for sophisti-
cated financial models and risk assessments. The motivation for using those two
forms becomes apparent when calculating the cumulative return of a security.
For a single time step, we can define both returns as follows:
pt − pt−1
Simple return: rsim,t = ,
pt−1
  (1)
pt
Log return: rlog,t = log ,
pt−1

where pt denotes the price of a security at time t. The aforementioned can easily
be generalized for returns over multiple time steps from t − L to t.
Understanding and analyzing financial returns is crucial for several reasons.
First, returns directly impact an investor’s wealth and financial planning, as
they determine the growth of investments over time. Second, returns are used
when assessing investment risks, and effective risk control is the key to ensur-
ing long-term investment success. Third, analyzing historical returns helps
investors identify trends and patterns, informing future investment decisions
and strategy development. Finally, financial institutions and fund managers
rely heavily upon return analysis to manage large portfolios and ensure they
meet their performance benchmarks. By examining returns, they can allocate
assets more effectively, diversify their portfolios, and carry out risk manage-
ment strategies that can protect profits against adverse market movements.
In summary, financial returns are a cornerstone of investment analysis and
decision-making. They provide a complete view of the performance and risk

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 8 Quantitative Finance

of financial assets, guiding investors and financial professionals in their pursuit

of optimal investment strategies and wealth maximization. Understanding the
different methods of calculating returns is important for anyone involved in the
financial world. In many of the data-driven examples that we cover in this Ele-
ment, a future return over a specific horizon serves as the target of a predictive
supervised learning model. It reflects the direction and extent of the expected
future price movement and plays a major role in portfolio optimization, which
will be discussed more in later sections.

2.2 Distributions of Financial Returns

Loosely speaking, a distribution describes the way in which values of a ran-
dom variable are spread or dispersed. Distributions are the foundation for the
domains of probability and statistics, and distributions can be either discrete,
where data points can take on values from a finite or a countable set, or con-
tinuous, where data points can take on any value within a given range.
Understanding distributions is useful for making inferences about populations
based on samples, assessing probabilities, and conducting various statistical
tests.
Mathematically, we represent the distribution of a discrete variable by a prob-
ability mass function (PMF) and that of a continuous variable by a probability
density function (PDF). The PMF simply indicates the probabilities of differ-
ent finite or countable outcomes, and the PDF presents how the probability
of a random variable taking values in a specific range is distributed. The key
properties of PMFs or PDFs are that they are nonnegative and sum to one over
the entire space of possible values. Taking a continuous variable X with the
PDF f(x) as an example, the probability that X lies within the interval [a, b] is
determined by the integral of f (x) over that range:
∫ b
P(a ≤ X ≤ b) = f (x)dx. (2)
a

It is important to understand the concept of the distribution of financial

returns as it gauges the quality and risk of investment performance. In practice,
financial returns do not typically follow a normal distribution, which would
suggest they tend not to be well behaved. Instead, they exhibit characteristics,
such as heavy tails which indicate that extreme values (large gains or losses)
are more frequent than would be predicted by a normal distribution.
There are several ways to understand a distribution. The most straightfor-
ward way is to use histograms to visually inspect the data distribution. For
example, the left plot of Figure 1 depicts the histogram for the simple daily
returns of Standard & Poor’s 500 (S&P500) since its creation. The distribution

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 9

Figure 1 Left: histogram for the return distribution; Right: QQ-plot.

is bell-shaped and appears similar to a normal (Gaussian) distribution. How-

ever, upon inspection, this distribution exhibits fatter tails and a sharper peak
compared to a normal distribution with the same mean and standard deviation
(plotted in black).3
QQ-plot (quantile-to-quantile plot) is another popular tool to check if a data
distribution is normal. A QQ-plot is made by plotting one set of quantiles
against another set of quantiles. A quantile denotes an input value, in this case
a return, such that a certain fraction (say 90%) of the data are less than or equal
to this value (in which case we call it the 90% quantile). A straight line would
be expected if two sets of quantiles are from the same distribution. The right
of Figure 1 is the QQ-plot for the return distribution of the S&P 500 versus
the assumed normal distribution. We can observe that the points align along a
straight line in the central portion of the figure, but curve off at the two ends.
In general, a QQ-plot like this indicates that extreme values are more likely to
occur than the assumed normal distribution. Code to create a histogram and a
QQ-plot in Python is shown next:

1 # ret is the return series of interest

2 import [Link] as stats
3 import [Link] as plt
4 import numpy as np
5

6 [Link](figsize =(12 ,4))

7 [Link] (121)
8 s = [Link](ret , bins =100, density=True)
9 xmin , xmax = [Link] ()
10 x = [Link](xmin , xmax , 100)
11 p = [Link](x, [Link](ret), [Link](ret))
12 [Link](x, p, "k", linewidth =2)
13

3 We formally define mean and standard deviation in the next section.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 10 Quantitative Finance

14 [Link] (122)
15 a = [Link](ret , dist="norm", plot=plt)
16 [Link]("Theoretical Quantiles")
17 [Link]("Ordered Values")

Empirical studies of financial markets have shown that the distribution of

returns is often better fit by distributions such as the Student’s t-distribution,
which better accounts for the heavy tails, or the Generalized Error Distribution
(GED). These distributions offer a more accurate depiction of the probability of
extreme events, and consequently better model financial time-series. We need
to be aware of such properties because a higher likelihood of extreme events
means that financial models have a higher potential for significant losses.
Overall, a good understanding of return distributions helps with financial
modeling, risk assessment, and strategic decision-making. By recognizing that
returns are not normally distributed and accounting for the actual distributional
characteristics, investors and analysts can develop more robust models that
better capture the risks and potential rewards of their investments. This knowl-
edge allows for more effective portfolio diversification, hedging strategies, and
overall risk management practices, ultimately leading to more informed and
potentially profitable investment decisions.

2.3 Statistical Moments

Statistical moments are sets of parameters used to describe a distribution. In
general, we can define the k-th central moment of a random variable X as:

µk = E[(X − µ)k ], µ = E(X ), (3)

for k ≥ 2 and µ1 = µ. Typically, attention is given to the first four moments –

mean (or expected value), variance, skewness, and kurtosis – as they capture
a distribution’s central tendency, spread, asymmetry, and peakedness, respec-
tively. Statistical moments help us understand the behavior of a distribution and
make predictions. For example, a normal distribution can be specified by giv-
ing a mean and standard deviation. However, these two moments alone might
not be enough to fully describe a return distribution. As an example, the return
distribution in Figure 1 exhibits heavier tails and a more pronounced peak than
a normal distribution. In this case, we need to check higher moments of the
distribution to better understand the data.
In statistics, skewness and kurtosis are the normalized third and fourth central
moments. Skewness measures the asymmetry of data about its mean. There are
two types of skewness: positive skew and negative skew. A symmetrical distri-
bution, such as a normal distribution, has no skewness. However, a distribution
that has larger values on the right tail is positively skewed. On the contrary, a

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 11

negatively skewed distribution has larger values in the left tail that are further
from the mean than those of the left tail.
In finance, skewness may stem from diverse market forces. Investor sen-
timent can lead to asymmetrical buying or selling pressures as market par-
ticipants overreact to news or trends. Economic news can also introduce
sudden, unidirectional shocks to asset prices as markets rapidly adjust to new
information. Market microstructure might also contribute to skewness when
imbalances in order flow, liquidity constraints, or trading mechanisms create
price distortions. There are many other possible causes for deviations from a
normal distribution in the returns.
The skewness of a return distribution can typically inform the reward profile
of a security or strategy. A canonical example of a strategy that is negatively
skewed is a reversion strategy. We can expect many small positive rewards
when assets revert as expected, but we can also suffer large losses if reversion
does not occur say due to an unexpected news event. Selling options and VIX
futures are other examples of strategies with negatively skewed return distribu-
tions. Vice versa, a positively skewed return distribution typically corresponds
to many small losses with a few large gains – a canonical example being
momentum strategies. The most favorable type of skewness depends upon the
risk preferences of investors.
Kurtosis is the fourth normalized statistical moment, describing the tail and
peak of a distribution. In particular, kurtosis informs us whether a distribution
includes more extreme values than a normal distribution. All normal distribu-
tions, regardless of mean and variance, have a kurtosis of 3. If a distribution is
highly peaked and has fat tails, its kurtosis is greater than 3, and, vice versa, a
flatter distribution has a kurtosis lesser than 3. Excess kurtosis can be attributed
to market shocks, economic crises, and other rare but impactful events that
significantly affect asset prices.

2.4 Statistical Hypothesis Testing

Hypothesis testing is another important concept in statistics, as it offers a sys-
tematic framework for making decisions and drawing conclusions about a
population using sample data. It is widely adopted in several fields, includ-
ing natural science, economics, psychology, and finance, where it is used to
evaluate hypotheses and determine the validity of claims or theories. For exam-
ple, we have already mentioned the fact that financial returns can have fat
tails compared to a normal distribution. To objectively assess this, we resort
to hypothesis testing which provides a formal framework for making infer-
ences and offers a structured method for evaluating claims. Consequently, this

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 12 Quantitative Finance

allows researchers and analysts to draw conclusions with a quantifiable level of

confidence. By using statistical techniques and predefined criteria, it also elim-
inates subjective biases and ensures that decisions are conditioned on empirical
evidence rather than intuition or guesswork.
The fundamental concept of hypothesis testing involves using sample data
to evaluate the evidence against a null hypothesis (H0 ), which functions as
the default or baseline assumption. The objective is to prove the alternative
hypothesis (H1 ), which constitutes the presence of a difference from the default
assumption. The process of hypothesis testing involves several key steps. First,
we need to define the null hypothesis and the alternative hypothesis. For
instance, in a test to determine whether the distribution of returns is normal, the
null hypothesis might state that the return distribution follows a normal distri-
bution, while the alternative hypothesis might posit that the return distribution
violates the assumption of normality.
We then need to select a significance level (α), typically 0.05, which delin-
eates the probability of erroneously rejecting the null hypothesis when it is in
fact true. The value of α reflects the strength of the evidence for rejection, and
thus a small α imposes a requirement of strong evidence for the null hypothesis
to be rejected. In addition, it is necessary to select an appropriate test statistic in
accordance with the nature of the data and the hypothesis under examination.
Popular test statistics include the z-score, t-score, F-statistic, and chi-square
statistic.4 Following is a complete example of one-sample hypothesis testing
to determine whether the mean of a population is zero. Suppose we have a ran-
dom sample X1, X2, · · · , Xn from some unknown distribution. We assume that
the sample mean is approximately normally distributed for large n, and we want
to test the null and alternative hypotheses:
H0 : µ = 0,
(4)
H1 : µ , 0,
where µ is the true mean of the population. To calculate the test statistic, we
use a t-statistic with n − 1 degrees of freedom:
X− µ X
T= √ = √ , (5)
s/ n s/ n

4 The Z-score quantifies the number of standard deviations by which a data point deviates from
the mean and is typically employed for large samples or when the population variance is known.
The t-score is used for small samples or unknown variances to compare means, while the chi-
square statistic tests categorical data for goodness-of-fit or independence. The F-statistic, used
in regression, evaluates variance ratios to assess group mean differences or model fit.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 13

where X and s are the sample mean and standard deviation. Under H0 , this
statistic approximately follows a t-distribution with n − 1 degrees of freedom.
After computing the test statistic using sample data, we derive a value that
can be compared against a critical value for a given alpha. The probability of
observing a test statistic, under H0 , that is more extreme than the one we com-
puted from our data is called the p-value. The p-value decides the statistical
significance of our results compared to the null hypothesis. In the previous
example, we can find p = 2 × P(Tn−1 > |tobs |) where tobs is the observed value
of the statistic, and Tn−1 denotes a random variable following the t-distribution
with n−1 degrees of freedom. If the resulting p-value is smaller than the signifi-
cance level, for example, α = 0.05, then we can say the test result is significant,
indicating strong evidence against the null hypothesis.
In the previous section, we use graphical tools to assess data distributions
but we can now also use statistical hypothesis testing to validate data proper-
ties. For instance, the Jarque-Bera test can be utilized to assess the validity of
the normality assumption. This widely recognized statistical method evaluates
whether the sample data skewness and kurtosis are consistent with those
expected in a normal distribution, thereby determining if the return distribution
adheres to normality.

2.5 Serial Covariance, Correlation, and Stationarity

For time-series forecasting, we make predictions based on historical obser-
vations from previous time stamps. While increments in financial time-series
are close to independent random variables, and are indeed often modeled by
stochastic processes such as Geometric Brownian Motion (GBM), it is still pos-
sible to identify and exploit small dependencies to make predictions. We can
measure such dependence with serial covariances (autocovariances) or serial
correlations (autocorrelations).
Serial covariance refers to the measure of how two variables change together
over time within a time-series. In the context of financial time-series, it spe-
cifically measures the covariance between different observations of the same
financial variable at different points in time. Intuitively, autocovariance tells us
how two instances of a time-series (xt, xs ) at different time points move together.
Understanding serial covariance is beneficial for identifying patterns and pre-
dicting future values based on historical data. For example, if the price of a
stock today is positively correlated with its price yesterday, this might indicate
a future upward trend.
Correlation quantifies both the magnitude and the orientation of the linear
association between two time-series. In contrast to covariance, correlation is

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 14 Quantitative Finance

both standardized and dimensionless, offering a uniform metric for assessing

the extent to which two variables vary together. The correlation coefficient
spans from -1 to 1, where 1 signifies a flawless positive linear association, -1
denotes a perfect negative linear relationship, and 0 implies the absence of any
linear relationship. Formally, we can define covariance and correlation as:

Covariance: Cov(Xs, Xt ) = E[(Xs − E(Xs ))(Xt − E(Xt ))],

= E(Xs Xt ) − E(Xs )E(Xt ), (6)
p
Correlation: Corr(Xs, Xt ) = CoV(Xs, Xt )/ Var(Xs )Var(Xt ).

Correlation is essential for many applications in financial time-series analy-

sis. By evaluating the extent to which two assets fluctuate together, investors
are able to design portfolios that reduce risk while optimizing returns. For
instance, when two assets exhibit low or negative correlations, a portfolio
combining the two can achieve lower overall volatility compared to portfo-
lios consisting of highly correlated assets. Correlation analysis is also used for
detecting market inefficiencies and arbitrage opportunities. For example, if two
assets are expected to be highly correlated due to economic or financial reasons
but deviate significantly at some point, one could speculate that this deviation
will shrink again.
We can check how a time-series relates to previous observations (at various
time lags) using Autocorrelation Function (ACF) plots and Partial Autocor-
relation Function (PACF) plots. The ACF measures the correlation between
a time-series and its own lagged (i.e., past) values. It indicates the degree to
which past values of a series influence its current values, providing insights
into the internal structure and patterns of the data. The ACF is especially effec-
tive in detecting trends, seasonal patterns, and various cyclical behaviors within
a dataset. Mathematically, the ACF at lag k for a time-series Xt is defined as:
Cov(Xt, Xt+k )
ρk = Corr(Xt, Xt+k ) = , (7)
σ2
where Cov(Xt, Xt+k ) is the covariance between Xt and Xt+k , and σ 2 is the vari-
ance of the time-series. The values of the ACF span from -1 to 1, where values
approaching 1 signify a robust positive correlation, those nearing -1 indicate a
strong negative correlation, and values around 0 imply minimal to no correla-
tion. A correlogram depicts the autocorrelation of a time-series as a function
of time lags. This plot can help identify the appropriate model for time-series
forecasting, such as an Autoregressive Moving Average model (ARMA), in
which the autocorrelation structure guides the selection of model parameters.
The PACF serves as an additional instrument in time-series analysis, quanti-
fying the correlation between a time-series and its lagged values while also

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 15

eliminating the linear effects of intermediate lags. Unlike the ACF, which
includes the cumulative effect of all previous lags, the PACF isolates the direct
effect of a specific lag. For instance, the PACF at lag k measures the correla-
tion between Xt and Xt+k after removing the effects of lags 1 through k − 1. This
allows for a clearer understanding of the underlying relationship at each spe-
cific lag, making it easier to identify the appropriate number of lags to include
in an autoregressive model.
Mathematically, we can define the PACF at lag k as the correlation between
Xt and Xt+k that is not accounted for by their mutual correlation with Xt+1, Xt+2,
· · · , Xt+k−1 . We can obtain PACF values by fitting a linear model with Xt and
the regressors standardized:

Xt = α + ϕk,1 Xt−1 + ϕk,2 Xt−2 + · · · + ϕk,k Xt−k, (8)

where ϕk,k is the PACF value for lag k, and ranges from −1 to 1. With stand-
ardization, the regression slopes become the partial correlation coefficient, as
correlation is effectively the slope we get when both the response and predic-
tors have been reduced to dimensionless “z-scores.” The PACF plot is used in
conjunction with the ACF plot to identify the order of an autoregressive (AR)
model. While the ACF helps in understanding the overall autocorrelation struc-
ture, the PACF helps pinpoint the specific lags that should be included in the
AR component of an ARMA model, ensuring a more accurate estimation.
In summary, the ACF and PACF are powerful tools that enable a deeper
understanding of time-series data, guiding the development of robust and
effective forecasting models. Their combined use allows for the precise iden-
tification of temporal structures, leading to improved predictions and better
decision-making in fields where time-series data is prevalent. Figure 2 shows
an example of ACF and PACF plots for the same underlying data. The shaded
area in the plot represents an approximate confidence interval around zero cor-
relation. In other words, it is a visual guide for checking which autocorrelation

Figure 2 Left: ACF plot; Right: PACF plot.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 16 Quantitative Finance

(or partial autocorrelation) lags are statistically significant from zero. We can
make these plots using the following code:
1 from [Link] import plot_acf , plot_pacf
2 import [Link] as plt
3

4 fig , ax = [Link] (1,2, figsize =(12 ,4))

5 plot_acf(df['ret'], lags =30, ax=ax [0]) # ACF
6 plot_pacf(df['ret'], lags =30, ax=ax [1]) # PACF
7 [Link] ()

Next, we discuss another concept that is commonly used in time-series

known as stationarity. Stationarity refers to the statistical property of a time-
series where its key characteristics, such as mean, variance, and autocovariance
structure, remain constant over time. This consistency makes stationary time-
series easier to analyze and model, as their behavior is predictable and stable.
In financial markets, where data often exhibit complex patterns, achieving
stationarity is helpful for accurate modeling, forecasting, and risk management.
There are two forms of stationarity: strict stationarity and weak stationarity.
We say that a time-series process is strictly stationary if the joint distribu-
tion f (Xt1 , · · · , Xtk ) is identical to the joint distribution f (Xt1 +τ , · · · , Xtk +τ ) for
all collections t1, · · · , tk and separate values τ. However, this assumption is
very restrictive and very few real-world examples meet this requirement.
Differently, we say that a time-series process is weakly stationary if:

• E(Xt ) = µ < ∞ where the mean of a time-series is constant and finite,

• Var(Xt ) = σ 2 < ∞ where the variance of a time-series is constant and
finite,
• the autocovariance and autocorrelation functions only depend on the lag:
γt,t+τ = Cov(Xt, Xt+τ ) = γτ ,
(9)
ρt,t+τ = Corr(Xt, Xt+τ ) = ρτ .
A considerable number of statistical and econometric models operate under
the assumption that the underlying time-series remains stationary. These mod-
els depend on the stability of statistical characteristics to generate precise
forecasts. In finance, stationarity ensures that historical risk measures, such as
volatility and correlation, remain relevant for future periods. In practice, finan-
cial time-series frequently display non-stationary characteristics as a result of
trends, seasonal patterns, and structural shifts. To achieve stationarity, ana-
lysts use various techniques. For example, they might use differencing to
subtract previous observation from the current observation with the aim of
removing trends and achieving a stationary series. Additionally, they might
remove a deterministic trend component from a series (detrending) or apply
transformations like logarithms to stabilize the variance.
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 17

2.6 Time-Series Models

We have introduced ACF and PACF which can be used to determine the order
of AR and ARMA models. But what exactly are these models? In simple words,
these are classical time-series models that form a crucial aspect of analyzing
sequential data, particularly in fields such as finance, economics, and envi-
ronmental science. By identifying patterns and correlations in historical price
data, we can employ time-series models to develop quantitative trading strate-
gies by exploiting patterns for profit. AR models, for instance, can help in
detecting momentum or mean-reversion patterns, which are commonly used
in algorithmic trading.
Beyond financial markets, time-series models are vital for economic policy
and planning. Central banks and governmental bodies employ these models
to predict key economic metrics, including GDP expansion, inflation levels,
and unemployment figures. Accurate forecasts help make informed policy
decisions that can stabilize the economy and promote growth. Within the spec-
trum of time-series models, the Autoregressive model (AR), moving average
model (MA), and Autoregressive Moving Average model (ARMA) models are
among the most fundamental due to their simplicity and effectiveness. These
models are capable of capturing the underlying patterns and dynamics of time-
series data, making them powerful tools for analysts and researchers aiming to
model and forecast temporal data accurately.
The AR is among the most basic and extensively employed time-series mod-
els. It describes the current value of a series as a linear aggregate of its past
values combined with an independent random error component. If an AR model
takes p previous observations, we denote the model as AR(p) and its functional
form is given as follows:
Xt = ϕ1 Xt−1 + · · · + ϕp Xt−p + ϵt, (10)
where Xt is the value at current time stamp t, ϕ1, · · · , ϕp are the model param-
eters that represent the effects of past values on the output value, and ϵt is an
error term. For a given order p, we can fit the model and find the optimal coef-
ficients ϕ1 in the same way as we would fit a linear regression, with the lagged
values (Xt−1, ..., Xt−p ) being the predictors and Xt the target. In order to decide
which order p to use, we can look at the PACF plot. For example, considering
the time-series presented in Figure 2, the PACF suggests that an AR(3) model
would probably capture most of the dependence, while it might be useful to
consider an AR(9) model to capture further dependence.
The AR model captures how the current observations depend on their past
values, making it suitable for modeling time-series with strong autocorrela-
tions. This model is particularly effective when the underlying process is driven
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 18 Quantitative Finance

by its own past values, which can be the case for stock prices or interest rates.
For example, an AR(1) model, where the output depends only on the immediate
past observation, is defined as:

Xt = ϕ1 Xt−1 + ϵt, (11)

where the model suggests that the current value of the time-series is influenced
directly by the observation at time t − 1.
Differently than the AR model, the MA model represents the current value
of a time-series as a linear combination of its previous error terms. The MA
model of order q, symbolized as MA(q), is defined as:

Xt = µ + ϵt + θ 1 ϵt−1 + θ 2 ϵt−2 + · · · + θ q ϵt−q, (12)

where µ is the mean of the series, ϵt, · · · , ϵq are error terms, θ 1, · · · , θ q are
the model parameters that represent the influence of past errors on the current
value. The MA model captures the influence of past shocks or disturbances on
the current observation, making it useful for modeling time-series with short-
term dependencies. This model is effective when the series is subject to random
shocks that have a lasting but diminishing impact over time. For instance, an
MA(1) model defines that the observation at time t is only influenced by the
immediate past error:

Xt = µ + ϵt + θ 1 ϵt−1, (13)

although, in practice, we can include several terms to model output. To decide

the order of a MA model, we check the ACF plot and obtain the estimated point
at which the correlation diminishes.
The ARMA model integrates features from both AR and MA models, offer-
ing a more adaptable and thorough methodology for time-series analysis. The
ARMA model with order (p, q), represented as ARMA(p, q), is defined as:
Õ
p Õ
q
Xt = ϕj Xt−j + ϵt + θ j ϵt−j, (14)
j=1 j=1

where an ARMA model proficiently models long-term dependencies via

its AR components and addresses short-term disturbances through its MA
components.
In financial time-series analysis, precise forecasting is important for
informed investment choices and the formulation of effective trading strate-
gies. AR, MA, and ARMA models provide systematic ways to predict future
price movements based on historical data. An AR model can forecast future
stock prices by considering the past price movements, while an MA model
can evaluate the impact of past market shocks on future prices. ARMA models

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 19

are often used to estimate future volatility, an essential component of pricing

derivatives and constructing risk-hedging strategies.
Given that the aforementioned models are linear, they should always serve as
a benchmark before testing any of the more complex, nonlinear deep learning
models that are described in later sections. Linear models also have the added
benefit of being easy to interpret. This helps form a better intuition for any
investment ideas, before moving to more powerful but less interpretable deep
learning models.

2.7 Extras
Alpha and Beta In quantitative finance, the notions of alpha and beta
are very important to understanding and evaluating the performance of invest-
ment strategies. These metrics are derived from the Capital Asset Pricing Model
(CAPM) and are used to measure the returns and risk associated with individ-
ual assets or portfolios relative to a benchmark, typically a market index. In
quantitative trading, where strategies are often driven by mathematical models
and algorithms, alpha and beta provide essential insights into the effectiveness
and characteristics of trading approaches.
Alpha assesses an investment’s performance relative to a benchmark index.
More specifically, it represents the surplus return that an investment or port-
folio achieves beyond the expected return predicted by the CAPM. In other
words, alpha signifies the additional value that a trader or investment strat-
egy contributes over what is anticipated based on the asset’s systematic risk.
Conversely, beta measures an investment’s responsiveness to market fluctua-
tions. It quantifies the relationship between the investment’s returns and those
of the overall market or benchmark, indicating the extent to which the invest-
ment’s returns are expected to vary in reaction to changes in the market index.
Mathematically, we define alpha (α) and beta (β) as:

α = Ri − [Rf + β(Rm − Rf )], (15)

where Ri is the return of the investment, Rf is the risk-free rate and Rm is the
return of the market. A positive alpha signifies that the investment has sur-
passed the benchmark, whereas a negative alpha indicates underperformance.
In quantitative trading, generating alpha is the primary goal as it reflects the
ability of a trading strategy to consistently beat its benchmark through superior
stock selection, timing, or other factors. Beta values have different meanings.
A beta exceeding 1 signifies that the investment is more volatile than the mar-
ket, indicating it tends to amplify market movements in response to changes.
Conversely, a beta below 1 indicates that the asset’s returns are less sensitive to
market movements than the market index itself. If an investment has a negative
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 20 Quantitative Finance

beta, it means that the investment moves inversely to the benchmark. We can
also think of beta as the covariance between strategy and market returns scaled
by the market’s variance.
A strategy that consistently generates positive alpha is considered successful,
as it indicates the ability to surpass the market performance on a risk-adjusted
basis. On the other hand, beta helps traders understand the risk profile of their
strategies and manage risk exposure to market volatility. For instance, a trader
seeking to minimize risk might construct a low-beta portfolio, while one aiming
for higher returns might opt for higher-beta assets. By utilizing alpha and beta
metrics, quantitative traders can make well-informed decisions and enhance
their trading performance.

Volatility Clustering Volatility clustering is an extensively observed phe-

nomenon in financial markets, characterized by sequences of high volatility
periods that are succeeded by similarly high volatility periods, and periods of
low volatility that are followed by similarly low volatility periods. This char-
acteristic implies that volatility is not constant over time but instead exhibits
temporal dependencies, forming clusters. This is one of the reasons why finan-
cial returns deviate from the normal distribution. This observation is known
as heteroskedasticity and describes the irregular pattern of the variation of a
process.
Figure 3 shows the returns of the S&P 500, and we can clearly see that large
returns tend to cluster. This means that large fluctuations in prices tend to occur
together, persistently amplifying the amplitudes of price changes. Such behav-
ior contradicts the assumption of constant variance in traditional models like
the classical linear regression model and calls for models that can accommodate
changing variances, in order to make reliable predictions.

Figure 3 Returns of the S&P 500 over 60 years.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 21

There are two popular models used to capture and analyze volatility cluster-
ing: the Autoregressive Conditional Heteroskedasticity (ARCH) model and the
Generalized Autoregressive Conditional Heteroskedasticity (GARCH) model.
These models help in capturing the changing variance over time and provide a
better fit for the distribution of returns. Instead of predicting returns Rt , we now
model the variance of returns. An ARCH(p) process of order p is defined as:

Var(Rt |Rt−1, · · · , Rt−p ) = σt2 = α0 + α1 R2t−1 + · · · + αp R2t−p, (16)

where the variance of the process at time t is determined by observations from

the earlier time step. Accordingly, the ARCH model allows for fluctuations in
conditional variance over time, effectively capturing volatility clustering.
The GARCH model extends the ARCH model by including past conditional
variances into the model, providing a more flexible and parsimonious model
for capturing volatility dynamics. We denote a GARCH(p, q) as:
Õ
p Õ
q
σt2 = α0 + αj R2t−j + βj σt−j
2
, (17)
j=1 j=1

where the GARCH model presents a dual dependence that is better at modeling
both short-term shocks and sustained persistence in volatility over time.
In practical terms, volatility clustering means that markets experience peri-
ods of turmoil and periods of calm. ARCH and GARCH models offer powerful
methods for analyzing this phenomenon, enabling more accurate forecast-
ing, risk management, and pricing of financial instruments. By recognizing
the temporal dependencies in volatility, these models enable us to better
understand market behavior and enhance decision-making in various financial
applications.

3 Supervised Learning and Canonical Networks

In this section, we explore the essential concepts of supervised learning, an
important subset of machine learning that identifies relationships between input
data and output labels using example input-output pairs. Supervised learning
is extensively applied in a variety of domains, including image recognition,
natural language processing, financial forecasting, and medical diagnosis. By
mastering the fundamentals of supervised learning, we can proficiently train
models to generate accurate predictions and make informed decisions based on
new, unseen data.
Supervised learning entails training a model on a labeled dataset, in which
each input is paired with its corresponding correct output. The model then
learns to associate inputs with outputs by minimizing the discrepancy between

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 22 Quantitative Finance

its predictions and the actual results. This methodology includes choosing
suitable algorithms, adjusting hyperparameters, and assessing the models’
effectiveness. In this section, we will examine these concepts comprehen-
sively, establishing a robust foundation for comprehending and implementing
supervised learning methodologies.
Additionally, we will introduce various neural network architectures, which
have become the cornerstone of modern machine learning. Neural networks,
modeled after the architecture of the human brain, are composed of intercon-
nected layers of nodes (neurons) that process and transform input data. We
will cover canonical neural network models, including feed-forward neural net-
works and state-of-the-art networks such as transformers, each designed for
specific types of data and tasks.
Upon finishing this section, you will have a detailed understanding of the
core concepts underpinning supervised learning and will better understand var-
ious types of neural networks. This knowledge will equip you with the skills
to apply these powerful techniques to a wide range of applications, unlocking
new possibilities in data analysis, prediction, and decision-making.

3.1 Supervised Learning: Regression and Classification

Supervised learning is at the core of machine learning and it is a process that
essentially learns, or in other words, fits a mapping between an input and an
output. Formally, for a regression task, it maps an input x ∈ Rd to an output
y ∈ R through a learned function by training on example input-output pairs.
We call this collection of example input-output pairs upon which the model is
fitted the training set, and it can be expressed as:

{(x1, y1 ), (x2, y2 ), · · · , (xN, yN )}. (18)

A supervised learning algorithm infers a function f that best defines the inter-
play between inputs and outputs by utilizing training data. The inferred function
can then be used to make estimates for new inputs. The function f can be as sim-
ple as a linear function or it can also be a highly nonlinear function as obtained
through deep learning models. During training, the true output values (labels)
are available, and our goal is to reduce the differences between the predicted
results and these actual labels. In mathematical terms, this reads:

min L(y,ŷ) with ŷi = f (xi ), (19)

where L is a choice of metric, known as a loss function or an objective func-

tion, that measures the difference between real outputs (y) and predicted values
(ŷ). After learning a functional mapping on the training set, we apply it to

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 23

unseen test data {x1′ , x2′ , · · · , xM

′ } and evaluate the performance of our learned

function. In general, a supervised learning problem goes through the following

steps:

1. Define the prediction problem,

2. Gather a training set that is representative of the application domain,
3. Carry out an exploratory analysis and select input features,
4. Choose the approximate learning algorithm and decide the model’s archi-
tecture,
5. Conduct model training on the training set and optimize hyperparameters
using a separate validation set,
6. Assess the effectiveness of the trained function using a test dataset.

Depending on outputs y, we can divide supervised learning algorithms into

two categories: regression and classification. When the output (y) takes con-
tinuous values, it is a regression problem. For example, stock prices and the
weights and heights of a person are all examples of continuous values that
would correspond to a regression task. A classification problem deals with
discrete outputs, such as whether an image contains a dog or not. The training
framework for regression and classification is very similar, with the exception
of the design of the objective function. We now discuss each problem type in
detail.

Regression One key aspect of supervised learning is the selection of an

appropriate loss function – also referred to as a cost or objective function – that
measures the discrepancy between a model’s predictions and the corresponding
true target values. The loss function guides the learning process by provid-
ing a measure that the model aims to minimize during training. As previously
noted, regression problems focus on the prediction of a continuous variable.
For regression problems, one of the most commonly used objective functions
is the mean-squared error (MSE):

1Õ
N
L(y,ŷ) = (yi − ŷi )2, (20)
N i=1
where the loss is merely the sum of residuals, ϵi = yi − ŷi , squared which we
aim to minimize to obtain a good fit to the data.5 The MSE loss is symmetric
and places greater emphasis on larger errors in the dataset.

5 We can also obtain the MSE for a linear model by a Maximum Likelihood approach, where we
Î
start with the likelihood of the data, i p(ϵi ), and assume that the distribution of the residuals
p(ϵi ) is Gaussian. Maximizing the likelihood of the data is equivalent to minimizing the log-
likelihood which, up to a constant, is equivalent to the MSE.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 24 Quantitative Finance

Table 1 Objective functions for regression problems.

Metrics Formula
q Í
N i (yi − ŷi )
1 N 2
Root mean squared error (RMSE)
1 ÍN
N i (ln(1 + yi ) − ln(1 + ŷi ))
Mean squared log error (MELE) 2

1 ÍN
Mean absolute error (MAE) N i |yi − ŷi |
Median absolute error (MedAE) median(|y1 − ŷ1 |, · · · , |yN − ŷN |)
( Í
N Íi 2 (yi − ŷi ) , if |yi − ŷi | ≤ δ
1 N 1 2
Huber loss (HL)
N i δ|yi − ŷi | − 2 δ , otherwise
1 N 1 2

There are numerous options for objective functions. For example, the mean-
squared logarithmic error can be applied to outputs that exhibit exponential
growth, imposing an asymmetric penalty that is less harsh on negative errors
than on positive ones. Both the mean absolute error (MAE) and the median
absolute error (MedAE) are symmetric and do not assign additional weight to
larger errors. Moreover, Huber loss, which merges aspects of the mean squared
error and the mean absolute error, is resistant to outliers and can be used to
stabilize training when working with noisy data. Table 1 summarizes some
common loss functions for regression problems. It is also very straightforward
to implement these losses:
1 from [Link] import (mean_squared_error ,
2 mean_squared_log_error ,
3 mean_absolute_error ,
4 median_absolute_error)
5 from [Link] import huber
6 import numpy as np
7

8 def huber_loss(y_true , y_pred , delta =1.0):

9 y_true = [Link](y_true)
10 y_pred = [Link](y_pred)
11 return [Link](huber(delta , y_true - y_pred))
12
13 mse = mean_squared_error(y_true , y_pred)
14 rmse = [Link](mse)
15 msle = mean_squared_log_error(y_true , y_pred)
16 mae = mean_absolute_error(y_true , y_pred)
17 med_ae = median_absolute_error(y_true , y_pred)
18 huber = huber_loss(y_true , y_pred , delta =1.0)

The best choice of objective function depends on the specific task. Some-
times, we can create customized loss functions to ensure that performance
metrics best reflect the consequences of incorrect predictions. For example,
in applications like medical diagnosis or fraud detection, false negatives may

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 25

be more costly than false positives. In such cases, loss functions can be tailored
to penalize certain types of errors more severely, aligning the model’s training
with the specific needs of the problem.

Classification Unlike regression, classification aims to place input data into

predefined categories or classes. It does so by analyzing a labeled dataset
where each example is matched with a class label. Once trained, a classifica-
tion model can predict labels for unseen data based on the learned patterns and
relationships. Classification techniques are commonly applied in fields such
as finance, healthcare, and marketing. Classification problems can be broadly
categorized into binary classification and multi-class classification. Binary
classification involves two distinct classes. Common examples of binary clas-
sification include determining whether an email is spam, predicting if a credit
card transaction is fraudulent, or diagnosing a patient as healthy or ill. Multi-
class problems involve more than two classes, such as classifying handwritten
digits (0–9), categorizing types of flowers (e.g., the Iris dataset), or classifying
news articles into different topics.
Since classification problems have discrete outputs, we first have to pro-
duce scores, or logits, to indicate the likelihoods that an observation belongs to
certain classes. These scores can then be normalized across all possible class
labels to obtain corresponding probabilities p̂i . After that, we use these scores
or probabilities p̂i to make actual predictions ŷi either by taking the class with
the highest score or by using threshold values. In the simplest case of a binary
classification problem, we first define the logistic function
1
σ(z) = , (21)
1 + e−z
which squashes any real-valued input into the open interval (0,1). We can then
model the probability of the prediction being positive as
p(yi = 1|xi ) = σ(wT xi ). (22)
The same mapping can be represented by
p̂i
logit(p̂i ) = log = wT xi, (23)
1 − p̂i
where the logit of the probabilities is a linear function of input features (xi ). The
objective functions for classification problems are different from regression
as there are instead finite distinct outcomes. In most cases, we choose cross-
entropy loss as the objective function for classification. In the binary case, the
cross-entropy is calculated as:

−(y log(p) + (1 − y) log(1 − p)), (24)

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 26 Quantitative Finance

where p is the probability of predicting y = 1. The loss function is then

computed by summing the cross-entropy of each data point:
Õ
N
L(y, p̂) = − (yi log(p̂i ) + (1 − yi ) log(1 − p̂i )). (25)
i=1

The functional form of the cross-entropy loss might be less intuitive than
the MSE, but it can still be understood within the context of maximum likeli-
hood estimation.6 If we deal with a multi-class classification problem (M > 2),
a separate loss is needed for each class label and a summation is taken at
the end:
Õ
M
− yoc log(poc ), (26)
c=1

where yo,c is a binary indicator (0 or 1) that is activated when the model assigns
the right label c for observation o, and p is the output from the algorithm which
indicates the predictive probability of observation o for class c. The loss of the
data is then obtained by summing the multi-class cross-entropy of each point.
Once the predicted probabilities are transformed into predictions of one class
or another, we can evaluate model performance through several metrics. To
illustrate those metrics we focus on binary classification problems for simplic-
ity. A frequently employed measure is the misclassification rate which can be
defined as the fraction of misclassified labels:
1Õ
N
Misclassification rate = Iy ,ˆy . (27)
N i=1 i i

The confusion matrix is another important tool that can be used to visu-
alize various metrics. Table 2 illustrates a confusion matrix that enumerates
the quantities of correct and incorrect predictions for every class. For exam-
ple, the False Positives in the top right corner represent errors where an actual
label is negative but a prediction is positive. In the context of a stock price
reversion example, this would be a case when a stock price does not revert
but we predicted that it would revert. Such an error is much more costly to us
than a False Negative, where a stock does actually revert but we predicted it
would not.
Following the notation in the confusion matrix, we can thus introduce other
popular evaluation metrics, which are shown in Table 3. For instance, accuracy
is computed by summing the diagonal entries in the confusion matrix and then

Î y
6 The likelihood of the data is i p̂i i (1 − p̂i )1−yi , which merely assigns a probability p̂i to each
point with label yi = 1 and a probability 1 − p̂i to each point with label yi = 0. The negative
log-likelihood then corresponds to our loss function.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 27

Table 2 Confusion matrix.

Actual
Positive Negative

Positive True Positive (TP) False Positive (FP)

Prediction
Negative False Negative (FN) True Negative (TN)

Table 3 Evaluation metrics for classification problems.

Metrics Formula
TP+TN
Accuracy TP+FP+TN+FN
TP
Precision TP+FP
TP
Recall TP+FN
F1 2 × Precision×Recall
Precision+Recall

dividing by the total number of predicted samples. Accuracy thus represents the
proportion of total predictions that are correct. Precision indicates the fraction
of predicted positives that are truly positive, while recall measures the fraction
of actual positives correctly identified. Lastly, the F1 score balances precision
and recall by using their harmonic mean.
It is very important to check all evaluation metrics when analyzing model
performance since a single performance metric can indicate misleading results.
For example, in an unbalanced data set, where 90% of labels are +1, we can
get an accuracy score of 90% by simply predicting everything as +1, even
though the model has not learned anything. Another issue arises when we assign
different importance to different types of errors. For example, a mean rever-
sion strategy usually makes frequent small gains but can make infrequent large
losses when a stock does not revert. Such a strategy might demonstrate a high
accuracy for predicting stock reversion but still lead to significant losses. To
implement these metrics, we can use the following code:
1 from [Link] import accuracy_score , f1_score ,
confusion_matrix , classification_report
2
3 acc = accuracy_score(y_true , y_pred)
4 f1_macro = f1_score(y_true , y_pred , average='macro ')
5 report = classification_report (y_true , y_pred)
6 cm = confusion_matrix(y_true , y_pred)
7
8 print("Accuracy:", acc)
9 print("F1 Score (macro):", f1_macro)

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 28 Quantitative Finance

10 print("Classification Report:")
11 print(report)
12 print("Confusion Matrix:")
13 print(cm)

Instead of using numerical values to assess model performance, we can

also use graphical tools. The receiver operating characteristics (ROC) curve
enables us to compare and choose models that are conditioned on their respec-
tive predictive performance. For this purpose, we need to compare predicted
probabilities with selected thresholds to decide final outcomes. Accordingly,
different thresholds yield different results. The ROC curve derives pairs of true
positive rates (TPR) and false positive rates (FPR) by examining every possible
threshold for classification, and then displays these pairs on a unit square plot.
We define TRP and FPR as the following:

TP
TPR = ,
TP + FN (28)
FP
FPR = .
FP + TN
Random predictions, on average, yield a diagonal line on the ROC curve
which has equal TPR and FPR rates. This diagonal line is the benchmark case,
so if the curve falls on the left side of the diagonal line, the learned model
is better than random guessing. The further from the margin, the better the
classifier (shown in Figure 4). We refer to the area under the ROC curve as

Figure 4 An example of different ROC curves.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 29

AUC and it is a summary measure that tells how good a classifier is. A higher
AUC score indicates a better algorithm.

3.2 Fully Connected Networks

After reviewing the basics of supervised learning, we now look at canonical
examples of neural network architectures. Fully connected networks (FCNs),
also known as multilayer perceptrons, are one of the earliest and most basic
neural networks. FCNs are indispensable in the field of deep learning, as they
provide the foundational architecture upon which other more sophisticated neu-
ral network models are built. FCNs have powerful abilities to approximate
complex functions and patterns within data, making them highly versatile and
applicable across various domains.
To gain a thorough understanding of FCNs, let us start with a simple linear
model, as a linear model can be viewed as a single neuron with a fully connected
input layer. Suppose we have an input of vector form x ∈ RNx and a scalar
output y ∈ R. A linear model posits that the prediction for an input vector x can
be determined by:
y = wT x + b. (29)
This is precisely how a single neuron (in many neural network frameworks)
computes its output – via a linear combination of inputs. In neural networks,
this single linear neuron can be extended by stacking many such layers (and
adding nonlinearities) to get more expressive models. However, at its core, a
single neuron’s linear component is identical to the linear regression formula.
For linear regression under a least squares objective, we can write the objective
function as:

1Õ
N
L= ( yi − ŷi )2, (30)
N i=1
where N is the number of sample points. We can therefore optimize the model
parameters by setting the partial derivatives of the L with respect to w and b
to 0:
∂L ∂L
= 0 and = 0, (31)
∂w ∂b
and the aforementioned can be solved analytically. In fact, by setting b = 0 for
simplicity, and writing X = (x1, ..., xN )T , the solution can easily be obtained as:

w = (XT X )−1 XT y. (32)

Moreover, one can directly recover the general case with b , 0 by noting that
we can always interpret the bias b as a weight w0 of a constant predictor x0 = 1.
While the analytical solution provides a direct method to find the optimal
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 30 Quantitative Finance

parameters, it may not be practical for large-scale problems as numerical matrix

inversion can become inefficient. As an alternative, gradient descent offers an
iterative approach to approximate the optimal parameters by minimizing the
MSE:
∂L
w=w−α ,
∂w (33)
∂L
b=b−α ,
∂b
where α is the learning rate. This is a rather simple application of gradient
descent, especially given that the problem is convex. However, even for a sim-
ple linear regression it can be beneficial to use gradient descent. For example,
using online gradient descent is a viable strategy for solving cases where the
data is so large that it does not fit in memory. Moreover, it is very easy to
substitute different loss functions, such as MAE or Huber loss in Table 1.
Note that the notion of learning parameters from data via gradient descent
is also the core concept behind more complex neural network training. Under-
standing how gradients are computed and used to update parameters in linear
regression aids in comprehending backpropagation in neural networks. Once
again, a single linear neuron is essentially a linear regression with optional
activation (defined in the next page). The extension to multiple neurons and
stacking them is the essence of deep neural networks.
One of the significant advantages of FCNs is their capacity for universal
approximation. According to the Universal Approximation Theorem (Hornik,
Stinchcombe, & White, 1989), a single-hidden-layer feed-forward network
with an adequately large neuron count can approximate any continuous func-
tion. This property makes FCNs incredibly powerful for tasks involving
function approximation. Further, many state-of-the-art models, such as con-
volutional neural networks (CNNs) and recurrent neural networks (RNNs),
incorporate fully connected layers as integral components. In CNNs, for
instance, FCNs are used in the final stages to consolidate the features from
the last hidden layers to make predictions.
In general, FCNs receive an input of vector form x ∈ RNx and map it to an
output (here a scalar) y ∈ R through a function y = f(x|θ). The vector θ com-
prises all model parameters, which we iteratively update to achieve the optimal
function approximation. Additionally, an FCN can form a chain structure by
stacking multiple layers sequentially. Each layer is a function of the previous
layer. The first layer can be defined as:
h(1) = g(1) (W (1) x + b(1) ), (34)
where h(1) ∈ RN1 designates the first hidden layer, containing N1 neurons.
Additionally, the quantities W (1) ∈ RN1 ×Nx and b(1) ∈ RN1 denote the associated
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 31

Figure 5 An FCN with two hidden layers in which each hidden layer has
five neurons.

weight matrix and bias vector, respectively. The function g(1) (·) is called the
activation function. We then define the following hidden layers as:

h(l) = g(l) (W (l) h(l−1) + b(l) ), (35)

where the l-th hidden layer (h(l) ∈ RNl ) has weights W (l) ∈ RNl ×Nl−1 and
biases b(l) ∈ RNl . To better illustrate this, we present an example of an MLP
in Figure 5. At its core, each hidden layer computes a linear transformation of
the previous layer’s output, followed by a nonlinear activation. The ultimate
output is determined by the target’s nature and is once again derived from the
preceding hidden layer. The discrepancy between the model’s predictions and
the true targets is quantified using a specified loss or objective function. Gra-
dient descent is then employed to adjust the model parameters in an effort to
minimize this loss. We can easily build a fully connected network with Pytorch
using the following code snippet:
1 import [Link] as nn
2

3 class MLP([Link]):
4 def __init__(self , seq_length , n_features):
5 super(MLP , self).__init__ ()
6 self.flat_dim = seq_length * n_features
7 [Link] = [Link](
8 [Link] (),
9 [Link](self.flat_dim , 64),
10 [Link] (),
11 [Link] (64, 32),
12 [Link] (),
13 [Link] (32, 1)
14 )
15 def forward(self , x):
16 return [Link](x)
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 32 Quantitative Finance

The activation function requires special attention, as its argument is just a

linear combination of the model inputs. Therefore, if the activation function is
also linear then the overall function represented by the neural network would
also be linear. Thus, making the activation function nonlinear is what allows us
to represent complex nonlinear functions with neural networks.7
A variety of activation functions exist and general choices include hyper-
bolic tangent function, sigmoid function, Rectified Linear Units (ReLU)
(Nair & Hinton, 2010), and Leaky Rectified Linear Units (Leaky-ReLU)
(Maas et al., 2013). Figure 6 plots some of these activation functions. The
ReLU function is prevalent in modern applications, and empirical research
advises initiating experimentation with ReLU while simultaneously evaluat-
ing other activation functions (Mhaskar & Micchelli, 1993). Leaky-ReLUs can
also be used to avoid some of the gradient issues caused by the flat part of the
ReLU. In broad terms, the selection of activation function to use is dictated by
the application context and must be substantiated through validation studies.
The same rationale for choosing activation functions likewise applies to other
network hyperparameters.

3.3 Convolutional Neural Networks

Convolutional neural networks (CNNs) constitute a class of deep learning
architectures meticulously engineered to process images and other structured
grid-like data. These networks have fundamentally changed the landscape of
computer vision, allowing machines to carry out tasks such as image recog-
nition, object detection, and image segmentation with performance levels that
rival human capabilities. Drawing inspiration from the visual cortex of ani-
mals, the architecture of CNNs enables them to automatically and adaptively
learn spatial hierarchies of features from input data, thereby enhancing their
effectiveness in various pattern recognition tasks.
CNNs are arguably the most important network structures as they inspired
much of the development of modern deep learning algorithms over the past
decade through initial breakthroughs in performance on image recognition
problems. Marking a pivotal breakthrough in computer vision, Krizhevsky,
Sutskever, and Hinton (2017) introduced the first convolutional neural network
successfully applied to large-scale image tasks. Often in image problems, fea-
tures learned by neural networks have intuitive interpretations such as edges or
surfaces. Whereas an FCN would need to relearn a feature for every part of an

7 Mathematical theory shows that a single hidden layer MLP with infinitely many neurons can
represent any continuous function. In practice, however, one aims for deeper, rather than wider,
networks as they make it easier to learn good feature representations.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Figure 6 Plots of various activation functions.
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 34 Quantitative Finance

image, the CNN architecture enables the model to learn the same feature for
different parts of an image by sliding a convolutional filter across it.
Subsequently, CNNs were studied and applied to many domain areas. We
now demonstrate how the same concept can be used for time-series problems.
Time-series data, characterized by its sequential and temporal nature, can ben-
efit from the unique ability of CNNs to detect patterns and trends over different
scales. By adapting the convolutional operations used in image processing to
time-series data, CNNs can effectively capture local dependencies and extract
representative features. These qualities make them strong tools for time-series
forecasting, anomaly detection, and classification.
Unlike MLPs that receive inputs in vector format, CNNs are adept at process-
ing grid-structured input data through the use of two specialized layer types:
convolutional layers and pooling layers. Convolutional layers constitute the
primary components of a CNN, with each convolutional layer containing multi-
ple convolutional filters designed to extract local spatial relationships from the
input data. Convolutional filters, also known as kernels or feature detectors,
are designed to traverse and transform input data by detecting specific fea-
tures or patterns. In essence, a convolutional filter is a diminutive weight matrix
that slides across the input data, performing a dot product with each localized
region of the input. This procedure is referred to as the convolution operation.
We denote a standard convolutional filter as K and it processes the input data
X ∈ R NT ×Nx by utilizing a convolution operation:
ÕÕ
M−1 N−1
S(i, j) = (X ∗ K)(i, j) = X(i + m, j + n)K(m, n), (36)
m=0 n=0

where S signifies the resultant matrix (feature map) and (i, j) correspond to the
indices of its rows and columns. We denote the convolution process as ∗.
A single convolutional layer is capable of containing multiple filters, each
of which convolves the input data using a distinct set of parameters. The matri-
ces produced by these filters are often termed feature maps. Similar to MLPs,
these feature maps can be transmitted to subsequent convolutional layers and
subjected to activation functions to incorporate nonlinearities into the model.
In time-series modeling, the primary strategy involves applying convolutional
filters along the temporal axis, thereby enabling the network to discern and
learn temporal dependencies and patterns.
Another crucial component of a CNN is the pooling layer, which also fea-
tures a grid-like structure. This layer condenses the information from specific
areas of the feature maps by applying statistical operations to nearby outputs.
For example, the widely used max-pooling layer (Y.- T. Zhou & Chellappa,
1988) selects the highest value within a designated region of the feature maps,

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 35

whereas average pooling computes the mean value of that region. The study
by Boureau, Ponce, and LeCun (2010) explores the application of various
pooling methods in different contexts. However, in most scenarios, selecting
the appropriate pooling technique necessitates domain expertise and empirical
experimentation.
Pooling layers are utilized across various applications to make the resulting
feature maps relatively invariant to small changes in the input data. This type of
invariance is beneficial when the focus is on detecting the existence of particu-
lar features rather than their exact positions (Goodfellow, Bengio, & Courville,
2016). For instance, in certain image classification problems, it is only neces-
sary to recognize that an image contains objects with specific characteristics
without needing to pinpoint their exact locations. Conversely, in time-series
analysis, the precise timing or placement of features is often essential, and
therefore the use of pooling layers must be approached with caution.
In addition to convolutional and pooling layers, a CNN has additional pos-
sible operations: padding and stride. Padding is employed to preserve the
dimensions of the feature maps, as convolution operations would otherwise
“shrink” the dimension of original inputs (demonstrated in Figure 7). Padding
solves this by adding, or “padding,” the original inputs with zeros around
the borders (zero-padding) so that the resulting feature maps have the same
dimension as before (the top-right figure of Figure 7).

Figure 7 Top: an illustration of padding; Bottom: an illustration of stride.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 36 Quantitative Finance

Figure 8 An example CNN network that first goes through a convolutional

layer and then a pooling layer with a fully connected layer at the end.

Separately, stride is a commonly utilized parameter that controls how the

convolutional filter moves across inputs and can reduce the dimensionality of
input data. More concretely, stride defines the number of steps the filter takes
as it slides over the input matrix, impacting the dimension of the output feature
map and the amount of computational work required. Recall that a convolu-
tional filter “scans” an input and by default moves by a step size of (1,1). Stride
defines a different step size. For example, a stride of (2,2) has the effect of mov-
ing the filter two steps and thus decreases the original input by half (as shown
in Figure 7).
Padding and stride are central concepts within CNNs that manage the spatial
dimensions of output feature maps and determine the network’s effectiveness in
capturing and retaining information from the input data. By carefully choosing
padding and stride values, one can balance the trade-offs between compu-
tational efficiency and the level of detail captured in the features extracted
by the network. Finally, we can combine all of these components to con-
struct a convolutional network. Figure 8 shows a typical example of a CNN
and possesses an architecture that is standard and highly popular in image
classification problems. Other famous networks for further independent study
include “AlexNet” (Krizhevsky, Sutskever, & Hinton, 2012) and “VGGNet”
(Simonyan & Zisserman, 2014).

3.4 WaveNet
CNNs are naturally desirable for dealing with stochastic financial time-series
as convolutional layers have smoothing properties that facilitate the extraction
of valuable information and discard the noise. In addition, a convolutional filter
can be configured to have fewer trainable weights than fully connected layers.
To some extent, this remedies the problem of overfitting (defined in Section 4).

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 37

However, a convolutional filter summarizes information for local regions of the

input, so the receptive field from convolutional layers is limited, and in order
to consider the entire input sequence, we have to use many layers and such
operations could be highly inefficient.
We now introduce WaveNet, an architecture that specifically addresses this
issue by using dilated convolutions. A WaveNet is a deep generative model
developed by DeepMind (Van Den Oord et al., 2016) which generates raw
audio waveforms and represents a substantial milestone in audio synthesis and
processing. It is capable of producing highly realistic human speech and other
audio signals by directly modeling the raw waveform of the audio, unlike tradi-
tional methods that rely on intermediate representations such as spectrograms
or parametric models.
WaveNet has proven to be a powerful tool that can be effectively adapted for
time-series analysis. Time-series data, characterized by sequential and temporal
dependencies, presents a set of challenges that a WaveNet architecture is well-
suited to address. By leveraging its strengths to study prolonged dependencies
and capture intricate temporal patterns, A WaveNet offers a robust framework
for modeling time-series processes. The core of a WaveNet is the dilated causal
convolutions which enable the network to consider a broad context of past
observations. In essence, the dilated convolutions skip some elements in the
input, which allows the networks to access a larger range of inputs. Apart from
dilated convolutional layers, other important components of a WaveNet include
residual and skip connections and gated activation units.
The work of Borovykh, Bohte, and Oosterlee (2017) proposes that the dilated
convolution for a time-series has a large receptive field. Specifically, a dilated
convolution is defined as:
Õ Õ
∞ M l−1

(wlh ∗d f l−1 )(i) = wlh ( j, m)f l−1 (i − d · j, m), (37)

j=−∞ m=1

where Ml denotes the number of channels and d is the dilation factor. A dilated
convolutional filter operates with every dth element in the input, therefore, it
can access a broad range of inputs compared to standard convolutional filters.
The causal nature of the convolutions ensures that the model does not violate
the temporal order of the time-series, making it suitable for prediction tasks.
We can stack multiple such layers to extract even longer dependencies. For
a network with L dilated convolutional layers, we increase the dilation fac-
tor by two at each layer, so that d ∈ [20, 21, · · · , 2L−1 ], and the filter size w
is 1 × k := 1 × 2. As a result, the dilation rate exponentially increases with
each layer, allowing the network to efficiently model prolonged dependencies
over sequences. An example of a dilated convolutional network that consists

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 38 Quantitative Finance

Figure 9 A WaveNet with three layers. The dilation factors for the first,
second, and third hidden layers are 1, 2, and 4 respectively.

of three layers is illustrated in Figure 9. To incorporate nonlinearity into the

model, we use activation functions after each layer to transform the resulting
representations. A WaveNet that uses ReLU has output from layer l is:
f l = [ReLU(wl1 ∗d f l−1 ) + b, · · · , ReLU(wlMl ∗d f l−1 ) + b], (38)

where ∗d refers to a convolution performed with a dilation factor of d, b ∈ R

is the bias term, and f l indicates the output generated from a convolution using
the filters whl for each h = 1, · · · , Ml within layer l.
In general, a deep network can suffer from an unstable training process if
backpropagation becomes unstable during the process of differentiation across
multiple layers. This problematic phenomenon is called the degradation prob-
lem and was discussed by He, Zhang, Ren, and Sun (2016). The work of
He et al. (2016) proposes residual connections to solve this limitation by forc-
ing the network to approximate H (x) − x instead of H (x) (the outputs from an
intermediate layer). They suggest that optimizing the residual mapping is easier
and they implement this technique by adding the inputs and outputs from a neu-
ral layer together. In a WaveNet (Figure 10), each dilated convolutional layer
is followed by a residual connection. Specifically, the outputs from the acti-
vation function undergo a 1 × 1 convolution (a point-wise convolution) before
the residual connection is applied. This approach ensures that both the resid-
ual path and the output from the dilated convolution have the same number of
channels, allowing multiple layers to be stacked effectively. Finally, we can
use WaveNet as an autoregressive model for time-series forecasting. In this
context, the expectation for predicting every t ∈ {0, ..., N} is:

E[xt+1 |xt, · · · , xt−r ] = β1 xt−r + · · · + βr xt, (39)

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 39

Figure 10 The network structure of a WaveNet. The input is convolved in the

first layer and then fed to the following network layer with a residual
connection. The Condition refers to any other external information that the
network uses. This operation is repeated until the output layer L(M) and the
final forecast is made.

where βi are parameters optimized via gradient descent. To create a one-step-

ahead prediction, we compute x̂t+1 for t + 1 ≥ r by inputting the sequence
(xt+1−r, . . . , xt ). These predictions can subsequently be reintroduced into the
network to formulate n-step-ahead forecasts. For instance, a two-step-ahead
out-of-sample prediction x̂t+2 is produced by using the input (xt+2−r, . . . ,x̂t+1 ).

3.5 Recurrent Neural Networks

Recurrent neural networks (RNNs) form a class of neural architectures specif-
ically devised to detect patterns within sequential data, including time-series,
natural language, and speech signals. In contrast to conventional feed-forward
neural networks which presume independence among inputs, RNNs possess an
internal memory structure that captures details of preceding inputs. This design
feature enables RNNs to maintain context and appreciate temporal dependen-
cies, rendering them highly suitable for scenarios where the sequential order
and context of data are integral.
RNNs have a rich history that dates back to the early days of artificial intel-
ligence and deep learning research. The foundational idea behind RNNs was
to create a network that could process sequences of data and retain a memory
mechanism to model time-based dependencies, which traditional feed-forward
networks could not handle effectively. Given this idea, time-series data that
have sequential structures are well suited to the framework of RNNs.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 40 Quantitative Finance

Figure 11 A recurrent network that processes information from the input and
the past hidden state.

When dealing with time-series data, we can take previous observations to

form inputs. For a multivariate time-series, a single input x1:T ∈ R NT ×Nx has two
dimensions where xt represents the features corresponding to time t, and T is the
length of the input’s look-back window. To employ an MLP for modeling these
processes, the inputs need to be flattened prior to feeding them into the hidden
layers. However, this operation could break the existing time dependencies in
inputs. To address this potential problem, RNNs have a structure that maintains
an internal representation that preserves important temporal relationships
The RNN architecture leverages hidden states that function as a memory
mechanism, and recursively update with each new observation at every time
step. Consequently, this structure naturally carries information forward from
earlier inputs to current ones, extracting temporal patterns from the data. We
can define a hidden state as:
ht = f (ht−1, xt |W, b),
(40)
= g(W h ht−1 + W x xt + b),

where W h ∈ RNh ×Nh , W x ∈ RNh ×Nx , b ∈ RNh constitute the linear weights
and biases for the hidden state, while g(·) designates the activation function.
The quantity Nh corresponds to the number of hidden units, and Nx refers to
the number of input features observed at any time t. An illustrative example of
such an RNN is depicted in Figure 11.
Nevertheless, due to the model’s recursive architecture, taking the derivative
of the objective function with respect to its parameters involves a sequence
of multiplicative terms that could lead to vanishing or exploding gradients
for RNNs (Bengio, Simard, & Frasconi, 1994). This issue complicates the
back-propagation of gradients, resulting in an unstable training procedure and
limiting RNNs’ effectiveness in modeling long-term dependencies.
A significant breakthrough came in 1997 when Hochreiter and Schmidhuber
(1997) introduced the Long Short-Term Memory (LSTMs) network. LSTM
addressed the vanishing gradient problem by introducing memory cells and
gating mechanisms that allowed the model to retain and selectively update

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 41

information over prolonged sequences. This innovation dramatically improved

the ability of RNNs to learn and remember over longer time periods, making
LSTMs a crucial development in the field.
Much like an RNN, an LSTM maintains a chain of hidden states that undergo
recursive updates. However, the LSTM architecture incorporates an internal
memory cell, along with three gates – specifically the input gate, forget gate,
and output gate – that oversee how information flows into and out of the
cell. These gates facilitate the long-term maintenance and adjustment of the
network’s state over lengthy sequences. We define the gates as follows:

Input gate: it = σ(W i,h ht−1 + W i,x xt + bi ),

where W i,h ∈ RNh ×Nh , W i,x ∈ RNh ×Nx and bi ∈ RNh ,
Output gate: ot = σ(W o,h ht−1 + W o,x xt + bo ),
where W o,h ∈ RNh ×Nh , W o,x ∈ RNh ×Nx and bo ∈ RNh ,
Forget gate: f t = σ(W f,h ht−1 + W f,x xt + bf ),
where W f,h ∈ RNh ×Nh , W f,x ∈ RNh ×Nx and bf ∈ RNh , (41)

where we define ht−1 as the LSTM’s hidden state at time step t−1 and apply the
sigmoid activation function σ(·). The parameters W and b represent the model’s
weights and biases. The resulting cell state and hidden state at the current time
step are then described by:
Cell state: ct = f t ⊙ ct−1 + it ⊙ tanh(W c,h ht−1 + W c,x xt + bc ),
(42)
Hidden state: : ht = ot ⊙ tanh(ct ),
where ⊙ is the element-wise product, W c,x ∈ RNh ×Nx , bc ∈ RNh , W c,h ∈ RNh ×Nh ,
and tanh(·) is the hyperbolic tangent activation function. Figure 12 plots an
LSTM cell with all gates mechanisms.
LSTMs have been applied successfully in numerous fields because of their
unique properties for dealing with prolonged sequences. For instance, LSTMs

Figure 12 An illustration of an LSTM cell.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 42 Quantitative Finance

are widely used in language modeling, where they predict the next word in a
sequence, as well as in machine translation, where they translate text from one
language to another. For financial applications, LSTMs are also well studied
and there exists a large amount of literature that applies LSTMs to predict finan-
cial time-series. Despite their success, LSTMs still suffer from several issues.
Firstly, due to the gating mechanism and cell structure, LSTMs are very com-
plex, which leads to a considerable amount of parameters that must be learned.
As a result, the problem of overfitting is severe in certain applications. LSTMs
are also computationally intensive, requiring lengthy training schedules.
In an effort to address the complications and drawbacks of LSTMs, Cho et al.
(2014) proposed the Gated Recurrent Units (GRUs) as a more straightforward
alternative. GRUs also aim to mitigate the vanishing gradient problem but
do so by utilizing a reduced parameter set. This makes them computationally
more efficient while often achieving performance on par with LSTMs. Unlike
LSTMs, GRUs merge the forget and input gates into an update gate and also
combine the cell and hidden states into a single vector. This results in fewer
parameters and a leaner design. In a GRU, there are two primary gates: the
update gate and the reset gate. We can summarize the operation of a GRU as
follows:
zt = σ(W z,h ht−1 + W z,x xt + bi ),
rt = σ(W r,h ht−1 + W r,x xt + bo ),
(43)
h̃t = tanh(W h (rt ⊙ ht−1 ) + W h,x xt ),
ht = (1 − zt ) ⊙ ht−1 + zt ⊙ h̃t,

where zt functions as the update gate, rt serves as the reset gate, h̃t denotes the
candidate hidden state, and ht corresponds to the new hidden state.
Overall, GRUs feature a more streamlined architecture than LSTMs, mak-
ing them less complex to implement and quicker to train. With fewer gates
and combined states, GRUs have fewer parameters, reducing the risk of over-
fitting. GRUs offer an efficient alternative that retains the key advantages of
LSTMs. Understanding the differences and trade-offs between LSTMs and
GRUs allows practitioners to choose the appropriate architecture for their
specific needs.

3.6 Seq2seq and Attention

In previous sections, we focus on a single-point estimation. In other words,
our model can only make predictions for a fixed horizon that is specified
beforehand. If we are interested in predictions at various horizons, several
models need to be fitted, with each requiring independent training. However,

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 43

information flows from past to future for time-series. We could thus expect
that features that are meaningful for short-term predictions could be used for
long-term predictions. Therefore it would be a waste to treat them independ-
ently. Here, we introduce the Sequence-to-Sequence model (Seq2Seq) and the
Attention mechanism that enable us to make multi-horizon forecasts. Both
models have an encoder-decoder structure and we can simultaneously forecast
all horizons of interest.

3.6.1 Sequence to Sequence Learning (Seq2Seq)

In Sutskever, Vinyals, and Le (2014), the Seq2Seq model was proposed as a
significant advancement in the realm of neural networks, especially for NLP
applications. The earliest Seq2Seq framework focused on machine translation,
whereby text is transformed from one language into another. Prior approaches,
such as statistical machine translation, had difficulty handling intricate lan-
guage patterns and preserving natural fluency. The Seq2Seq model, with its
encoder-decoder architecture, provides a more robust framework for handling
such tasks.
A standard Seq2Seq model is comprised of two core components: an encoder
that encodes an input sequence into a fixed-dimensional representation, and
a decoder that leverages this representation to generate an output sequence.
Early Seq2Seq implementations typically employed RNNs for both encoding
and decoding. However, these RNN-based models struggled with longer input
sequences, largely because of issues like vanishing gradients. This led to the
adoption of LSTM networks and GRUs, which provided better handling of
long-term dependencies.
Seq2Seq architectures were also soon applied to financial time-series. Nota-
bly, Z. Zhang and Zohren (2021) introduced an application of Seq2Seq and
Attention models in the context of financial time-series. Consider an input
sequence x1:T = (x1, x2, · · · , xT ) ∈ RT×m , where each xt ∈ Rm is an m-
dimensional feature vector at time t and T is the total length of the sequence.
The encoder processes these vectors step by step to derive meaningful features,
and the resulting context vector captures the relevant information gathered by
the encoder. After that, the decoder utilizes the information from the context
vector and generates the output y1:k = (y1, y2, · · · , yk ) ∈ Rk×n where k is the
furthest prediction point. Specifically, given a single x1:T , we can derive the
hidden state (ht ) with the previous hidden state and current observations (xt ):
Encoder: ht = f (ht−1, xt ), (44)
where f can be a simple RNN model or complex recurrent network. The encoder
iterates over the input sequence until it reaches the final time step, and its last

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 44 Quantitative Finance

Figure 13 A typical example of a Seq2Seq network.

hidden state serves as a summary of the entire input. In Seq2Seq models, this
final hidden state is often taken as the context vector c, functioning as the
“bridge” between the encoder and decoder. For the decoder, the hidden state
ht′ is defined as:

Decoder: ht′ = f(ht−1

′
, yt−1, c), (45)

and the distribution for output yt is:

P(yt |yt−1, yt−2, · · · , y1, c) = g(ht′, c), (46)

where f and g can be various functions but g needs to produce valid prob-
abilities, which could be achieved through a softmax activation function
(Equation 47). Figure 13 shows an example of a standard Seq2Seq network.
exp(zi )
softmax(z)i = ÍK , i = 1, · · · , K. (47)
j exp(zj )

Seq2Seq models have advanced a wide array of tasks in NLP and other
fields. In machine translation, Seq2Seq revolutionized the field by providing
more accurate and fluent translations compared to traditional methods. These
abilities enabled the generation of concise summaries from long documents,
aiding in information extraction and content curation. Further, Seq2Seq mod-
els powered early chatbots and virtual assistants, allowing for context-aware
responses in dialogues.
A primary drawback of traditional Seq2Seq architectures is that they com-
press the entire input sequence into one fixed-dimensional context vector.
For short sequences, this approach works reasonably well. But, for longer
sequences, it becomes problematic as the context vector may not encapsulate all
the relevant information. This can potentially lead to a loss of important details
and degrade the quality of the generated output. Consequently, the decoder
can find it challenging to generate precise and coherent outputs, especially

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 45

for tasks that require retaining detailed information over extended sequences.
These shortcomings led to the subsequent development of the attention
mechanism.

3.6.2 Attention
The attention mechanism, introduced by Bahdanau, Cho, and Bengio (2014),
enables a model to dynamically attend to different parts of an input sequence
instead of relying solely upon a single fixed-size context vector. This approach
leverages a system of alignment scores, attention weights, and context vectors
to provide the model with greater flexibility, thereby improving its ability to
handle longer sequences effectively.
In attention-based models, alignment scores are first calculated to assess the
relevance of each encoder hidden state to the current decoder state, indicat-
ing how each input token influences the token being generated. These scores
are then normalized with a softmax function to yield attention weights, which
dynamically control how much emphasis each input token receives at each
decoding step. Next, a weighted sum of the encoder hidden states is taken
according to these attention weights, resulting in a context vector that highlights
the most pertinent aspects of the input. This context vector is then used by the
decoder to generate the next token in the output sequence.
The attention mechanism also follows an encoder-decoder architecture. We
can denote the encoder’s hidden state at time t by ht :
Encoder: ht = f(ht−1, xt ), (48)
where f is a non-linear function that is similar to a Seq2Seq model. The dif-
ference lies in the decoder structure as we now need to compute attention
weights, alignment scores, and context vectors. Specifically, we define the
context vector ct and attention weights at the time stamp t as:
Õ
T
Context vector: ct = αt,i hi,
i=1
′
(49)
exp(e(ht−1 , hi ))
Attention weight: αt,i = ÍT ′
,
j=1 exp(e(ht−1, hj ))
′
where e(ht−1 , hi ) is the attention score that indicates the weights placed by the
context vector on each time step of the encoder. The work of Luong, Pham, and
Manning (2015) introduces three methods to compute the score:

 ′
hTi ht−1

 dot,
′


′
e(ht−1, hi ) = hTi W a ht−1 (50)
general,



 tanh(W a [hT ; h′ ]) concatenate.
 i t−1

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 46 Quantitative Finance

Figure 14 An example of Attention.

Finally, similar to the process for a Seq2Seq model, the context vector ct is
fed to the decoder:
Decoder: ht′ = f(ht−1
′
, yt−1, ct ),
(51)
P(yt |yt−1, yt−2, · · · , y1, ct ) = g(ht′, ct ),

where ht′ denotes the hidden state at time t and the activation function is denoted
by g. An illustrative example of the Attention mechanism is shown in Figure 14.
In essence, the attention mechanism was conceived to address the drawbacks
of Seq2Seq models – namely, their dependence on a fixed-size context vector
and the ensuing information bottleneck. By enabling the model to selectively
focus on different regions of the input sequence, attention mechanisms sub-
stantially improve the handling of lengthy inputs and the retention of crucial
contextual information.
By granting the decoder access to all the encoder hidden states, rather than
relying on a single fixed-size context vector, the information bottleneck issue is
significantly alleviated. Moreover, attention mechanisms promote better gradi-
ent flow during training, helping to mitigate the vanishing gradient problem in
RNNs and enhancing the model’s capacity to capture long-range dependencies.

3.7 Transformers
The attention mechanism is very powerful as it enables context vectors to incor-
porate information across longer sequences. However, such a model possesses
a chain structure that is very slow to train. This problem worsens as input
lengths increase. To address this issue, the Transformer network was designed
by Vaswani et al. (2017) and represents a major advancement in leveraging
attention mechanisms.
Unlike Seq2Seq models which have a recurrent structure that is slow to train,
the Transformer architecture introduces a parallelizable attention mechanism

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 47

that allows it to process entire sequences simultaneously. Transformers sub-

stantially reduce training and inference times by leveraging parallel processing,
which makes them more efficient when dealing with large datasets. The parallel
efficiency of the Transformer model originates from the self-attention mech-
anism (also referred to as scaled dot-product attention). As the foundational
component of the Transformer architecture, self-attention enables the model to
evaluate the relationships between all tokens within the input sequence at once,
independent of their positional distance. This capability is especially beneficial
for capturing long-range dependencies and makes Transformers highly scalable
and versatile.
Recent architectures have consistently delivered state-of-the-art results and
top-tier performance across a variety of tasks, including language translation,
text summarization, and question-answering. Such models can contain billions
of parameters and handle extremely large-scale datasets. For example, BERT,
introduced by Devlin, Chang, Lee, and Toutanova (2018), was a groundbreak-
ing model in natural language processing that revolutionized the way language
models understand text. OpenAI’s GPT series comprises autoregressive lan-
guage models that have also markedly pushed forward developments in the
field of natural language generation.
In this section, we carefully introduce the Transformer designed by
Vaswani et al. (2017). A solid grasp of the foundational Transformer compo-
nents is crucial, as many cutting-edge models extend from them. A typical
Transformer architecture follows an encoder-decoder design. The encoder
includes input embedding, positional encoding, and an attention mechan-
ism. Specifically, the attention module consists of a stack of N layers, each
containing a multi-head self-attention sub-layer and a position-wise fully con-
nected feed-forward sub-layer. The decoder mirrors this structure with its
own stack of layers – often of the same depth as the encoder – but each
decoder layer incorporates three main sub-components: masked multi-head
self-attention, encoder-decoder attention, and a position-wise feed-forward net-
work. Figure 15 shows various components of transformers and we discuss
each in detail.

3.7.1 Encoder
In traditional machine learning models, data is often represented in raw, high-
dimensional forms. For transformers, this raw data is transformed into dense,
continuous representations known as Input Embeddings. In the context of NLP
applications, the input embedding layer transforms discrete tokens – like words
or subwords – into dense vectors of a predefined dimension dmodel . These vec-
tors capture semantic relationships and contextual meanings of the tokens,

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 48 Quantitative Finance

Figure 15 The Transformer model architecture as first introduced in

Vaswani et al. (2017).

allowing the model to learn complex dependencies within the data. By map-
ping tokens into a continuous space, embeddings facilitate more efficient and
effective learning and processing by the model. For time-series, we can take,
for example, 1-D convolutional layer to carry out the embedding step.
Unlike RNNs or CNNs, transformers do not inherently process data in a
sequential manner. This poses a challenge for capturing the order of inputs
in a sequence. Transformers address this need by employing Positional Encod-
ings, which are combined with the input embeddings to inject positional context
into the model. These encodings are designed to be unique for each position in
the sequence and can be generated using various methods, such as sinusoidal
functions:

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 49

PEpos,2i = sin(pos/100002i/dmodel ),
(52)
PEpos,2i+1 = cos(pos/100002i/dmodel ),
where i is the dimension and pos is the position. This function is used as we
inspect that this simple form allows the model to study the relative position
of inputs. Position encodings ensure that the model can distinguish between
different positions in the sequence, thereby preserving the order and relational
information that is vital for understanding positional information.
Together, input embeddings and position encodings enable transformers to
handle sequential data with high flexibility and efficiency. They transform raw
features into meaningful representations and incorporate positional informa-
tion, allowing transformers to model intricate interconnections and dependen-
cies. This combination is a key factor behind the impressive performance of
transformers across a variety of tasks in NLP problems and beyond.
The core strength of the Transformer is rooted in the attention mechanism,
specifically self-attention, which enables the model to assign varying levels of
significance to different elements of the input sequence when encoding each
token. Leveraging well-established mathematical foundations, this mechanism
effectively manages long-range dependencies. Within the encoder, the Self-
Attention Mechanism allows the model to assign varying degrees of importance
to different segments of the input sequence for each token. The first step in this
process consists of linear projections:
Qi = W Q x i ,
K i = W K xi, (53)
V i = W xi,V

where each token’s embedding is transformed into Query (Q), Key (K), and
Value (V) vectors via learned weight matrices. Here xi represents the internal
representation of a single token for NLP tasks or a single timestamp for time-
series problems.
Attention scores are determined by the dot product of the Query and Key
vectors, scaled by the square root of the Key vector dimension dk to keep the
variance close to 1. These scaled scores are then passed through a softmax
function to produce the attention weights:
 QT K j 
attention weights(xi, xj ) = softmax √i , (54)
dk
where the final representation for each token is computed as a weighted sum
of the Value vectors:
Õ
outputi = attention weights(xi, xj ) · V j . (55)
j

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 50 Quantitative Finance

To capture multiple aspects of token relationships, Transformers use multi-

head attention. Each head performs self-attention with different sets of weight
matrices:

headh = Attention(Qh, K h, V h ), (56)

where outputs from each attention head are concatenated and then passed
through a learned weight matrix:

Multi Head = Concat(head1, · · · , headH )W O, (57)

where H indicates the parallel heads. After the self-attention step, each token
in the sequence is independently processed by a position-wise feed-forward
network, introducing additional nonlinearity and enhancing the transformer’s
capacity to capture complex features beyond what self-attention alone can
achieve. This network is composed of two linear transformations with a ReLU
activation in between. Formally, for each token, this can be represented as:

FFN(x) = max(0, xW 1 + b1 )W 2 + b2, (58)

where W 1 and W 2 are trainable weight matrices, b1 and b2 are learnable biases,
and max(0, ·) represents the ReLU activation function. The FFN is applied inde-
pendently to each position (i.e., each token embedding) and transforms the
embeddings into a different feature space. In a two-layer FFN, the first linear
transformation is often used to expand the dimensionality, while the second
rescales it back to the original size.
Both the self-attention and feed-forward sub-layers incorporate residual con-
nections and layer normalization, which help stabilize training and enhance
overall performance. Layer normalization is a technique used to stabilize train-
ing by normalizing activations within each training example across the features
of a given layer. Through these residual connections, the input to each sub-layer
is added directly to its output, and layer normalization is applied to the sum to
maintain numerical stability and convergence:

LayerNorm(x + Sublayer(x)). (59)

Together, these components enable the Transformer encoder to effectively

process and encode sequences. By leveraging self-attention to capture depen-
dencies and FFNs to learn complex feature mappings, the Transformer architec-
ture establishes a versatile and powerful foundation for numerous applications.

3.7.2 Decoder
The decoder in a Transformer architecture is integral to generating output
sequences for purposes such as machine translation, text generation, and

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 51

sequence-to-sequence tasks. The decoder processes representations from the

encoder and produces sequential output tokens. Its design integrates key
elements such as masked self-attention, multi-head attention, and position-
wise feed-forward networks, all of which are instrumental in enhancing the
decoder’s overall effectiveness.
The decoder’s masked self-attention mechanism enforces that the prediction
at any position in the sequence relies only on the previously observed positions,
thus maintaining the autoregressive property essential for sequence generation.
To achieve this, a mask is applied to the attention weights to block the model
from attending to future tokens. At each position i in the decoder, the attention
scores are calculated as:

score(xi, xj ) = QTi K j, (60)

where Qi and K j are the Query and Key vectors, respectively. To prevent the
leakage of future information, a mask M is applied to the scores:

QT K j
masked score(xi, xj ) = √i + Mi,j, (61)
dk
where Mi,j is −∞ if j > i, ensuring that the softmax function will yield zero
weights for future tokens:

attention weights(xi, xj ) = softmax(masked score(xi, xj )), (62)

where the output for each token xi is computed as:

Õ
outputi = attention weights(xi, xj ) · V j . (63)
j

In the decoder, multi-head attention helps the model capture various aspects
of the relationships between the decoder’s tokens and the encoder’s output.
Through cross-attention, the decoder focuses on the encoder’s output. For each
head h, the cross-attention mechanism computes:

headh = Attention(Qhdec, K henc, V henc ), (64)

where Qhdec are the Query vectors from the decoder, and K henc and V henc are the
Key and Value vectors from the encoder. We can concatenate the outputs from
each head and transform them as:

multi head output = Concat(head1, · · · , headh )W o . (65)

Each decoder position is individually passed through a position-wise feed-

forward network to improve its representational ability. This network is

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 52 Quantitative Finance

generally comprised of two linear layers with a ReLU activation in between.

Formally, for each token xi , the transformation is:

FFN(xi ) = max(0, xi W 1 + b1 )W 2 + b2, (66)

where the network enables the decoder to capture complex feature interactions.
Residual connections are applied around each sub-layer (self-attention, cross-
attention, and feed-forward network), followed by layer normalization. For a
given sub-layer output, the layer normalization is:

LayerNorm(x + SubLayer(x)). (67)

By integrating masked self-attention, multi-head attention, and position-

wise feed-forward networks, the decoder’s architecture empowers it to produce
coherent and contextually appropriate output sequences. By attending to past
and present tokens and incorporating information from the encoder, the decoder
can handle long-range input sequences. The final outputs are generated from
the decoder autoregressively, meaning that the model relies upon the encoded
input sequence and previously generated tokens to produce each subsequent
token.

3.7.3 Transformers-Based Time-Series Models

The usage of transformers for time-series analysis has attracted great popular-
ity. Because time-series problems come with their own unique challenges, such
as temporal dependencies, irregular sampling, and varying sequence lengths,
there has been a surge of research exploring how to tailor transformers spe-
cifically for these use cases (see Y. Wang et al. (2024) for a recent review).
Broadly, one can categorize transformers designed for time-series tasks into
two main groups: one based on the application domain, and another based on
the underlying network architecture. Figure 16 shows the groupings of these
models and their respective sub-fields.
From an application-domain perspective, time-series transformers are typi-
cally tailored toward four major tasks: forecasting, imputation, classification,
and anomaly detection. Our main focus here is on forecasting, but we briefly
mention all tasks for completeness. Forecasting involves predicting future val-
ues based on historical data patterns. Transformers excel here by capturing
long-range temporal dependencies that classical models might miss. Imputa-
tion deals with filling in missing or corrupted data points. The self-attention
mechanism helps the model learn relationships across different time steps
to recover lost information. For the classification of sequences, transformers
can leverage learned feature embeddings to distinguish between categories
or labels associated with entire sequences or specific time intervals. Finally,

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 53

Figure 16 Groups of time-series transformers based on application domains

and attention modules.

anomaly detection involves identifying unusual patterns or outliers in the data.

Here, transformers can highlight subtle deviations from normal behavior by
comparing attention-weighted signals across multiple timestamps.
We can also group models in terms of how attention is employed. Here
we distinguish between point-wise, patch-wise, and series-wise. Point-wise
attention treats each time step as an individual token, learning direct pairwise
relationships across all time steps. This granular approach can capture intricate
local patterns, though it might become computationally heavy for very long
sequences. Patch-wise attention groups consecutive time steps into patches or
segments, reducing the overall sequence length before applying attention. This
strategy trades some resolution for improved efficiency and can still preserve
local correlations within each patch. Finally, series-wise attention considers the
entire sequence as a single unit, using more global operations to learn high-level
representations of the data. While this can be extremely efficient, it risks los-
ing some of the fine-grained temporal detail that is critical to many time-series
applications. Figure 17 illustrates these attention modules.
We list each attention module with one representative work. For point-wise
attention, the vanilla transformer and the Informer designed by H. Zhou et al.
(2021) are illustrative examples of earlier transformer models for time-series. A
good example for patch-wise attention is PatchTST designed by Nie, Nguyen,
Sinthong, and Kalagnanam (2022). The series-wise attention can be found in
the recently introduced iTransformer (Y. Liu et al. (2023)). There is a trade-off
between the benefits from moving from point-wise to sequence-wise attention.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 54 Quantitative Finance

Figure 17 Categorization based on attention modules: point-wise,

patch-wise, and series-wise. Examples of models that employ each of those
attention types respectively are the Informer, PatchTST, and iTransformer.

Patch-wise strikes a good balance in this trade-off which has led PatchTST to
outperform many other architectures in forecasting benchmarks.
Besides, there are many other models that apply one of these atten-
tion modules, including the Autoformer (H. Wu et al., 2021), Crossformer
(Y. Zhang & Yan, 2023), and others. Readers might also find some of the earlier
works in this area useful. These are widely covered in an earlier survey paper
by Lim and Zohren (2021). Note that there are still interesting modules that we
have not covered here. For example, the work of Lim, Arık, Loeff, and Pfister
(2021) on the Temporal Fusion Transformer (TFT) designed a transformer
architecture specifically for multi-horizon forecasting, combining the strengths
of transformers with recurrent layers to handle both static and time-varying
features.
Overall, we think that it is important to recognize the scope and prog-
ress that has been made on the development of transformers for time-series
applications. It has become clearer which Transformers are best suited for
specific time-series challenges, whether that involves capturing nuanced local
trends for anomaly detection or learning broad seasonal patterns for long-
term forecasting. The interplay between the nature of the data and the chosen
architecture continues to shape ongoing innovations in transformer-based time-
series modeling. This will continue to pave the way for increasingly accurate
and robust models. For interested readers, the aforementioned recent review
paper (Y. Wang et al., 2024) is a good place to start any further reading.

3.8 Graph Neural Networks and Large Language Models

In this section, we introduce some recent developments that have gained great
popularity. Note that these materials are more advanced and we include them to
demonstrate some directions of future development for applying deep learning
models to quantitative finance. In this section, we discuss the intuition of the
usage of these methods and introduce various promising resulting applications.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 55

3.8.1 Graph Neural Networks

In the realm of machine learning, the advent of graph neural networks (GNNs)
has marked a significant evolution in our ability to process, analyze, and derive
insights from data that can be modeled by graphs. Graph representations,
with nodes that represent entities and edges that represent their relationships,
pervade numerous domains including social networks, molecular chemistry,
transportation systems, and communication networks. Traditional neural net-
work models, despite their prowess, fall short when it comes to capturing the
dependencies and relational information inherent in graph data. GNNs are a
groundbreaking class of neural networks engineered to explicitly handle graph
structures and have led to a leap forward in areas such as node classification,
link prediction, and graph classification.
In finance, we can often naturally represent the interactions among enti-
ties (such as individuals, institutions, and assets) as graphs. GNNs provide a
framework to process such graph-structured data. Research on the application
of GNNs in quantitative finance has been active in the past few years. The
works of Pu, Roberts, Dong, and Zohren (2023); C. Zhang, Pu, Cucuringu, and
Dong (2023) adopt GNNs to build momentum strategies and to forecast mul-
tivariate realized volatility. Soleymani and Paquet (2021); Sun, Wei, and Yang
(2024) combine GNNs with reinforcement learning to tackle the problem of
portfolio construction. For a nice review on GNNs in various financial applica-
tions, interested readers are pointed to the work of J. Wang, Zhang, Xiao, and
Song (2021). Here, we introduce the basics of networks and, in particular, we
describe the most prevalent GNN model, graph convolutional neural networks
(GCNs).

Basics of Networks and Graphs The core strength of GNNs stems from their
capacity to learn representations of nodes (or entire graphs) that encapsulate not
only their features but also the rich context provided by their connections. Such
operations are achieved through mechanisms like message passing, aggregating
information across neighboring nodes, and iteratively refining their represen-
tations. This process allows GNNs to capture both local structures and global
graph topology, offering a nuanced understanding of graph-structured data.
Before delving into GNNs, we need to understand the basics of networks and
graphs. To start, we define a graph G as:
G = (V, E), (68)
where V = {v1, · · · , vn } denotes the set of n nodes and E represents the set
of edges. An edge eij = (vi, vj ) ∈ E indicates a connection between nodes vi
and vj .
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 56 Quantitative Finance

Nodes (also called vertices) represent the entities or objects in a graph. In

different contexts, a node could represent a computer in a network, a person in
a social network, a city in a transportation map, or a neuron in a neural network.
Edges (also called links) represent the connections or relationships between
these nodes. Edges can be undirected, indicating a bidirectional relationship,
or directed, indicating a one-way relationship (these form a directed graph or
digraph). Edges may also have weights, which quantify the strength or capacity
of the connection, such as the distance between cities, bandwidth in a network,
or the strength of a social tie.
In order to describe a graph, we need a way to represent nodes and edges in
a compact form. This is done in the form of an adjacency matrix. An adjacency
matrix A is a n × n matrix, where Aij indicates the connectivity status between
node vi and vj . Depending on the nature of the problem, there are many types
of graphs:

• Undirected graphs: These are graphs with edges that lack direction, mean-
ing each connection between two nodes is inherently bidirectional.
• Directed Graphs (Digraphs): Graphs in which edges carry a direction, rep-
resenting a one-way relationship from one node to another. For example,
eij = (vi, vj ) ∈ E denotes an edge pointing from node vi to node vj .
• Bipartite Graphs: This is a distinct type of graph in which nodes are divided
into two separate groups, and every edge connects a node from one group
to a node in the other group, with no edges existing within the same group.
• Homogeneous graphs: Graphs where all nodes and edges are of a single
type.
• Heterogeneous Graphs: Graphs that contain multiple types of nodes and/or
edges. For example, we can denote a graph as G = (V, E, t : V → A, τ :
E → R), where each node vi ∈ V is assigned a type ai ∈ A by function t
and each edge eij ∈ E is assigned a type rij ∈ R by function τ.
• Dynamic graph: A dynamic graph is defined as a sequence of graphs G seq =
{G1, · · · , GT }, where each Gi = (Vi, Ei ) for i = 1, · · · , T. In this sequence,
Vi , Ei represent the sets of nodes and edges for the i-th graph, respectively.

Graph Convolutional Neural Networks In the work of Z. Wu et al. (2020),

GNNs are classified into four main categories: convolutional graph neural
networks, graph auto-encoders, recurrent graph neural networks, and spatial-
temporal graph neural networks. In this section, we introduce graph convolu-
tional neural networks (GCNs) which have become the most widely adopted
and extensively utilized GNN models.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 57

As previously mentioned, CNNs have seen tremendous success in handling

data with established grid-like structures, such as images. The fundamental
principle of CNNs lies in the convolution operation, which entails moving a
filter (or kernel) across the input data (e.g., an image) to generate a feature
map. This feature map highlights the presence of particular features or patterns
at various positions within the input. This process is inherently suited to data
with a regular, grid-like structure where the relative positioning of data points
(e.g., pixels in images) is consistent and meaningful.
GCNs (Kipf & Welling, 2016) extend the concept of convolution to graph-
structured data, where the data points (nodes) are connected by edges in a non-
Euclidean domain. Unlike the regular, grid-like topology of images, graphs are
irregular, and the number of neighbors for each node can be different. GCNs
address this by defining convolution in terms of feature aggregation from a
node’s neighbors, allowing them to capture the structural information of the
graph. The fundamental operation in a GCN is this aggregation of features from
a node’s neighbors, and it can be mathematically represented as:

Hl+1 = σ(D̂− 2 ÂD̂− 2 H(l) W(l) ),

1 1
(69)
where the matrix of node features at layer l is denoted as H(l) and H(0) is the input
feature matrix X. We write  = A + IN as the addition between the adjacency
matrix A and identity matrix IN , allowing nodes to consider their own features
Í
in aggregation. D̂ is the degree matrix of Â, where D̂ii = j Âij , and W(l) is the
weight matrix for layer l. Here σ denotes a nonlinear activation function, such
as ReLU.
Simply put, much like traditional CNNs are constructed from convolu-
tional layers, GCNs are built by stacking multiple graph convolutional layers.
Each graph convolutional layer receives the node vectors from the previous
layer (or the initial input feature vectors for the first layer) and generates
new output vectors for each node. To illustrate this process, Figure 18 depicts
how a graph convolutional layer aggregates the vectors from each node’s
neighbors.
In Figure 18, the vector for node A, labeled xA , is combined with the vectors
of its neighboring nodes, xB and xC . This combined vector is then transformed
or updated to produce node A’s vector in the next layer, denoted as hA . This
process is uniformly applied to every node in the graph. This technique is
commonly referred to as message passing, where each node “passes” its vec-
tor to its neighbors to facilitate the updating of their vectors. The “message”
from each node is its associated vector. The specific rules for aggregation and
updating are detailed in Equation 69.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 58 Quantitative Finance

Figure 18 A graph convolution layer that pools information for node A from
its neighbors.

Once we stack several graph convolutional layers, we form a typical GCN

as shown in Figure 19. The output of a GCN depends on the problem at hand.
Graph prediction tasks are commonly classified into three categories: graph-
level, node-level, and edge-level. When dealing with a node-level task, such
as classifying individual nodes, the vectors generated for each node can serve
as the final outputs of the model. In the case of node classification, these out-
put vectors may represent the probabilities that each node belongs to specific
classes. This is illustrated in the top section of Figure 20.
Alternatively, we might focus on a “graph-level” task, where the objective
is to generate a single output for the entire graph rather than producing out-
puts for each individual node. For example, the goal could be to classify entire
graphs instead of classifying each node separately. In this case, the vectors from
all nodes are collectively input into another neural network (such as a simple
multilayer perceptron) that processes them together to produce a single output
vector. This is illustrated in the bottom part of Figure 20.
Edge-level tasks focus on predicting properties or attributes related to the
edges in a graph. These tasks are important for understanding and interpret-
ing the relationships between entities represented by the nodes in a graph.
For example, we can predict whether a link (edge) should exist between two
nodes, even if it is not present in the observed data. This task is fundamental in

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 59

Figure 19 A GCN that consists of multilayers.

Figure 20 Top: a “node-level” prediction task; Bottom: a “graph-level”

prediction task.
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 60 Quantitative Finance

various applications, such as recommending friends in social networks, predict-

ing interactions between proteins in biological networks, or inferring missing
connections in knowledge graphs.

3.8.2 Large Language Models and Generative AI

Recent developments of the GPT-3 and GPT-4 models behind ChatGPT, a

breakthrough generative language model, stand as a pinnacle of innovation in
the domain of large language models (LLMs), offering a compelling glimpse
into the future of human–computer interaction. ChatGPT, along with other
models such as Bard and Claude, without a doubt, has revolutionized not only
our daily lives but also the way we work, ushering in a transformative period
for human-computer interaction.
By leveraging the power of large language models, ChatGPT has opened
up unprecedented possibilities in the financial sector and has inspired a wave
of research and development focused on applying LLMs to tackle complex
challenges in finance. Some examples include automating customer service,
market analysis, fraud detection, and more. We next introduce you to the evo-
lution of large language models and discuss the rationale behind the strength of
LLMs. However, the development and cointegration of LLMs and quantitative
finance are still in the early stages. We also present some potential limitations
of applying current state-of-the-art LLMs to quantitative finance and potential
future directions of work.

Evolution of LLMs The evolution of language models has been marked by

significant milestones, primarily advancements in neural network design and
learning methodologies. Starting with RNNs, the journey to develop models
like BERT and the GPT series showcases a remarkable trajectory of innovation,
with each leap strengthening the models’ capabilities to capture and understand
language at scale.
RNNs were among the first neural architectures used to handle sequential
data, such as text. Their design allows information to persist through the net-
work’s hidden states, theoretically enabling them to remember long sequences
of inputs. Nonetheless, in practical applications, RNNs encountered challenges
with maintaining long-term dependencies due to problems like vanishing and
exploding gradients. This made it challenging to capture context over large
spans of text. LSTMs and GRUs, variants of the basic RNN, introduced gat-
ing mechanisms to better control the flow of information. This addresses the
issue of long-term dependencies to a significant extent. These improvements
allowed for more effective learning from longer sequences, leading to better
performance on a wide range of NLP tasks.
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 61

Subsequently, BERT introduced an innovative method by pre-training a

deeply bidirectional model that simultaneously incorporated both left and right
contexts across all layers. This approach marked a significant shift from earlier
models, which typically processed text in only one direction. Leveraging the
Transformer architecture’s attention mechanism, BERT effectively understands
a word’s context by taking into account its entire surrounding environment.
This led to notable improvements in tasks like question answering and lan-
guage inference. Finally, GPT series, starting with GPT-1 and extending to
GPT-3 and beyond, emphasized generative pre-training of transformer-based
models (only using the decoder part) on a diverse corpus of text, followed by
fine-tuning on specific tasks. GPT models demonstrated remarkable text gener-
ation capabilities, understanding and generating human-like text across various
genres and styles. Their scalable architecture enabled them to learn from vast
datasets, capturing deep linguistic patterns.
The evolution from RNNs to sophisticated models like BERT and GPT
demonstrates a quantum leap in the field of NLP. Each step in this journey
introduced innovations that significantly expanded the capabilities of language
models, moving from basic text processing to understanding context, nuance,
and even generating coherent and contextually relevant text. These advance-
ments have not only pushed the boundaries of what is possible with machine
understanding of language but have also opened up new avenues for human-
computer interaction, making machines better conversationalists, writers, and
analysts. The future of language models promises even greater integration into
daily technology use, making it difficult to distinguish between content created
by humans and that generated by machines

What Made LLMs So Powerful? LLMs have become incredibly power-

ful due to a combination of factors that include algorithmic developments,
advancements in computational capabilities, and access to extensive amounts
of textual data for training. Following are the primary factors that enhance the
effectiveness of large language models.
LLMs are trained on extensive corpora that encompass a wide range of
human knowledge and language use, from literature and websites to scientific
articles and social media content. This broad coverage enables the models to
learn a diverse set of language patterns, idioms, and domain-specific knowl-
edge. The sheer volume of data ensures that the model encounters numerous
examples of language use, facilitating the learning of complex linguistic struc-
tures and nuances. In addition, LLMs are often pre-trained on a general corpus
and then adapted for particular applications using smaller, specialized datasets.
This approach, called fine-tuning, allows the models to apply their broad

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 62 Quantitative Finance

understanding of language to particular domains or applications, significantly

enhancing effectiveness in activities such as text classification, responding to
questions, and generating written content.
The creation of advanced neural network frameworks, especially the
Transformer architecture, has been pivotal. Transformers utilize self-attention
mechanisms to handle data sequences. This allows the model to assess the
significance of various words within a sentence or document. This ability to
understand context and relationships between words significantly enhances
the model’s understanding of language. Specifically, the adoption of attention
mechanisms allows LLMs to focus on relevant parts of the input data when
making predictions or generating text. This capability allows the model to study
the broader context of a word or phrase and lead to more accurate and coherent
outputs.
Training large language models demands considerable computational power,
often utilizing clusters of GPUs or TPUs for periods that extend from weeks
to several months. Advances in hardware and the availability of cloud com-
puting resources have made it feasible to train models with billions or even
trillions of parameters. The scale of these models allows them to capture a vast
range of linguistic patterns and knowledge, contributing to their effectiveness.
Furthermore, the iterative development of LLMs, in which each new version
builds upon the learnings and feedback from previous iterations, has stead-
ily improved their performance. Additionally, the engagement of the research
community and industry in developing, testing, and deploying these models
has led to rapid advancements and innovative applications.
In summary, the power of large language models lies not just in their size but
in the convergence of these technological and methodological advancements.
They represent a synthesis of data, computational resources, and cutting-edge
algorithms, resulting in tools that are able to comprehend and create text that
closely mimics human language with exceptional skill.

LLMs for Time-Series Forecasting and Quantitative Finance LLMs have

already sparked significant interest in quantitative finance and time-series anal-
ysis in general. Although LLMs models are trained largely on textual corpora,
researchers have begun adapting them for forecasting problems by “translat-
ing” numerical or temporal patterns into a format that LLMs can process. The
core idea is that LLMs have learned powerful sequence-modeling capabilities,
which can be harnessed beyond natural language. By carefully encoding time-
series data as a pseudo-text input, an LLM can in principle capture long-range
dependencies, temporal structures, and contextual nuances in much the same
way it understands linguistic patterns (X. Zhang et al., 2024).

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 63

One notable approach along these lines is Time-LLM (Jin et al., 2023). This
method cleverly reprograms a large language model to treat time-series obser-
vations as tokens in a sequence. Specifically, the time-indexed data points are
formatted into a textual prompt in which the LLM is asked to “complete” the
sequence effectively performing a forecast. Despite being originally designed
for language tasks, this work provides an example of how the LLM’s internal
attention mechanisms and capacity for pattern recognition can be extended to
temporal prediction. Time-LLM has shown promising results on a variety of
benchmarks, demonstrating that large language models can be repurposed for
time-series forecasting with relatively minimal changes to their architecture.
By leveraging training on massive text corpora, Time-LLM highlights a new
direction for cross-domain learning, where the underlying skills of an LLM are
refocused on numerical patterns and trends over time.
Despite their advanced capabilities, the usage of LLMs in quantitative
finance is still in the early stage. In the second part of this Element, we
will discuss how LLMs can be used for volatility forecasting and portfolio
optimization. In this section, we discuss some limitations that LLMs face when
applied to the domain of quantitative finance. These limitations stem from the
unique challenges and requirements of the financial sector, including the need
for precise numerical analysis, real-time decision-making, and understanding
of complex financial instruments and markets.
LLMs excel at processing and generating text but often struggle with under-
standing and manipulating numerical data to the extent required in quantitative
finance. Financial analysis often involves complex mathematical models and
statistical methods that are beyond the current capabilities of language-based
models. Integrating LLMs with specialized numerical processing systems
remains a challenge. Furthermore, the financial markets are dynamic, with
conditions that change rapidly. LLMs trained on historical data may not
adapt quickly enough to real-time data or sudden market shifts. The latency
in processing new information and updating models can be a limitation in
time-sensitive financial applications.
On the one hand, LLMs can be fine-tuned with financial texts to understand
domain-specific language. However, truly grasping the intricacies of financial
instruments, regulatory environments, and market mechanisms requires a level
of expertise that LLMs may not achieve solely through language training. This
gap can lead to inaccuracies or oversimplified analyses when processing com-
plex financial scenarios. On the other hand, there is a continual concern with
respect to overfitting, a scenario where a model excels on its training dataset but
fails to perform well with new, unseen data as future market conditions can dif-
fer significantly from historical patterns. Ensuring that LLMs generalize well to

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 64 Quantitative Finance

new, unseen market conditions without overfitting to past data remains a chal-
lenge. Also, most existing LLMs are trained up-to-date, so they can not be used
for historical backtests because of the look-ahead bias due to the information
leakage problem.
While large language models hold strong potential for revolutionizing many
aspects of quantitative finance, addressing these limitations is important for
their effective and responsible application. Ongoing research and development
efforts are focused on overcoming these challenges and show promise for the
improvement of the capabilities of LLMs in financial analysis, prediction, and
decision-making.

3.8.3 Other Recent Developments: State-Space Models and xLSTM

State-space models provide a framework for modeling dynamic systems by
representing a system’s evolution over time with a set of latent variables.
Also, state-space models are the underlying mathematical framework that
the Kalman filter (Kalman, 1960) operates on. These models are also very
popular in time-series analysis because they can seamlessly incorporate var-
ious sources of uncertainty and are adaptable to complex systems. Recent
work, Mamba (Gu & Dao, 2023) marks a noteworthy breakthrough in the area
of sequence modeling, particularly for time-series data. Traditional sequence
models often face challenges because of their computational complexity and
capturing long-range dependencies in data. Mamba addresses these issues by
leveraging selective state spaces to model sequences efficiently.
The core of Mamba is the concept of selective state spaces, which enables
the model to hone in on the sequence’s most critical elements while discarding
extraneous or less important information. This selective attention mechanism
is key to Mamba’s ability to operate in linear time, a crucial feature for handling
large-scale time-series data where computational efficiency is paramount. By
narrowing the state space to only the most important components, Mamba can
maintain high accuracy in sequence predictions while significantly reducing
the computational overhead.
The linear-time complexity of Mamba is particularly beneficial for real-
world applications where speed and scalability are critical. For instance, in
financial markets, where vast amounts of noisy high-frequency data need to be
processed in real-time, Mamba’s approach allows for rapid and accurate mod-
eling of sequences without sacrificing performance. Additionally, the model’s
ability to selectively study relevant states makes it robust to noise and capable
of adjusting to a wide range of time-series data types, from economic indicators
to sensor readings in IoT devices.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 65

Another recent work proposed by Beck et al. (2024) designed xLSTM which
expands upon a traditional LSTM network to tackle certain inherent short-
comings of standard recurrent networks while enhancing their capabilities for
complex sequence modeling tasks. While LSTMs have shown to be very effec-
tive in sequential modeling, they can sometimes struggle with certain types
of data patterns, especially when tackling very long sequences or when the
relationships between data points are highly nonlinear and intricate.
One of the key innovations in xLSTM is the ability to dynamically adjust
its memory and learning mechanisms based on the complexity and nature of
the data it encounters. Traditional LSTMs use fixed gates for controlling the
flow of information, which can be limiting when faced with varying data char-
acteristics. In contrast, xLSTM introduces adaptive mechanisms that allow the
network to modulate its memory retention and forgetfulness more effectively.
This adaptability enables xLSTM to maintain a high level of performance even
when dealing with sequences that have non-stationary patterns or when the rele-
vant information spans a wide range of time steps. By extending the core LSTM
architecture, xLSTM is better equipped to capture complex dependencies that
might be missed by more rigid models.
The introduction of xLSTM is a significant breakthrough in the ongo-
ing development of neural network architectures for sequence modeling.
Kong, Wang, et al. (2024) builds on xLSTM to particularly model multivariate
time-series. They improve and revise the memory storage of xLSTM to fit with
time-series analysis and adopt patching techniques to ensure that long-term
dependencies can be studied.

4 The Model Training Workflow

Having discussed basic descriptive analyses of financial time-series as well
as supervised learning frameworks in the context of financial applications, we
now present a comprehensive pipeline for the model training workflow. Over-
all, developing a quantitative trading strategy with deep networks requires a
systematic approach to properly evaluate model performance and adjust model
configurations. Ensuring that this procedure is transparent and replicable is cru-
cial for successful deployment. This section covers common frameworks to
design and train networks in various settings.
Essentially, we can divide the whole process into six parts: problem setup,
data collection and cleaning, feature extraction, model construction, cross-
validation and hyperparameter tuning, and final deployment. This framework,
as illustrated in Figure 21, outlines the essential steps of formulating, training,
tuning, and evaluating model performance in a systematic way. We now briefly

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 66 Quantitative Finance

Figure 21 Key steps of the model training workflow.

introduce each of these steps and discuss cross-validation and hyperparameter

tuning specific to financial time-series in detail.

4.1 Problem Setup

The starting point for any application is to scope out the ultimate objective and
clearly define the stages of the work process. If, for example, our goal is to
generate a predictive signal, we need to consider various aspects of the desired
signals, including frequency, asset type, and turnover. Even if it is difficult to
build an exhaustive list, it is always better to consider these points beforehand.
In this section, we provide a prediction task as an example and introduce the
stages of generating a signal for trading. Note that the introduced workflow
is not tied only to prediction problems but can also be used as a framework
for many other applications. In particular, the sections on cross-validation and
hyperparameter tuning can be applied across problems.
In the previous section, we explored how supervised learning can be grouped
into two main categories: regression and classification. Depending on the
desired outputs, the task may be framed as a regression problem aiming to
predict returns or as a classification problem placing stocks into performance
categories such as return quantiles. Knowing the target format helps us to
choose objective functions, features, and proper evaluation metrics.

4.2 Data Collection and Cleaning

After defining our objective, we need to choose an appropriate dataset and carry
out cleaning processes to make sure the dataset represents the application of
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 67

our interest. We have introduced several methods to source market price data in
Appendix B and, importantly, we need to choose the frequency of interest and
price formats. High-frequency microstructure market data or down-sampled
price and volume data are two possible examples specific to quantitative trad-
ing. Different formats might influence network architectures and change the
amount of training data required.
Beyond obtaining the right dataset, data preparation is a vital step and might
affect our model performance in unexpected ways if it is carried out poorly.
Missing data is one of the common problems that we might encounter when
dealing with time-series. Hence, we need be extremely careful to make sure
that there is no leakage of future information (also known as a look-ahead bias)
when choosing to impute these missing values. Having access to future infor-
mation might erroneously boost training performance but will lead to very poor
out-of-sample results.
Additionally, it is important to store data in a format that permits swift
exploration and iteration. Beyond databases, popular choices are pickle, HDF,
or Parquet formats – each with its own advantages and disadvantages. For
data exceeding available memory or requiring distributed processing across
multiple machines, parallel computing can also be employed.

4.3 Feature Extraction

One of the primary advantages of employing neural networks is the ability to
automate feature extraction. However, a model still might not be able to gener-
alize well on out-of-sample data if we feed networks with excessive irrelevant
information, especially when the signal-to-noise ratio is very low as is typical
in financial applications. We should first get a sense of data that we are work-
ing with and understand the relationship between targets and variables. This
can help us choose the most appropriate algorithm and carry out transforma-
tions as needed. As introduced in Section 2, we can use visualizations such as
histograms or QQ-plots to examine our data.
As an example, we aim to predict next-day trading volumes Vt+1 . When
looking at a histogram we observe several issues. Firstly, trading volumes are
always positive. To deal with this, one might choose to model the logarithm
log Vt+1 . Furthermore, when plotting this over time we might observe non-
stationarities. It is thus advisable to model the next day’s volume normalized
by a trailing measure of volume, such as the 20-day median volume at day t,
Ṽt . We would thus choose to build a model to predict log(Vt+1 /Ṽt ) and choose
similar normalizations for input features that refer to past volumes, such as
log(Vt−d /Ṽt ) for d = 0, 1, 2, ....
We can also calculate numerical metrics such as correlation and nonlin-
ear statistics such as the Spearman rank correlation coefficient. A systematic
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 68 Quantitative Finance

and thorough exploratory data analysis is the basis of building a successful

predictive signal. Information-theoretic approaches, including mutual informa-
tion (Vergara & Estévez, 2014), can also be used to better understand variable
relationships. After exploring the data, we can begin to design and properly
normalize features from a meaningful set that could boost model performance,
speed up the training process, and help with convergence.
Nevertheless, feature engineering is a complex process that draws on domain
expertise, statistical and information theory principles, and creative insights.
It involves clever data transformations aimed at uncovering the systematic
links between input and output variables. Practitioners can employ numerous
approaches, such as outlier detection and remediation, functional transforma-
tions, and integrating multiple variables. We can also even leverage unsuper-
vised learning. Although the focus of this Element is not on feature engineering,
we emphasize that this plays a central role in building quantitative trading
strategies, and in practice, it is sometimes learned through trial and error.

4.4 Model Construction

We have introduced a wide range of neural network architectures, ranging from
canonical examples such as multilayer perceptrons and CNNs to state-of-the-
art transformer-based architectures for time-series. In general, neural networks
are flexible function approximators that require few assumptions about data
distributions. However, they often need a large dataset to calibrate model
weights to successfully model the relationships between inputs and targets.
When constructing a network, one of the most important factors to consider is
the bias-variance trade-off. Our goal is to evaluate and adjust the model’s com-
plexity using estimates of its generalization error. In order to properly tune the
model to obtain decent out-of-sample performance, we need be aware of how
the bias-variance trade-off relates to under and overfitting. In general, we can
break down prediction errors into reducible and irreducible parts. The irreduc-
ible part is due to random variation (noise) in the data such as natural variation
or measurement errors. This type of error is out of our control and cannot be
reduced by model choice.
The reducible portion of generalization error can be divided into bias and
variance errors. Both types arise from differences between the true functional
relationship and the model’s approximation. If a model is too simple to capture
a dataset’s complex structure, we might get poor results due to the model’s
inability to capture the complexity of the true functional form. This type of error
is called bias. For example, if a true relationship is quadratic, but our model is
linear, even an infinite amount of data would not be enough to recover the true
relationship. This is exactly the bias part of the bias-variance trade-off.
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 69

Figure 22 A visual example of under and overfitting with polynomials.

On the contrary, if a model is too complex, it shows superior performance

on training data, but might end up overfitting as it starts extracting information
from the noise instead of learning true patterns in the data. As a result, it learns
idiosyncrasies from training data which likely would not be found in the testing
set, and consequently, the out-of-sample predictions will vary widely. This is
the variance part of the bias-variance trade-off.
Figure 22 illustrates the concepts of under and overfitting. We assess the
in-sample errors when approximating a sine function using polynomials of
increasing complexity. Specifically, we generate thirty random samples with
added noise and fit polynomials of varying degrees to these data points. The
model then makes predictions on new data, and we record the mean-squared
error for these forecasts. In the left panel of Figure 22, a first-degree polyno-
mial is fitted to the data, clearly demonstrating that a straight line does not
adequately capture the true function. However, the estimated lines remain rela-
tively consistent across different samples drawn from the underlying function.
It thus has high bias and low variance. The right panel shows a polynomial of
degree 15 fitted to the same data. It closely matches the small sample data but
fails to accurately estimate the true relationship because it has overfitted upon
the random variations in the sample points. As a result, the learned function is
highly sensitive to the specific sample, exhibiting low bias and high variance.
The middle panel illustrates that a fifth-degree polynomial provides a reasona-
bly accurate approximation of the true relationship within the interval. It is the
Goldilocks example which is just right. It simultaneously has a variance that is
only slightly higher than the model on the left and a bias that is only slightly
higher than the model on the right, so that the sum of the two yields the lowest
generalization error.

4.5 Cross Validation

Once we train multiple models, we need to compare them and choose the most
appropriate one. Recall that the ultimate goal for any supervised learning algo-
rithm is to make good predictions on testing data, and that this requires models
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 70 Quantitative Finance

to generalize performance from the training set to unseen instances. In order to

fulfill this goal, we often split data into three sets: training, validation, and test-
ing sets. Model weights are first calibrated on the training data. Then we can
take a subset of training data (not used during the training process) to form
the validation set which evaluates model performance so that we can com-
pare different algorithms and select the best model architecture based on the
bias-variance trade-off.
The reason for using the validation set in addition to the test set is to pre-
serve the test set and not touch it until the final evaluation. Otherwise, we could
artificially boost model performance. This occurs because, each time we use
our test set for evaluating a model, we are effectively learning from that test
set. The more frequently we do this, the more the model learns from the test
set and is corrupted. This type of information leakage is especially detrimen-
tal for financial time-series as we are attempting to model causal relationships.
Besides comparing different network architectures on a fixed validation set, we
often resort to systematic cross-validation to perform hyperparameter optimi-
zation. Deep networks are sensitive to many hyperparameters, for example, the
number of neurons in a layer, learning rate and batch size. We will discuss the
exact techniques for choosing hyperparameters in the next section.
In general, K-fold cross-validation is a standard technique used for tuning
hyperparameters. However, cross-validation for time-series is nontrivial. The
first and most important difference for time-series is that we can not randomly
assign samples to either training or validation set because, if so, we might end
up training with future information. In other words, temporal dependency exists
between observations and we must be sure not to include this during the train-
ing process. Otherwise, we could obtain a seemingly “superior” model that in
reality has poor generalization ability. To understand this point better, let us
give an example. Imagine we had intraday price data at either tick or 1 second
frequency and our aim was to make predictions of the future price move over
the next 5 minutes, that is, 300 seconds. The sample which includes the future
return from t to t + 300, and the sample one second later, which includes the
return from t + 1 to t + 301, are highly correlated if not virtually the same.
Randomly shuffling the data could end up placing one of these samples in the
training data and one in the test set. The effect of this is similar to testing on
in-sample data because of the high correlation. In order to solve the aforemen-
tioned limitation, we utilize hold-out cross-validation in which samples are
chronologically fed into validation sets after being used for training. Specif-
ically, we can start cross-validation on a rolling basis. Figure 24 shows this
process where the validation set comes chronologically after the training sub-
set. Note that it is not necessary to gradually increase the training set. We can

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 71

Figure 23 Left: traditional U-shaped overfitting curve; Right: double

descent error curve.

implement this rolling forward cross-validation by Listing 1 in Appendix D and

more information can found in our GitHub repository.8
Recently, there has been research on a concept called Double Descent which
is a phenomenon observed in over-parameterized deep networks where the test
error curve exhibits two distinct “descent” phases. Double descent is a phe-
nomenon observed in modern, over-parameterized deep learning models where
the test error curve, expressed as a function of model complexity or training
time, exhibits not just one but two distinct “descent” phases. Traditionally, one
might expect the test error to reach a minimum at some intermediate model
complexity and then increase (due to overfitting) as complexity grows. How-
ever, with double descent, after initially declining and then peaking around
the point where the model just fits the training data (the “interpolation thresh-
old”), the test error goes down a second time as model complexity or training
continues to increase. In effect, very large or heavily trained models often
end up generalizing better than smaller ones. Figure 23 compares the tradi-
tional U-shaped error curve and the double descent error curve. This concept is
counter to classic underfitting/overfitting intuition. We have not observed sim-
ilar patterns within financial time-series but nonetheless it might inspire new
possibilities.

4.6 Hyperparameter Tuning

We have briefly mentioned hyperparameters, but we now take a careful look at
them as they are important components in the construction of a successful deep
neural network. When creating a network, we are presented with many choices.
It is nontrivial to set hyperparameter values beforehand and we need a systemic

8 See [Link] or [Link]

Trading.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 72 Quantitative Finance

Figure 24 Cross-validation for time-series.

way to search for optimal parameters. This process of searching for optimal
parameters is called hyperparameter tuning. These parameters are not part of
the so-called “inner” optimization of the model, such as learning the weights
of the neural network using gradient descent. During this inner optimization,
the hyperparameters are kept fixed. In hyperparameter optimization, which is
sometimes also called “outer” optimization we now repeat the inner optimi-
zation multiple times for different choices of hyperparameters with the aim of
finding the model with the lowest cross-validation error. There are many ways
to search for optimal hyperparameters and we introduce three popular methods
here.
The most basic hyperparameter tuning method is grid search in which we fit
a model for each possible combination of hyperparameters over a grid of pos-
sible values. Obviously, if we have a large number of hyperparameters to tune,
this method would be extremely time-consuming and inefficient. An alternative
to grid search is random search. Random search is different from grid search
in the sense that we do not come up with an exhaustive list of combinations.
Rather, we can give a statistical distribution for each hyperparameter and sam-
ple a candidate value from that distribution. This gives better coverage on the
individual hyperparameters. Indeed, empirical evidence suggests that only a
few of the hyperparameters matter which makes grid search a poor choice for
dealing with a larger number of candidates.
The previous two methods perform individual evaluations of hyperparame-
ters without learning from previous hyperparameter evaluations. The advantage
of these approaches is that they allow for trivial parallelization. However, we
discard the information from previous evaluations that could otherwise be used
to inform regions where we are more likely to find better hyperparameters.
For example, if initial evaluations show that the generalization error plateaus
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 73

quickly after reducing the learning rate, it might be less likely to find better
models when reducing the value even further.
Bayesian optimization (Frazier, 2018) is a sequential global optimization
(SMBO) algorithm that can be used to inform at which point in the hyperparam-
eter space to evaluate a model’s performance next given generalization errors
obtained from previous evaluations. It is specifically designed for scenarios
in which each evaluation of a target function is complex or expensive to run.
To implement such an approach, we first construct a model with some hyper-
parameters and obtain a score v according to some evaluation metric. Next, a
posterior distribution of the hyperparameter is computed and the choices for
the next experiment can be sampled according to this posterior expectation.
We would then repeat this process until convergence.
In practice, Gaussian Processes (GPs) are often used to model the objective
function. An intuitive way of thinking about a GP is as a Gaussian distribution
over continuous functions. Any finite number of points on this function are
distributed according to a multi-variate Gaussian – thus another way of thinking
about the GP is as a multi-variate Gaussian where the number of possible points
goes to infinity. The correlation between points is given by a kernel function
which depends on the distance between the points. Thus, the closer the points
the more correlated they are, which enforces the continuity of the GP.
In Bayesian optimization, one typically starts by specifying a GP prior over
the model’s generalization error across the hyperparameter space, often with
a zero mean and constant variance for simplicity. An initial evaluation is
performed on a random hyperparameter setting, after which the posterior dis-
tribution is updated based on the observed outcome. This updated posterior
then guides the selection of the next hyperparameters to explore, aiming to
efficiently locate optimal configurations. Intuitively, when choosing the next
point to evaluate the model, we have to trade off exploration and exploitation:
It makes sense to search further in regions where the GP indicates that the objec-
tive function is improving (exploitation). However, we also want to search in
areas where the uncertainty is large and we have no knowledge yet regarding
how good the objective might be (explorations).
In practice, we can carry out hyperparameter tuning by using Optuna
(Akiba et al., 2019), which is an open-source optimization framework designed
for hyperparameter tuning. It leverages techniques such as Bayesian optimi-
zation to systematically explore large search spaces and find optimal con-
figurations. We can easily integrate it with cross-validation to ensure that
optimizations are evaluated on multiple splits of data for reliable results. By
intelligently selecting the most promising hyperparameter settings to evaluate
at each step, Optuna minimizes the amount of training required and reduces the
need for extensive manual tuning.
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 74 Quantitative Finance

4.7 Setting Up Model Pipelines in Practice

The last step before deploying models to production is to have a pipeline that
encapsulates the processes mentioned previously and to build a robust frame-
work. It is also important to consider the capability of handling distributed
computing, which enables the scalability of our infrastructure. It is thus possible
to build the entire framework ourselves if we possess the necessary knowledge
and we can tailor each step based upon specific requirements. Otherwise, we
can also resort to established tools to build our frameworks. There are gener-
ally three popular frameworks, Ray, Dask, and Apache Spark, that facilitate the
construction of model pipelines. Each has its own strengths and use cases. The
three platforms have different design goals so it is difficult to say which is the
best in general. To better understand each platform, we compare them based on
performance, scalability, ease of use, ecosystem, and use cases.
Ray is designed for high-performance computing and excels in scenarios
requiring real-time execution, such as online learning, reinforcement learning,
and serving models. It is highly scalable and capable of handling millions of
tasks over thousands of cores with minimal overhead. Dask provides scalable
analytics and is optimized for computational tasks that fit into the Python eco-
system, including data manipulation with Pandas and NumPy. It is particularly
effective for parallelizing existing Python code and workflows. Apache Spark
is renowned for its speed in batch processing and its ability to handle streaming
data, courtesy of its in-memory computing capabilities.
In terms of ease of use, Ray offers a Python-native interface that is easy to
use for those familiar with Python programming. Its API is flexible, allowing
for straightforward integration with other machine learning and deep learn-
ing libraries. Dask integrates closely with Python’s data science stack, making
it accessible to data scientists and analysts already working with Pandas,
NumPy, or Scikit-learn. Its lazy evaluation model allows for efficient com-
putation. Apache Spark, while powerful, may have a steeper learning curve,
especially for users not familiar with its RDD and DataFrame APIs. However,
it provides good documentation and a vast array of functionalities beyond data
processing.
Ray has a growing ecosystem and is particularly strong in AI applications
with libraries like Ray Tune for hyperparameter tuning and Ray Serve for model
serving. It is also part of the Anyscale platform, which simplifies deployment
and scaling. Dask is part of the larger Python ecosystem, making it easy to inte-
grate with existing data science and machine learning workflows. It does not
have as wide an array of dedicated tools as Spark but excels because of its sim-
plicity and flexibility. Apache Spark boasts a mature ecosystem with built-in
libraries for various tasks, including Spark SQL for processing structured data,
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 75

MLlib for machine learning, and GraphX for graph processing. Its widespread
adoption ensures a wealth of resources and community support.
The choice between platforms depends on the problem to be solved and, in
some cases, you might even integrate these platforms to achieve desired out-
comes. For example, we could use Ray to build a start-to-finish framework for
deep learning models. For initial data preprocessing, we can use Ray’s remote
functions (@[Link]) to parallelize data fetching and use libraries like Ray
Pandas to normalize or extract meaningful features. Such libraries provide us
with Pandas-like operations but on a much larger scale. For deep learning mod-
els, Ray integrates seamlessly with frameworks like TensorFlow and PyTorch,
distributing the training process and making efficient use of available compu-
tational resources. In terms of hyperparameter tuning and cross-validation, Ray
Tune is an excellent tool that empowers us to distribute the search for the best
model parameters across multiple workers simultaneously. This is particularly
beneficial when experimenting with large models or when you need to iterate
quickly over many hyperparameter combinations.

PART II: APPLICATIONS

5 Enhancing Classical Quantitative Trading Strategies

with Deep Learning
In this section, we embark on an exploration of classical quantitative trading
strategies, dissecting their mechanics, applications, and the unique market con-
ditions they respectively best serve. Given the breadth and diversity of these
strategies, we divide this journey into three distinct parts.
The first part focuses on CTA-style futures and FX strategies in the com-
modities and foreign exchange markets. To start, we introduce the idea of
“volatility targeting,” a risk management technique that adjusts investment
exposure based on changing market volatility, with the objective of maintaining
a consistent risk profile throughout different market conditions. Next, we delve
into “time-series momentum” and “trend-following” strategies, as well as sim-
ple reversion models. These methods exploit the persistence of price trends over
time, whether by capitalizing on the continuation of current market directions
or by anticipating reversals. By analyzing historical price data, these strategies
seek to predict and profit from future price movements, making them partic-
ularly suited to the futures and FX markets where trends can be pronounced
and prolonged. We then round out the first part of our exploration by inves-
tigating the “carry” strategy. This approach seeks to profit from the interest
rate spread between different currencies, capturing the “carry” earned when
holding higher-yielding assets financed by borrowing lower-yielding ones.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 76 Quantitative Finance

This strategy highlights the importance of interest rates and funding costs in
trading decisions.
The second part of the section shifts focus to classical cross-sectional
strategies, which are important in the equity market. We explore the “long-
short” strategy via cross-sectional momentum, in which long positions are
taken in stocks showing strong performance and short positions in those with
weak performance. This method aims to capitalize on the relative momentum
across different securities, hedging market-wide risk by maintaining bal-
anced portfolio-level long and short exposures. We next discuss “Statistical
Arbitrage” (StatArb) strategies, which involve employing statistical models to
identify and exploit price inefficiencies between closely related assets. By ana-
lyzing historical price relationships and using statistical methods to identify
deviations from expected values, traders can execute high-frequency trades to
take advantage of temporary mispricings, all while managing risk and exposure
through sophisticated mathematical models.
The third part is the core of this section, in which we address the trans-
formative potential of deep learning to refine and revolutionize such classical
quantitative strategies. By leveraging deep learning algorithms, with their
ability to analyze vast datasets, traders can uncover complex nonlinear pat-
terns, and improve the predictive accuracy of models. This section covers how
deep learning can be integrated into both futures/FX and equity strategies,
from augmenting trend analyses in CTA-style strategies to refining the selec-
tion process in long-short equity approaches and improving the detection of
arbitrage opportunities in StatArb.
By providing insights into these cutting-edge techniques, this section aims
to equip readers with the knowledge to harness the power of deep learning,
pushing the boundaries of traditional quantitative trading strategies to achieve
enhanced performance and risk management in an increasingly complex mar-
ket environments.

5.1 Overview of Classical Quantitative Trading Strategies

5.1.1 Classical CTA-Style Futures and FX Strategies
Commodity Trading Advisors (CTAs) play an influential role in futures and
foreign exchange (FX) markets, employing a variety of strategies to generate
returns and manage risk. This section delves into classical CTA-style strategies,
focusing on long-only benchmarks, volatility targeting, time-series momentum,
and trend-following strategies. The explanation of each strategy is accompa-
nied by its mathematical underpinnings, so as to provide a deeper understanding
of its operational mechanics.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 77

Before diving into specific trading rules, we include a brief introduction

to futures contracts. We have also included an extended discussion of futures
contracts in Appendix B. Futures possess unique characteristics that must be
considered when performing data preprocessing. Futures contracts are stan-
dardized legal agreements to buy or sell an asset at a predetermined price on a
specified future date, and they have different end dates. Difficulties can arise
when joining futures contracts with different settlement dates. There are gen-
erally two ways to combine futures contract time-series: nearest futures and
continuous futures approaches.
The nearest futures approach is quite straightforward. To start, we select the
price series of a contract until its expiration, the next contract is then directly
selected, and so on until all contracts of consideration have been selected. How-
ever, the time-series generated with the nearest futures approach can not be used
for back-testing purposes because it includes significant price distortions due
to the price gaps on expiration dates. Figure 25 shows an example with such a
distortion where the nearest futures chart shows a large apparent price jump on
July 22, 2021. However, this price jump never took place because this is due
to contract expiration. In reality, all outstanding contracts are liquidated on (or
before) their respective settlement date. To maintain a position, a trader must

Figure 25 Top: price series generated by a nearest futures contract approach;

Bottom: price series generated by a continuous futures contract approach.
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 78 Quantitative Finance

“roll forwar” the contract by closing the one set to expire and opening a new
one with a future expiry. In essence, if the new contract is 20% cheaper, you
would be able to buy 25% more of those, for the same dollar amount.
On the contrary, the continuous futures approach reflects actual price move-
ments by linking successive contracts in a way that eliminates price distortions
(price gaps) at rollover points. This alternate linked-contract representation can
thus be used for back-testing and more accurately reflects the hypothetical gains
and losses of a trader. However, the trade-off is that the price series from con-
tinuous futures contracts will not match actual historical prices whereas those
generated by the nearest futures approach do. In some cases, we might even
observe the negative price series from the continuous futures approach. As a
result, the appropriate method to join futures contracts together depends on
the specific use case. Generally, the nearest futures contracts should be used if
the actual historical price is important, but if the goal is to simulate the gains
and losses of a strategy, the continuous contracts approach should be adopted
instead.

Long-Only Benchmark The long-only benchmark strategies are common in

investment management, and particularly relevant in futures and FX trading.
The default position for such strategy is the respective benchmark (S&P 500,
BTC, etc.) and that the trader tries to reallocate positions so as to achieve a
better (risk-adjusted) return than this benchmark. The strategy is thus evalu-
ated based on its relative performance to the benchmark rather than its absolute
performance.
By comparing the returns of actively managed portfolios against a long-only
benchmark, investors can gauge the value added by portfolio managers through
active selection and timing decisions. Also, the performance of long-only
portfolios can reflect broader market sentiments and trends. In bull markets,
long-only strategies are likely to perform well, capturing upside potential.
Conversely, their performance can suffer in bear markets, highlighting their
sensitivity to overall market conditions.
Correlations of these strategies with their benchmarks are also important.
For example, pension funds can achieve broad long-only market exposure cost-
effectively through cheap passive instruments. However, adding small alloca-
tions to uncorrelated strategies like time-series momentum, despite potentially
higher fees, enhances diversification and introduces the potential for excess
returns due to their differentiated risk-reward profiles.

Volatility Targeting Volatility targeting is a dynamic position sizing method

that can be used within strategies. It adjusts the exposure of an asset based on the
current or forecasted volatility of that asset or broader market. This ensures that
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 79

the level of risk remains stable over time. This method is especially significant
in the management of futures, FX trading, where market conditions fluctuate
significantly. By considering volatility – a primary measure of risk – investors
can potentially enhance risk-adjusted returns and better manage the drawdowns
associated with periods of high market turbulence.
The core idea behind volatility targeting involves scaling an asset’s invest-
ment exposure according to the ratio of a target volatility level to the current or
expected volatility of that asset. This adjustment factor can be defined as:
σtarget
A= , (70)
σcurrent
where σcurrent is the current asset volatility typically estimated using the stand-
ard deviation of historical returns over a specified look-back period. The target
volatility (σtarget ) is a predetermined level of risk that the investor aims to main-
tain. Its determination is guided by the investor’s risk tolerance, investment
timeline, and perspective on market conditions.
The trading positions are then scaled by the adjustment factor A to align
the volatility with the target level. Hence, if an asset’s current volatility is
higher than the target, its exposure is reduced (and vice versa), thereby aiming
to stabilize the risk profile. Figure 26 shows an example of a long-only S&P
500 benchmark strategy which has a Sharpe ratio of 0.461. It also includes a
version of the strategy that uses volatility targeting (to an annual volatility of
σtgt = 15%) to scale positions and consequently increases the Sharpe ratio to
0.632.
In practice, implementing a volatility targeting strategy involves continuous
monitoring of market conditions and trading performance. As market volatility

Figure 26 Long-only benchmark S&P 500 strategy and an accompanying

version that incorporates volatility targeting of 15% annual standard
deviation.
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 80 Quantitative Finance

changes, the risk exposure must be periodically adjusted to maintain the tar-
get risk level. This dynamic rebalancing requires a disciplined approach and
an efficient execution mechanism to minimize transaction costs and slippage.
Moreover, investors often employ advanced forecasting tools that consider fac-
tors like market sentiment, economic metrics, and geopolitical conditions, that
allow them to adjust their risk exposure in anticipation of potential volatility.
These models can range from simple historical volatility measures to complex
GARCH models and machine learning algorithms.
As we discuss in greater detail in the next section, volatility targeting across
multiple instruments can also be interpreted as a simple form of portfolio con-
struction. In particular, when assuming that the covariance matrix of portfolio
constituents is a diagonal matrix with respective variances on its diagonal
entries, then a standard mean-variance portfolio reduces to volatility targeting.
While assuming a diagonal covariance matrix tends to be a poor assumption for
equity markets, we can see that the covariance matrix of a universe of future
contracts is roughly block-diagonal with very small terms in the off-diagonals
(Figure 27).

Time-Series Momentum and Trend Following Time-series momentum

(TSM) and trend-following are quantitative trading strategies designed to profit
from ongoing market trends. Their core assumption is that assets showing
robust performance over a certain timeframe will likely maintain that momen-
tum, while assets underperforming during the same period will continue to lag.
These strategies are regularly applied across multiple asset classes – such as
futures, foreign exchange, equities, and commodities – and they have played
an important role in systematic trading.
A TSM strategy focuses on the autocorrelation of returns. It involves tak-
ing long positions in assets that have demonstrated rising price trends over a
predefined look-back period and short positions in assets that have shown a
downward trends. Specifically, a simple TSM strategy implementation could
be implemented as follows: for each instrument s, we assess whether the excess
return from the previous (k) periods is positive or negative. If it is positive, we
enter a long position; if it is negative, we take a short position. In both cases,
the position is maintained for (h) months.
According to Moskowitz, Ooi, and Pedersen (2012), which demonstrates an
example of a 12-month (k = 12) TSM strategy with a 1-month holding period
(h = 1), we can define the return of a time-series strategy (TSMOM) as:
σtgt
rTSMOM,i = sign(rit−k:t ) rit,t+1, (71)
t,t+1
σti

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 81

Figure 27 A heatmap of correlation matrix among various futures contracts.

where σtgt is the annualized volatility target and σti is an estimate of current
market volatility, which can be calculated by using an exponentially weighted
moving standard deviation on rit,t+1 . Note that in the previous formulation, when
working with returns (and ignoring or only using linear transaction costs), the
result does not depend on the actual overall position size. However, in practice,
one would actually target a dollar volatility, such as an volatility of σtgt USD =

10 million USD, rather than a percentage volatility of say σtgt = 15%. Then
USD /σ i,USD would correspond to the actual target trading position
sign(rit−k:t )σtgt t
in USD.
In this case, sign(rst−k:t ) is essentially the time-series momentum factor, where
we go long if the 12-month return is positive and vice versa. In practice, there
are various ways to decide the direction of our positions and we use Yt to
indicate trading directions in a more general case. We here introduce two pop-
ular trend-following strategies: simple moving-average crossover (SMA) and

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 82 Quantitative Finance

moving average crossover divergence (MACD) strategies. The SMA crossover

strategy utilizes two SMAs with different look-back periods K1, K2 (K1 < K2 ):
Yt = SMA(t, K1 ) − SMA(t, K2 ),
1Õ
K−1 (72)
SMA(t, K) = pt−i,
K i=0
where pt is the price of an instrument at time t, and we long if Yt > 0 and short
if Yt < 0. The formation of MACD uses an exponentially weighted moving
average (EWMA) to capture trends and momentum defined as:

Yt (S, L) = MACD(t, S, L),

MACD(t, S, L) = EWMA(t, S) − EWMA(t, L), (73)
EWMA(t, S) = αpt + (1 − α)EWMA(t − 1, S),

where a MACD signal has two time-scales S, which captures short-term move-
ment and L, which captures the long-term trend. α is the smoothing factor
(0 < α ≤ 1), which controls the degree of the weighting decrease for the
EWMA and we can define α in terms of a span S via α = S+1 2
. We can further
improve the signal by combining multiple MACD signals together. In such a
case, each MACD signal has a different time-scale and a final position could
be decided according to:
Õ
3
Ỹt = Yt (Sk, Lk ), (74)
k=1

where, for example, Sk ∈ {8, 16, 32} and Lk ∈ {24, 48, 96} days. Note that the
long look-back is often chosen to be roughly three times the short look-back.

Carry Carry trading is predominantly employed in the foreign exchange

market to exploit interest rate differentials between currencies. The strategy
involves going long on a currency that offers a higher interest rate while short-
ing a currency with a relatively low interest rate. Traders profit from the interest
spread, provided that the exchange rate remains favorable. The carry trade strat-
egy gained significant attention in the 1990s and early 2000s, with the Japanese
yen (JPY) often selected as the funding currency due to Japan’s low interest
rates. Accordingly, traders would leverage the low borrowing costs in JPY and
allocate those funds into higher interest rate currencies, such as the Australian
dollar (AUD) or the New Zealand dollar (NZD).
Persistent interest rate differentials can exist due to many reasons, including
differing economic policies, growth rates, and inflation levels across countries.
Carry traders that take advantage of these interest rate differences expect that a

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 83

higher-yielding currency will not depreciate against a lower-yielding currency

by an amount greater than the interest rate spread. If we are trading a currency
pair, say Currency A (with interest rate iA ) and Currency B (with interest rate
iB ), the interest rate differential (IRD) is:

IRD = iA − iB, (75)

for which going long Currency A and shorting Currency B, we will earn interest
on Currency A and pay interest on Currency B. The net interest earned per day
(I) on a notional amount of capital C can therefore be calculated as:

(iA − iB ) × C × l
I= , (76)
365
where l is the leverage that magnifies both potential profits and potential losses.
While the interest differential might be positive, there remains a risk that the
currency pair’s exchange rate moves against the position. If Currency A depre-
ciates against Currency B, it can negate the interest earnings or even lead to a
net loss. Accordingly, carry trading in FX markets involves not only a simple
interest rate arbitrage but also entails significant exchange rate risk. Traders
thus need to account for the possibility that currency movements could wipe
out the interest gains. Additionally, leverage, which is frequently employed in
carry trades, can magnify returns, but also heightens the potential for losses.
This makes it crucial to manage risk effectively in carry trading strategies.

5.1.2 Classical Equity Strategies

In the realm of quantitative finance, strategies in equity markets are pop-

ular tools, particularly for hedge funds and institutional investors. Among
these, classical equity strategies like long-short, cross-sectional momentum,
and statistical arbitrage stand out for their approaches to capturing alpha while
managing risk. Before discussing these strategies individually, we first intro-
duce the concept of portfolio optimization, as these strategies are mostly traded
in the form of a portfolio. A portfolio is the group of assets and the primary
goal of managing a portfolio is to balance risk and return in accordance with
the investor’s specific objectives. By distributing investments across a vari-
ety of asset classes (such as stocks, bonds, and real estate), different sectors,
geographic regions, and investment strategies, a portfolio can minimize idio-
syncratic risk (also called diversifiable or specific risk). This strategy, known as
diversification, helps mitigate the impact of poor performance of any individual
investment on the overall portfolio.
The success of a portfolio depends on the allocation of its assets. There are
various ways to determine the weightings of a portfolio’s constituent assets.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 84 Quantitative Finance

As a simple example, we present an equally weighted long-only portfolio with

volatility targeting:

1 Õ σtgt i
N
p
rt = r , (77)
N i=1 σti t,t+1

where Nt denotes the total number of assets within the portfolio, and rit repre-
sents the return of the asset i. The upcoming sections will outline traditional
trading strategies and illustrate how deep learning models can be utilized to
enhance these methodologies.

Equity Long-Short via Cross-Sectional Momentum A popular form of

long-short equity strategy involves buying undervalued (long positions) and
selling overvalued (short positions) stocks. Long-short strategies seek to gen-
erate returns in both upward and downward market conditions, achieving a
balance that reduces market exposure and captures alpha through stock selec-
tion. In fundamental long-short strategies investors or fund managers conduct
thorough research to choose stocks that are undervalued for purchasing and
those that are overvalued for short selling. The strategy often employs a funda-
mental analysis approach, looking at company financials, industry conditions,
and economic factors.
By maintaining long and short positions simultaneously, the strategy aims to
hedge market risk. Many funds aim for market neutrality by targeting a zero net
exposure, which is the difference between long and short exposures and can be
defined as:
(Value of Long Positions - Value of Short Positions)
Net Exposure = . (78)
Portfolio Value
This is also called a market-neutral strategy. The portfolio return (rp ) is
determined by taking the weighted average of the returns from the long posi-
tions (rL ) and subtracting the weighted average of the returns from the short
positions (rS ):

rp = wL · rL − wS · rS, (79)
where wL and wS are the weights of the long and short positions, respectively.
Another strategy for stock selection is the cross-sectional momentum strategy,
which capitalizes on the momentum factor across different stocks or sectors.
The underlying concept is that stocks that have outperformed their competi-
tors in the past are expected to sustain their strong performance in the short
to medium term, while those that have underperformed are likely to continue
struggling.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 85

Specifically, this strategy involves ranking stocks based on their past returns
and taking long positions in those within the top percentile while shorting those
within the bottom percentile. Mathematically, the strategy first ranks stocks
based on rit−1 , which is the return in the previous period. It then goes long stocks
with rit−1 in the top x% and short stocks in the bottom x%, with a typical value
for x% being 10%. To avoid sector biases and sector-specific exposure, the
strategy can be applied within sectors, buying the best performers and selling
the worst performers within each sector. Momentum strategies can exhibit con-
siderable variation in their effectiveness based on the chosen time frame for
measuring past returns, and often require back-testing to determine optimal
parameters. These strategies are staples in the quantitative trading world and
are widely applied in today’s trading markets.

Statistical Arbitrage (StatArb) Statistical Arbitrage, often referred to as

StatArb, is a sophisticated financial strategy that seeks to exploit statistical
mispricings of one or more often-related assets. Rooted in the principles of
mean reversion and quantitative analysis, StatArb involves complex mathe-
matical and computational techniques and is a subset of arbitrage strategies,
which aims to profit from price differences between markets or securities with-
out taking significant risk. Statistical Arbitrage has its roots in the convergence
trading strategy developed at Morgan Stanley in the 1980s. The approach was
pioneered by a group led by Nunzio Tartaglia, a physicist and mathematician.
Initially, it focused on pairs trading, which involves taking opposing positions
in two co-integrated stocks. Cointegration is a concept in time-series analysis
that applies to nonstationary series whose linear combination turns out to be
stationary. More concretely, consider two nonstationary time-series Xt and Yt .
If there exists some constant β such that Xt − βYt is stationary, we say that Xt
and Yt are co-integrated. Over time, Statistical Arbitrage evolved to incorpo-
rate multiple assets and use more sophisticated statistical models, leading to its
increased usage in quantitative trading.
In its simplest form, StatArb involves identifying pairs of co-integrated
stocks (pairs trading). When the price relationship between such a pair diverges,
the trader sells the overperformer and buys the underperformer, betting on the
j
convergence of their prices. For example, if pit and pt are the prices of two co-
integrated stocks i and j at time t, we would look for significant deviations in
j
their price ratio or difference. If the price ratio pit /pt increases so that it deviates
significantly from its historical mean, traders might short stock i and go long
on the stock j, betting on the ratio of their prices to revert toward the mean. In
more sophisticated multivariate approaches, a StatArb strategy might involve

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 86 Quantitative Finance

modeling pt , a vector of stock prices at time t, using a vector autoregressive

or deep learning model. By identifying complex relationships among multiple
stocks, traders can then construct portfolios that are expected to be market-
neutral and profit from mean reversion across related assets. Because StatArb
strategies rely upon subtle, unstable price relationships, they require rigorous
and active risk management.

5.2 Enhancing Time-Series Momentum Strategies

with Deep Learning
In the previous sections, we introduced several classical trading strategies. We
now demonstrate how to combine these strategies with deep learning models
to obtain better performance. By incorporating deep learning, we can better
analyze, model, and trade markets. Notably, time-series momentum trading,
which capitalizes on the continuation of asset price trends over time, greatly
benefits from deep learning’s ability to analyze extensive historical data and
uncover complex patterns that simpler algorithms might miss.
First, we present an end-to-end framework proposed by Lim, Zohren, and
Roberts (2019) which utilizes networks to directly optimize performance met-
rics. This framework, termed Deep Momentum Network, builds upon ideas
from time-series momentum strategies (Moskowitz, Ooi, and Pedersen, 2012).
In these strategies, a network is trained by optimizing the Sharpe ratio and
directly outputs trade positions. Second, we extend Deep Momentum Net-
work with transformers, as proposed by Wood, Giegerich, Roberts, and Zohren
(2021). In this framework, the transformers help to extract long term depen-
dencies and can be interpreted to a certain degree by their attention weights.
Third, we present an approach designed by Poh, Lim, Zohren, and Roberts
(2021a) which further extends cross-sectional momentum trading strategies.
In particular, they improve cross-sectional portfolios by integrating learning-to-
rank algorithms, recognizing that the effectiveness of a cross-sectional portfolio
relies heavily upon accurately ranking assets before portfolio construction.
Traditionally, quantitative trading is often a two-step optimization problem
where we first decide the direction and then the positions of the trades. The first
step is essentially a prediction problem and various methods, like the previously
introduced trend-following strategies can be used to predict price directions.
The second step is to determine positions based on these predictive signals and
similarly there are established methods for doing this. For example, we could
simply select the direction based on the signal’s sign and scale the size of the
position based on the signal’s magnitude.
With deep learning, we can bypass this two-step optimization problem by
concurrently learning trend analysis and determining position sizes within a

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 87

single function. The Deep Momentum Networks (DMN) framework, intro-

duced in Lim et al. (2019), directly output positions based on the objective of
maximizing strategy metrics, like returns or Sharpe ratio. Instead of outputting
a predictive signal like a standard supervised learning task, we use a network f
to output positions wit at any time point for asset i:

wit = f (uit ; θ), (80)

where uit are market features and θ are network parameters. In particular, we
aim to optimize the average return and the annualized Sharpe ratio using the
following loss functions:
Lreturns (θ) = −µR,
1Õ
=− R(i, t),
N Ω
√
µR × 252 (81)
Lsharpe (θ) = − q ,
Í
( Ω R(i, t)2 )/N − µ2R
σtgt
R(i, t) = wit i rit,t+1,
σt
where µR represents the average return across the entire universe Ω of size
N and R(i, t) denotes the return generated by the trading strategy for asset i at
time t. We can employ different network architectures to model the relationship
between the position wit and the market features uit . The entire computational
process is differentiable, which allows for the use of gradient ascent to max-
imize the objective functions. In practice, we multiply the loss functions by
minus one and use gradient descent to minimize them. The following code
snippet demonstrates how to construct a negative Sharpe ratio loss function in
Pytorch:
1 import torch
2 import [Link] as nn
3

4 def Neg_Sharpe(portfolio):
5 return -[Link](portfolio) / [Link](portfolio)
6
7 class SharpeLoss([Link]):
8 def __init__(self):
9 super ().__init__ ()
10

11 def forward(self , outputs_prev , future_rets):

12 portflio = outputs_prev * future_rets
13 loss = Neg_Sharpe(portflio)
14 return loss

We include results from Lim et al. (2019) to demonstrate the effectiveness

of DMNs. The authors assessed multiple different network architectures by

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 88 Quantitative Finance

back-testing their performance across eighty-eight ratio-adjusted continuous

futures contracts sourced from the Pinnacle Data Corp CLC Database. These
contracts contained price data spanning from 1990 to 2015 for a diverse set
of asset classes, including commodities, fixed income, and currency futures.
The following metrics were used to gauge the trading performance: expected
returns (E(R)), volatility (Std(R)), downside deviation (DD), the maximum
drawdown (MDD), Sharpe ratio, Sortino ratio, Calmar ratio, the percentage
of positive returns observed (% of +Ret) and the average profit over the aver-
age loss (Ave.P/Ave.L). The exact definitions of these metrics can be found in
Appendix C.
In Table 4, we present the experimental results alongside three classical
trading benchmark strategies: long-only, using the sign of past returns for
time-series momentum strategies (Sgn(Returns)) and MACD signals. We also
test on different network architectures, including a simple linear model, MLP,
WaveNet, and an LSTM. The complete testing period extends from 1995 to
2015, during which we optimize the performance metrics for the strategy’s
returns as outlined in Equation 80. In Table 5, volatility scaling is applied to
adjust the overall strategy returns to align with the volatility target (15%). The
rescaling of volatility should, in general, increase Sharpe ratio and facilitate
comparisons between different strategies.
When reviewing the raw signal outputs (Table 4), the LSTM model opti-
mized for the Sharpe ratio delivers the highest performance, exceeding the
Sharpe-optimized MLP by 44% and the Sgn(Returns) strategy – the top clas-
sical approach – by more than double. Additionally, the DMN enhances the
Sharpe ratio for both the linear and MLP models. This suggests that models
capable of capturing nonlinear relationships can achieve superior results by
utilizing extended time histories through an internal memory state.
We report the results with the addition of volatility scaling in Table 5. The
results clearly demonstrate that the addition of volatility scaling improves
performance ratios across strategies. Specifically, the volatility scaling has a
greater positive impact on network-based strategies compared to the classi-
cal strategies for which the Sharpe-optimized linear models beat reference
benchmarks. In terms of risk evaluation metrics, the adjusted volatility also
makes downside deviation and maximum drawdown comparable across strate-
gies. The LSTM models optimized for the Sharpe ratio maintain the lowest
maximum drawdown among all models and consistently achieve superior
risk-adjusted performance metrics.

The Momentum Transformer In the previous section, we observe that net-

work architectures can be successfully used for momentum strategies, and

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
use, available at [Link] [Link]
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of

Table 4 Performance metrics – raw signal outputs.

% of Ave.P
E(R) Std(R) DD MDD Sharpe Sortino Calmar +Ret Ave.L

References
Long Only 0.039 0.052 0.035 0.167 0.738 1.086 0.230 53.8% 0.970
Sgn(Returns) 0.054 0.046 0.032 0.083 1.192 1.708 0.653 54.8% 1.011
MACD 0.030 0.031 0.022 0.081 0.976 1.356 0.371 53.9% 1.015

Linear
Sharpe 0.041 0.038 0.028 0.119 1.094 1.462 0.348 54.9% 0.997
Ave. Returns 0.047 0.045 0.031 0.164 1.048 1.500 0.287 53.9% 1.022

MLP
Sharpe 0.044 0.031 0.025 0.154 1.383 1.731 0.283 56.0% 1.024
Ave. Returns 0.064 0.043 0.030 0.161 1.492 2.123 0.399 55.6% 1.031

WaveNet
Sharpe 0.030 0.035 0.026 0.101 0.854 1.167 0.299 53.5% 1.008
Ave. Returns 0.032 0.040 0.028 0.113 0.788 1.145 0.281 53.8% 0.980

LSTM
Sharpe 0.045 0.016 0.011 0.021 2.804 3.993 2.177 59.6% 1.102
Ave. Returns 0.054 0.046 0.033 0.164 1.165 1.645 0.326 54.8% 1.003
use, available at [Link] [Link]
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of

Table 5 Performance metrics – rescaled to target volatility.

% of Ave.P
E(R) Std(R) DD MDD Sharpe Sortino Calmar +Ret Ave.L

References
Long Only 0.117 0.154 0.102 0.431 0.759 1.141 0.271 53.8% 0.973
Sgn(Returns) 0.215 0.154 0.102 0.264 1.392 2.108 0.815 54.8% 1.041
MACD 0.172 0.155 0.106 0.317 1.111 1.622 0.543 53.9% 1.031

Linear
Sharpe 0.232 0.155 0.103 0.303 1.496 2.254 0.765 54.9% 1.056
Ave. Returns 0.189 0.154 0.100 0.372 1.225 1.893 0.507 53.9% 1.047

MLP
Sharpe 0.312 0.154 0.102 0.335 2.017 3.042 0.930 56.0% 1.104
Ave. Returns 0.266 0.154 0.099 0.354 1.731 2.674 0.752 55.6% 1.065

WaveNet
Sharpe 0.148 0.155 0.103 0.349 0.956 1.429 0.424 53.5% 1.018
Ave. Returns 0.136 0.154 0.101 0.356 0.881 1.346 0.381 53.8% 0.993

LSTM
Sharpe 0.451 0.155 0.105 0.209 2.907 4.290 2.159 59.6% 1.113
Ave. Returns 0.208 0.154 0.102 0.365 1.349 2.045 0.568 54.8% 1.028
 Deep Learning in Quantitative Trading 91

that LSTM networks generally outperform the other networks in DMNs.

Nonetheless, LSTMs can struggle to handle long-term patterns and react to
major events like market crashes. In time-series contexts, attention mecha-
nisms and learnable attention weights can be used to assess the relevance of
past timestamps. This enhances the model’s ability to capture and consider
long-term dependencies. This approach also allows the model to focus, or place
higher attention, on significant events and regime-specific temporal dynamics.
Furthermore, the use of multiple attention heads allows for the examination of
multiple regimes that occur simultaneously across different timescales.
The works of Wood, Giegerich, et al. (2021); Wood, Kessler, Roberts, and
Zohren (2023); Wood, Roberts, and Zohren (2021) follow the DMN framework
and design Transformer-based networks that incorporate attention mechanisms.
One of the core attributes of the Momentum Transformer (TFT) is its ability to
effectively capture attention patterns in time-series data, segmenting the input
sequence into distinct regimes. This segmentation process allows the model to
focus on specific temporal windows in which market behavior exhibits con-
sistent momentum, thus enabling the TFT to learn and predict trends more
effectively. The attention mechanism in TFT dynamically adjusts to different
market regimes, helping to distinguish between periods of significant market
movements and noise. This segmentation process not only enhances predictive
accuracy but also provides insight into how different time periods contribute to
overall forecasts, allowing for more interpretable trading strategies.
TFT does not only use attention models, but rather is constructed from a
combination of LSTM and attention architectures. In the context of financial
markets, where low signal-to-noise ratios persist, the LSTM serves as a tool to
summarize local patterns and capture short-term dependencies and trends prior
to the application of an attention mechanism. In other words, the LSTM layer
acts as a filter that distills relevant information, allowing the attention mech-
anism to operate more efficiently on a more structured, cleaner representation
of the time-series. This approach differs from other applications of transform-
ers, such as in NLP, where raw sequence data might contain more immediately
discernible patterns. However, in the context of noisy, stochastic financial mar-
kets, the combination of LSTM for local pattern summarization and attention
for regime-based segmentation enables the TFT to outperform conventional
transformers.
Specifically, the aforementioned authors study the following Transformer
architectures: Transformer, Decoder-Only Transformer, Convolutional Trans-
former, Informer (H. Zhou et al., 2021) and Decoder-Only TFT (Lim et al.,
2021). They adhere to the experimental framework outlined in Lim et al.
(2019) by examining the test results over three periods: overall performance

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 92 Quantitative Finance

from 1995 to 2020 to evaluate general performance; underperformance period

from 2015 to 2020, a timeframe during which both classical strategies and
LSTM-optimized DMNs exhibited underperformance; and the COVID-19
crisis period spanning the COVID-19 pandemic period, characterized by mar-
ket regime shifts that included a market crash and a subsequent bull market.
In Table 6, we display the experimental findings in terms of the performance
metrics of the strategy over the portfolio of 88 futures for the aforementioned
periods. Notably, the Decoder-Only TFT achieves the highest performance
across all risk-adjusted evaluation metrics for both scenarios 1 and 2. When
compared to the LSTM-optimized model, the Sharpe ratio increases by 50% for
the period from 1995 to 2020 and by 109% for the period from 2015 to 2020.
During the COVID-19 crisis, the LSTM model experienced significant losses
but transformer models still deliver decent results, however, the TFT performs
less well than the other transformer models. Overall, the different variants of
the momentum transformer have higher returns and lower risks, indicating their
ability to better model price dynamics.
In Figure 28, we display the return plots for experimental periods 2 and 3.
These plots clearly show that LSTM models were ineffective during the period
of market instability from 2015 to 2020 and throughout the COVID-19 cri-
sis. In contrast, transformer architectures were capable of adapting smoothly
to sudden changes in market regimes, and outperformed LSTM models sig-
nificantly. Additionally, the nonhybrid transformer models excelled during the
Bull market that followed the COVID-19-induced market crash, capitalizing
on this sustained momentum of this regime.
To trade with DMNs, it is important to comprehend the reasoning behind
how the model selects positions. The attention mechanism within the TFT not
only highlights important segments of the time-series but also assigns greater
weight to specific key dates, such as those when significant market events
occurred. This feature provides transparency into the model’s decision-making
process, making it easier for traders and analysts to understand why partic-
ular predictions were made. By focusing on key dates, the TFT helps users
interpret how past market events influence future predictions, offering valua-
ble insights into the driving forces behind market momentum. In Table 7, we
present the significance of input variables for the Decoder-Only TFT, for the
years 2015–2020 and the COVID-19 crisis period. Overall, across both periods,
the daily return feature is assigned the highest weight, indicating that the TFT
pays the most attention to market movements over that past day as compared
to longer lookbacks.
It is also notable that daily data plays a less significant role during the
COVID-19 crisis than it does during the 2015–2020 period. This is likely

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
use, available at [Link] [Link]
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of

Table 6 Performance metrics – raw signal outputs.

% of Ave.P
E(R) Std(R) DD MDD Sharpe Sortino Calmar +Ret Ave.L

Average 1995 - 2020

Long Only 0.024 0.049 0.035 0.125 0.51 0.73 0.21 52.4% 0.988
TSMOM 0.043 0.044 0.031 0.063 1.03 1.51 0.94 54.2% 1.002
LSTM 0.027 0.016 0.011 0.021 1.70 2.66 1.68 55.1% 1.091
Transformer 0.031 0.024 0.016 0.029 1.14 2.13 1.53 54.7% 1.051
Decoder-Only Trans. 0.029 0.026 0.017 0.034 1.11 1.69 1.09 53.5% 1.051
Conv. Transformer 0.029 0.027 0.018 0.038 1.07 1.60 0.98 53.5% 1.041
Informer 0.023 0.013 0.008 0.014 1.72 2.67 1.79 54.8% 1.103
Decoder Only TFT 0.040 0.015 0.009 0.013 2.54 4.14 3.22 57.3% 1.154

Average 2015 - 2020

Long Only 0.017 0.050 0.035 0.114 0.37 0.51 0.15 51.9% 0.982
TSMOM 0.009 0.043 0.031 0.082 0.24 0.33 0.12 52.8% 0.931
LSTM 0.012 0.018 0.013 0.035 0.82 1.19 0.66 53.3% 1.004
Transformer 0.019 0.012 0.008 0.010 1.53 2.32 1.86 54.7% 1.071
use, available at [Link] [Link]
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of

Table 6 (cont.)
% of Ave.P
E(R) Std(R) DD MDD Sharpe Sortino Calmar +Ret Ave.L

Decoder-Only Trans. 0.013 0.019 0.013 0.026 0.72 1.03 0.60 52.7% 1.012
Conv. Transformer 0.018 0.019 0.007 0.031 0.98 1.47 0.77 52.9% 1.056
Informer 0.016 0.011 0.008 0.017 1.51 2.30 1.44 54.3% 1.089
Decoder Only TFT 0.019 0.012 0.006 0.017 1.71 2.61 2.06 55.7% 1.073

COVID-19
Long Only −0.014 0.067 0.056 0.123 −0.19 −0.22 −0.12 57.2% 0.720
TSMOM 0.009 0.047 0.031 0.041 0.21 0.32 0.22 50.0% 1.041
LSTM −0.041 0.028 0.025 0.053 −1.50 −1.67 −0.78 52.2% 0.643
Transformer 0.042 0.012 0.008 0.008 3.38 5.55 7.31 64.8% 1.066
Decoder-Only Trans. 0.080 0.025 0.014 0.010 3.01 5.55 8.56 58.8% 1.243
Conv. Transformer 0.031 0.019 0.014 0.016 1.81 2.74 3.17 57.4% 1.058
Informer 0.043 0.016 0.010 0.010 2.71 4.45 4.28 59.6% 1.137
Decoder Only TFT 0.018 0.017 0.013 0.021 1.22 1.74 1.57 60.3% 0.831
 Deep Learning in Quantitative Trading 95

Table 7 Decoder-Only TFT average variable importance.

2015–2020 COVID-19

rday 30.8% 24.4%

rmonth 13.6% 10.6%
rquarter 8.9% 14.0%
rbiannual 8.9% 8.5%
rannual 11.9% 13.5%
M(8, 24) 9.1% 9.7%
M(16, 48) 10.3% 11.9%
M(32, 98) 6.5% 7.3%

Figure 28 These figures compare the performance of variants of the

momentum transformer strategy with benchmarks for the 2015–2020 period
(left) and the COVID-19 crisis (right). In each plot, we display cumulative
returns adjusted to an annualized volatility level of 15%.

because the 2015–2020 timeframe was highly non-stationary, whereas 2020

included a substantial market crash followed by a distinct upward trend.
Accordingly, it is not surprising that during the COVID-19 crisis, the TFT
assigned quarterly returns greater weight. Additionally, MACD (M(S, L)) indi-
cators of all lookbacks were assigned above-average importance, further
demonstrating the TFT’s ability to adapt to each specific scenario.
To further demonstrate the interpretability of transformers we show feature
importances for forecasting Cocoa futures prices in Figure 29. Cocoa provides
a representative example of the model’s behavior when trading a commodity
future, displaying a series of well-defined regimes throughout the observed
period. This variation in variable importance for trading Cocoa futures is
illustrated over time from 2015–2020 in Figure 29. These varying feature
importances are a result of the model’s ability to effectively combine differ-
ent components at different moments, adapting its approach in response to
significant events.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 96 Quantitative Finance

Figure 29 Variable importance for Cocoa futures during out-of-sample

forecasting from 2015 to 2020 is illustrated in the accompanying figures. The
upper plot displays the price series, while the lower plot showcases the
Decoder-Only TFT model. To emphasize the most significant features, we
highlight the seven variables with the highest average weights.

5.3 Enhancing Cross-Section Momentum Strategies with

Deep Learning
As previously discussed, cross-sectional strategies are a widely adopted form of
systematic trading, that can be applied to many asset classes. These strategies
aim to capture risk premia by engaging in relative trading between assets –
purchasing those with the highest expected returns while shorting those with
the lowest. For a portfolio of securities that is rebalanced on a monthly basis,
the returns for a cross-sectional momentum (CSM) strategy at time t can be
represented as follows:

1 Õ i σtgt i
N

t,t+1 =
rCSM X r , (82)
N i=1 t σti t,t+1

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 97

t,t+1 represents the realized portfolio returns from time t to t + 1,

where rCSM
N denotes the number of stocks within the portfolio, and Xit ∈ {−1, 0, 1}
defines the cross-sectional momentum signal or trading rule for security i. The
overarching framework of the CSM strategy consists of the following four
components:
Score Calculation: Yti = f (uit ), (83)
where uit denotes the input vector for asset i at time t , and the strategy’s pre-
dictive model f generates the corresponding score Yti . For a cross-sectional
universe consisting of N assets, the collection of scores for the assets of con-
sideration is represented by the vector Y t = {Y1t , · · · , YNt }. The second step
involves ranking these scores. Each score ranking can be determined as:
Score Ranking: Zit = R(Yti ), (84)

where Zti ∈ {1, · · · , N} signifies the ranking position of asset i after the scores
are sorted in ascending order using the operator R(·). The third step is the selec-
tion process and typically involves applying a threshold to retain a specific
proportion of assets, which are then used to construct the corresponding long
and short portfolios. Equation 85 follows the assumption that the strategy uti-
lizes standard decile-based portfolios, meaning that the top and bottom 10% of
assets are selected:

 −1 Zti ≤ (0.1 × N),




Security Selection: Xti = 1 Zti > (0.9 × N), (85)



0
 Otherwise.
The last step is portfolio construction. For example, we might construct
an equally weighted portfolio scaled by volatility targeting as shown in Equa-
tion 82. Most cross-sectional momentum strategies conform to this framework
and are generally consistent in the final three steps: ranking scores, select-
ing assets, and constructing the portfolio. However, it can differ in the choice
of prediction models f used to calculate the asset scores, ranging from sim-
ple heuristic methods to advanced models that incorporate a wide array of
macroeconomic inputs. While there are numerous techniques available for
scores computation, we typically focus on three primary approaches: classical
momentum strategies, Regress-then-Rank, and Learning to Rank.
For classical momentum strategies, we calculate scores with time-series
momentum factors or signals, such as MACD. Equation 86 illustrates how an
asset could be scored based on its raw cumulative returns calculated over the
preceding 12 months:
Score Calculation: Yit = rit−252,t, (86)
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 98 Quantitative Finance

where rit−252,t represents the unadjusted returns of asset i over the 252-day
period ending at time t.
Differently, the regress-then-rank approach first requires a predictive model,
such as a standard regression or deep learning model. A score is then calculated
so that:

Score Calculation: Yti = f (uit ; θ), (87)

where f denotes a prediction model that receives an input vector uit and is param-
eterized by θ. We then designate a target variable, such as volatility-normalized
returns, and train the model by minimizing the MSE loss:

1 Õ i rt,t+1 2
i
L(θ) = (Y − ), (88)
N Ω t σti

where Ω denotes the collection of all N possible forecasts and target pairs across
the set of instruments and their corresponding time steps.
Learning to Rank (LTR) (T.- Y. Liu et al., 2009) is a research domain in
Information Retrieval that emphasizes the use of machine learning techniques
to develop models for executing ranking tasks. To introduce the framework
of LTR, we borrow examples from document retrieval. For training purposes,
we are provided with a collection of queries Q = {x1, · · · , xN }. Each query
xi is linked to a set of documents {x1i , · · · , xm i } that must be ranked according
to their relevance to the respective query. An accompanying set of document
labels yi = {y1i , · · · , ymi } is provided to indicate the relevance scores of the doc-
uments. The goal of LTR is essentially to learn a ranking function f that takes as
j j
input a pair (xi, xi ) and outputs a relevance score f(xi, xi ) that can then be used to
rank the j-th item for query i. There are several ways to train LTR algorithms,
but we choose to introduce the framework here using the point-wise approach.
j
We can treat each query-item pair (xi, xi ) as an independent instance and train
the model with the objective of minimizing the mean squared error between
the estimated scores and the actual relevance scores, expressed formally
as:
Õ
j j
Lpoint wise = ( f (xi, xi ) − yi )2 . (89)
i,j

The studies by Poh et al. (2021a); Poh, Lim, Zohren, and Roberts (2021b,
2021c); Poh, Roberts, and Zohren (2022) adopt the concept of Learning to Rank
and introduce a framework for integrating LTR models into cross-sectional
trading strategies. To apply this framework to momentum strategies, we can
equate each query to a portfolio rebalancing event. In this analogy, each asso-
ciated document and its corresponding label can be viewed as an asset and its

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 99

Figure 30 LTR for cross-sectional momentum strategy.

designated decile for the next rebalance. This decile is based on a performance
metric, typically returns.
Figure 30 illustrates a schematic representation of this adaptation. Follow-
ing this framework, for the training process, let B = {x1, · · · , xN } represent a
sequence of monthly rebalancing events. At each rebalancing point xi , there
is a collection of equity instruments xi = {x1i , · · · , xm
i } along with their corre-
sponding assigned deciles yi = {yi , · · · , yi }. With all rebalance-asset pairs, we
1 m
j
can form the training set {(xi, xi ), yi }i=1
N to obtain a trained function g to produce

scores. During testing, we inject out-of-sample data to obtain scores and then
rank these scores to select securities. Accordingly, we construct portfolios that
invest in the assets projected to deliver the highest returns and divest from those
expected to generate the lowest.
As a concrete example, Poh et al. (2021a) applied this approach to actively
trade companies listed on the NYSE from 1980 to 2019. At each rebalancing
interval, 100 stocks – representing 10% of all tradable stocks – were selected
and actively traded according to multiple different LTR algorithms. These
include RankNet (RNet), LLambdaMART (LM), ListNet (LNet), and ListMLE
(LMLE). To verify the effectiveness of LTR, they include four benchmarks: a
random selection of stocks (Rand), classical time-series momentum strategies
that use past returns (TM) or MACD signals (MACD) to calculate scores, and
a regress-then-rank technique that uses a MLP network (MLP).
The out-of-sample effectiveness of these different strategies can be evaluated
by the results shown in Figure 31 and Table 8. Figure 31 displays the strategies’
cumulative returns, while Table 8 presents the strategies’ principal financial
performance indicators. To enhance the comparability of each strategy’s perfor-
mance the overall returns are standardized to an annualized 15% portfolio-level
volatility target for all strategies. In this analysis, all returns are calculated

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
use, available at [Link] [Link]
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of

Table 8 Performance metrics – rescaled to target annualized volatility of 15%.

% of Ave.P
E(R) Std(R) DD MDD Sharpe Sortino Calmar +Ret Ave.L

Benchmarks
Rand 0.024 0.156 0.106 0.584 0.155 0.228 0.042 54.5% 0.947
TM 0.092 0.167 0.106 0.328 0.551 0.872 0.281 58.2% 1.114
MACD 0.112 0.161 0.097 0.337 0.696 1.157 0.333 59.1% 1.184
MLP 0.044 0.165 0.112 0.641 0.265 0.389 0.068 55.1% 1.001

Learning to Rank Models

RNet 0.243 0.162 0.081 0.294 1.502 3.012 0.828 69.3% 1.407
LM 0.359 0.166 0.067 0.231 2.156 5.321 1.555 76.2% 1.594
LNet 0.306 0.155 0.068 0.274 1.970 4.470 1.115 71.5% 1.679
LMLE 0.260 0.162 0.071 0.236 1.611 3.647 1.102 68.1% 1.534
 Deep Learning in Quantitative Trading 101

Figure 31 Cumulative returns – rescaled to target volatility annualized

volatility of 15%.

without accounting for transaction costs, focusing on the models’ inherent pre-
dictive capabilities. Both the graphical data and the statistical metrics clearly
indicate that the LTR algorithms surpass the benchmark group across all per-
formance criteria, with LambdaMART achieving the highest scores on the
majority of the evaluated metrics.
More generally, the ranking algorithms notably enhance profitability,
demonstrating both higher expected returns and the rate percentages. Even
the least effective LTR model significantly surpasses the top reference bench-
mark across all evaluated metrics. Although all models have been adjusted to
maintain similar levels of volatility, LTR-based strategies tend to experience
fewer severe drawdowns and reduced downside risks. Moreover, the lead-
ing LTR model achieves substantial improvements across various performance
indicators. This pronounced difference in performance highlights the value of
learning cross-sectional rankings, as it can lead to better results for momentum
strategies.

6 Deep Learning for Risk Management

and Portfolio Optimization
In this section, we will introduce concepts and practical tools for evaluating
risk in financial markets, as well as techniques to optimize portfolios for various
objectives. We begin by examining traditional risk metrics, such as the standard

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 102 Quantitative Finance

deviation and Value at Risk (VaR), which have long been employed to capture
both the volatility and potential downside of asset returns. These metrics are
foundational concepts for understanding market risk and inform a wide range
of decision-making processes in financial institutions. Next, we delve into clas-
sical models for volatility forecasting – covering established approaches such
as the HAR (Heterogeneous Auto-Regressive) model – that provided finan-
cial practitioners with insights into how market fluctuations evolve over time.
While these methods remain useful, they may not always capture the complex
structures present in modern, high-frequency financial market data. Conse-
quently, we also introduce deep learning models for volatility forecasting,
emphasizing how neural networks can learn intricate, nonlinear dynamics from
large datasets in ways that traditional econometric tools often cannot.
Following this discussion of measuring and forecasting risk, we shift our
focus to portfolio optimization strategies. The essence of portfolio optimization
is to find an asset allocation that optimizes for some investment performance
criteria. For example, a portfolio manager might aim to minimize volatility
or maximize the Sharpe ratio. The main benefit of investing in a portfolio
is the diversification which decreases overall volatility and increases return
per unit risk. We continue by exploring the classic mean–variance framework
pioneered by Markowitz (1952), which remains a foundational element of
modern portfolio theory. This approach weighs expected returns against the
portfolio’s variance (risk), enabling investors to construct an efficient frontier
of optimal risk–return trade-offs. We then discuss maximum diversification,
a strategy designed to spread risk across diverse assets or factors, and con-
sequently achieve a more stable performance profile across varying market
conditions.
Moving beyond these traditional methods, we next demonstrate how deep
learning algorithms can be applied to portfolio optimization. Based on two
works C. Zhang, Zhang, Cucuringu, and Zohren (2021); Z. Zhang, Zohren,
and Roberts (2020), we present an end-to-end approach that leverages deep
learning models to optimize a portfolio directly. Instead of predicting returns
or constructing a covariance matrix of returns, the model directly optimizes
portfolio weights for a range of objective functions, such as minimizing vari-
ance or maximizing the Sharpe Ratio. Deep learning models are adaptable to
portfolios with distinct characteristics, allowing for short selling, cardinality,
maximum position, and several other constraints. All constraints can be encap-
sulated in specialized neural network layers, enabling the use of gradient-based
methods for optimization.
By bringing risk measurements, volatility forecasting, and portfolio optimi-
zation together in one section, we underscore the integral connection between

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 103

these topics. Accurately forecasting volatility is vital not only for effective risk
management but also for informing the dynamic allocation of assets in a port-
folio. When market volatility patterns are well understood, practitioners can
align their portfolio strategies in a way that accounts for fluctuating levels of
uncertainty. In other words, volatility forecasting is not merely an isolated exer-
cise and it provides a predictive lens through which portfolio decisions can be
refined. Combining these topics ensures a holistic perspective, from quanti-
fying and forecasting market risk to deploying those insights in a systematic
strategy that seeks to balance returns and risk.

6.1 Measuring Risk

We start this section by reviewing the main concepts of risk in quantitative
trading, as risk measurement is crucial for developing, evaluating, and exe-
cuting trading strategies. There are many different ways to quantify risk, for
example, Value at Risk (VaR), expected shortfall, drawdown, Sharpe ratio,
and Sortino ratio. Each metric provides us with a unique perspective to under-
stand the potential losses of a trading strategy. This section combines risk and
portfolio optimization. We typically view risk in the context of portfolio opti-
mization as the uncertainty of returns, focus on the variability of asset prices
and the potential for investment loss. To do so, we tend to look at the following
metrics:

• Standard deviation (Volatility): The standard deviation of returns remains

one of the most widely used risk metrics in portfolio optimization. It mea-
sures how much returns fluctuate from their mean, indicating the variability
of performance. Hence, a larger standard deviation implies a higher level
of risk.
• Covariance and Correlation: These metrics capture the relationship
between the movements of two assets. Covariance indicates the direc-
tion of the relationship, while correlation provides both direction and
strength. Understanding the relationships between assets is crucial for
diversification, and is accordingly a key concept in portfolio optimization.
• Beta: Beta measures how an asset’s returns move in response to overall
market changes. When beta is above 1, the asset’s returns tend to amplify
market movements, while a beta below 1 means the asset’s returns are less
sensitive to market swings. In a portfolio context, assessing the portfolio’s
overall beta offers insights into its exposure to market fluctuations.
• Value at Risk (VaR) and Conditional Value at Risk (CVaR): VaR quanti-
fies the maximum potential loss over a specified time horizon at a chosen

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 104 Quantitative Finance

confidence level. In other words, with probability (for example) 95%, losses
will not exceed the VaR figure. When VaR is used as a risk constraint, it
effectively places a limit (threshold) on the acceptable level of potential
loss. CVaR extends this by indicating the expected loss beyond the VaR.
Thus, CVaR focuses specifically on the distribution’s tail, capturing the
average magnitude of losses that surpass the VaR.
• Downside Risk: This measures the potential for loss in adverse scenarios,
focusing on negative returns. Metrics like the Sortino ratio, which is the
ratio of the asset’s return relative to its downside risk, are particularly useful
in this context.

These risk metrics are popular indicators in both academia and industry.
Thus, a good understanding of these metrics provides us with the foundation for
managing our portfolio risks. Note that risk measurement is not a set-and-forget
process. Continuous monitoring is vital as market conditions, asset correlations,
and volatilities evolve. Consistent reviews are imperative to maintain align-
ment between the portfolio and an investor’s risk preferences and objectives.
By applying diverse risk metrics and regularly monitoring and adjusting their
holdings, investors can improve the likelihood of meeting their financial targets
while effectively managing their risk exposure.

6.2 Classical Methods for Volatility Forecasting

We focus on volatility, as it is one of the most commonly used risk mea-
sures. Volatility is computed via the standard deviation of price changes. These
changes can be either percentage changes in price, that is, returns, or price dif-
ferences. The former yields a volatility in percentage terms, for example, 15%
volatility per year, while the latter yields a dollar volatility, that is, $10 mil-
lion volatility per year. Different methods can be used to estimate standard
deviations. For example, we can simply take the weighted average of historical
volatility to represent current market conditions. We can also resort to sampling
methods such as Monte Carlo which relies upon random sampling and statisti-
cal techniques to approximate the probability distribution of returns. To go yet
a step further, we can use predictive models to forecast the future variability of
returns on a financial instrument.
We first delve into some classical methods to predict future volatility. In par-
ticular, we focus on the Heterogeneous Autoregressive (HAR) model for daily
volatility forecasts and the HEAVY (High-frequency based Auto-regressive
and Volatility) model, which further utilizes intraday data to improve its
forecasts.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 105

The HAR model is a popular approach for forecasting daily volatility. It

was introduced to account for the regularly observed enduring memory effect,
which suggests that volatility shocks can have prolonged effects over time. The
HAR model incorporates lagged values of daily, weekly, and monthly volatil-
ities to predict the next day’s volatility. In doing so, the model’s structure
acknowledges that market participants operate on different “heterogeneous”
time horizons. The predicted daily volatility vd,t+1 is denoted as:

vd,t+1 = β0 + β1 vd,t + β2 vw,t + β3 vm,t + ϵt, (90)

where vd,t, vw,t, vm,t represent the daily, weekly, and monthly volatilities respec-
tively, and β coefficients are the parameters that need to be inferred. The linear
HAR model is straightforward to estimate and interpret, which makes it a
valuable tool for capturing the dynamics of financial market volatilities.
While the HAR model is an effective method, the HEAVY model is designed
to forecast volatility using high-frequency data, which provides more granu-
lar insights into the market’s behavior compared to traditional low-frequency
data. HEAVY models are commonly used for modeling volatility from high-
frequency data like tick-by-tick or minute-by-minute price movements. In order
to estimate volatility from high-frequency data, we introduce the notion of
realized volatility (RVt ). A common estimate for realized volatility is:
v
tm
Õ
RVt = (r2ti ), (91)
i

where r2ti are the high-frequency returns, and m represents the number of high-
frequency intervals within a day (e.g., minutes). RV is used as a measure of the
total variance in asset prices over a specific time interval, and the idea is that
volatility can be obtained from the squared returns of the high-frequency price
series. We can then express the HEAVY model as the following:

vd,t+1 = α + βvd,t + γRVt + ϵt, (92)

where the realized volatility is used to capture the short-term volatility from
high-frequency data and the lagged volatility component is used for the long-
term trends. The benefits of leveraging high-frequency data allow for more
accurate volatility estimation, and HEAVY models are well-equipped to handle
the phenomenon of volatility clustering.
However, microstructure noise exists and the HEAVY model remains sensi-
tive to certain market effects, such as bid-ask spreads, and sampling frequency.
It is nontrivial to eliminate the noise, which could affect the accuracy of
volatility estimates and predictions.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 106 Quantitative Finance

6.3 Deep Learning Models for Volatility Forecasting

While largely effective, traditional volatility models sometimes struggle to cap-
ture complex, nonlinear patterns in financial time-series data. Deep learning
offers an alternative approach that is able to model such intricate dependencies
and nonlinearities. One of the innovative applications in this domain is HARnet
(Reisenhofer, Bayer, & Hautsch, 2022), which integrates the WaveNet architec-
ture with the original HAR model to enhance volatility forecasting. In the orig-
inal HAR model, daily, weekly, and monthly features are used to increase the
receptive field of the model. Similarly, as previously covered in the Founda-
tions section, WaveNet is a deep learning architecture with similarly capabili-
ties, but a higher level of complexity that enables it to capture non-linearities.
The proposed HARnet is able to leverage both models’ strengths to capture
sequential patterns with long memory properties in volatility data.
A typical HARnet comprises an input layer, stacked dilated convolutions,
skip connections, and a final output layer. A HARnet receives different lagged
volatilities (daily, weekly, and monthly) and injects this input data to stacked
dilated convolutions which process input volatilities to capture dependencies
across different time scales. Because of its dilation factor, WaveNet can expo-
nentially extend its receptive field as the depth of the network increases. This
facilitates the incorporation of information from long sequences without a sig-
nificant increase in computational complexity. After a subsequent series of
neural layers, a final output layer is used to predict future volatility. While the
aforementioned WaveNet-based architecture is interesting, its usage shows few
improvements, as it still only incorporates daily data. It is also somewhat coun-
terintuitive to simultaneously use WaveNet and daily, weekly, and monthly
inputs.
A much better use case for WaveNet is forecasting volatility from intra-
day data. First, the dynamic range is much larger, stretching from minutes to
months. Second, intraday data has more potential to contain interesting non-
linear patterns that can be exploited. Accordingly, Moreno-Pino and Zohren
(2024) introduces DeepVol, a model that adopts the WaveNet architecture to
predict volatility from high-frequency financial data. This network is composed
of a stack of dilated causal convolution layers and a subsequent dense layer that
produces the volatility forecast. The use of dilated convolutions allows the net-
work to efficiently increase its receptive field (i.e., the range of time steps it
can consider when making a prediction) without having to increase the number
of layers or computational cost. The work’s results show that these properties
allow the model to efficiently make use of past intraday data to enhance its
predictions.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 107

The application of deep learning to volatility forecasting represents a sig-

nificant advancement in financial modeling. By leveraging architectures like
WaveNet, deep learning-based models can better handle nonlinear relation-
ships in data, a valuable attribute when dealing with financial markets. Deep
networks are also particularly adept at capturing long memory characteristics
of volatility through dilated convolutions and attention layers. As financial data
continues to expand in volume and complexity, deep learning will likely play
an increasingly central role in the development of such advanced analytical
tools.

6.4 Classical Methods for Portfolio Optimization

Portfolio optimization is a cornerstone of modern finance, providing a struc-
tured approach to selecting investments that balance risk and return according
to an investor’s objectives. In this section, we review two important approaches:
mean-variance optimization and maximum diversification. Each method offers
a unique perspective on how to construct a portfolio to achieve specific invest-
ment goals. Mean-variance optimization, developed by Markowitz (1952) in
the 1950s, is a foundational component of modern portfolio theory. This
approach focuses on constructing a portfolio that aims to maximize returns for
a specified level of risk or, conversely, minimize risk for a specified level of
return. They introduce the concept of the efficient frontier, which represents
the portfolio construction that yields the highest expected return for each level
of risk. This is achieved through mathematical optimization by maximizing
returns and minimizing variances (or volatility which represents risk).
To construct such a portfolio, consider a set of n assets, where each asset has
an expected return ri and variance σi2 . The portfolio weights are represented
by w = (w1, w2, · · · , wn )T where wi is the proportion of the portfolio invested in
asset i. One way to formulate the optimization problem is to identify portfolios
that achieve the highest possible expected returns for a specified level of risk.
This can be expressed as:

max rp = wT r, s.t.,
w
(93)
σp2 = wT Σw = σp,0
2
, and wT 1 = 1,

where Σ is the covariance matrix of asset returns with σij representing the covar-
iance between asset i and j. σp,0
2 denotes a target level of risk and 1 is a vector

of ones ensuring that the weights sum to 1 (i.e., fully invested portfolio). In
solving the constrained maximization problem outlined earlier, one determines
the optimal portfolio weights that maximize returns while keeping risk at a
given level. An alternative formulation of the mean-variance problem focuses

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 108 Quantitative Finance

on minimizing portfolio risk while enforcing a specified target for expected

returns. This would be expressed as:

min σp2 = wT Σw, s.t.,

w
(94)
rp = wT r = rp,0, and wT 1 = 1,

where rp,0 denote a target expected return level. To solve this formulation of
the constrained optimization problem, we introduce Lagrange multipliers λ and
γ for the constraints. The Lagrangian L is:

L(w, λ, γ) = wT Σw + λ(rp,0 − wT r) + γ(1 − wT 1). (95)

To find the optimal solution, we take the partial derivatives of L with respect
to w, λ and γ and set each of them to zero:
∂L
= 2Σw − λrp,0 − γ1 = 0,
∂w
∂L
= rp,0 − wT r = 0, (96)
∂λ
∂L
= 1 − wT 1 = 0,
∂γ
where the solution obtained provides the optimal portfolio allocation that min-
imizes the portfolio risk for a given expected return. By setting the partial
derivatives equal to zero, we are essentially finding the point where the rate
of change of the objective function with respect to each asset weight is zero,
implying that the portfolio has reached an optimal balance between risk and
return. The Lagrange multiplier in this context represents the trade-off between
the expected return and the risk of the portfolio. It provides insight into how
much additional return can be achieved by increasing the overall level of risk in
the portfolio. The solution essentially tells us the proportion of wealth to allo-
cate to each asset in order to achieve the best risk-return trade-off, considering
both the covariance between asset returns and the constraints set.
The strategy of maximum diversification is based on the premise that a
portfolio that diversifies across a wide range of assets will typically have a
lower risk than the sum of its individual components. Accordingly, the objec-
tive is to trade a selection of assets that effectively lowers unsystematic risk,
thereby minimizing the overall portfolio’s volatility. As a result, maximum
diversification considers the correlations between assets rather than just their
individual risks. By holding assets with low or negative correlations, the aggre-
gate risk of a portfolio can be meaningfully reduced. A central measure for
this approach is the diversification ratio (DR), defined as the ratio between the
sum of the individually weighted asset volatilities and the total volatility of

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 109

the portfolio. A larger value therefore indicates more effective diversification.

In mathematical terms, we define DR as:
wT σ
DR = √ , (97)
wT Σw
where w is again the weight vector and wT σ represents the weighted sum of
the individual asset volatilities. The goal is to maximize the diversification ratio
with respect to the weights vector w, and can be expressed as:

wT σ
max √ , s.t. wT 1 = 1, (98)
w wT Σw
where 1 is a vector of ones, ensuring the weights sum to 1. The optimization
problem is a fractional programming problem due to the ratio in the objective
function. We can simplify this by maximizing the numerator while holding the
denominator constraint. This reformulation constrains the portfolio variance to
be constant (usually set to 1), and focuses on maximizing the weighted average
volatility. One approach to tackle this optimization problem is to again employ
Lagrange multipliers:

L(w, λ, γ) = wT σ − λ(wT Σw − 1) − γ(wT 1 − 1), (99)

where λ and γ are Lagrange multipliers for the constraints. We optimize the
portfolio weights by differentiating with respect to w, λ and γ:
∂L
= σ − 2λΣw − γ1 = 0,
∂w
∂L
= wT Σw − 1 = 0, (100)
∂λ
∂L
= wT 1 − 1 = 0.
∂γ
Intuitively, an investor allocates capital across a variety of assets that have
low or negative correlations with each other to achieve maximum diver-
sification (MD). Following the same logic, this strategy aims to minimize
unsystematic risk, capitalizing on the unique price movements of each asset.
The key advantage of MD is risk reduction without a proportional decrease in
potential returns, which is particularly appealing during turbulent market con-
ditions. This diversification can help protect against significant downturns in
any single investment or asset class as negatively correlated assets are unlikely
to all move in the same direction.
Although MPT and MD are popular, their underlying assumptions have
been widely questioned and frequently do not hold true in real financial mar-
kets. In particular, MPT presupposes normally distributed asset returns and

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 110 Quantitative Finance

assumes that investors are rational, risk-averse, and chiefly focused on mean
and variance. Nevertheless, financial datasets frequently exhibit highly erratic
behavior, making them prone to deviating from these assumptions, particularly
during episodes of sharp market fluctuations (see for instance (Cont & Nitions,
1999; Z. Zhang, Zohren, & Roberts, 2019b)). Additionally, MPT assumes a
static view of risk and return, ignoring the dynamic nature of asset perfor-
mance and market conditions. The estimates of expected returns, variances, and
covariances are also very difficult to obtain, and small errors in these estimates
can lead to significant discrepancies in the model results and consequently over-
or under-allocation to certain assets.

6.5 Deep Learning for Portfolio Optimization

We now tackle the problem of portfolio optimization from the perspective of
deep learning. Returns and their covariance matrices are often unstable and
difficult to estimate. However, these challenges can be addressed through the
application of deep learning algorithms. Specifically, C. Zhang et al. (2021)
and Z. Zhang et al. (2020) propose an end-to-end framework that leverages
deep learning models to directly optimize a portfolio. Unlike the conventional
two-step process of forecasting and then optimizing, this approach bypasses
the need for estimating future returns and the covariance matrix by directly
outputting optimal portfolio weights.
The proposed framework is highly flexible and capable of dealing with dif-
ferent objectives and constraints. We take a general mean-variance problem as
an example to demonstrate the framework. In a general mean-variance problem
that permits short selling, we seek to maximize:
Objective (1) Mean Variance Problem:
λ Õ n
maxE(rp,t+1 ) − Var(rp.t+1 ), s.t. ||wt ||1 = |wi,t | = 1, (101)
wt 2 i=1

where rp,t+1 = wTt rt+1 represents the portfolio return, while rt = (r1,t, · · · , rn,t )T
denotes the vector of returns for n assets at time t, with ri,t referring to the
return of asset i (i = 1, · · · , n). The index t can be any chosen interval, such as
minutes, days, or months. λ is the risk aversion rate that controls the trade-off
between returns and risks, and wt = (w1,t, · · · , wn,t )T represents the portfolio
weights that need to be optimized. In order to obtain wt , we adopt a deep neural
network (f) that outputs portfolio weights:
wt = f (X t ) , (102)

where X t denotes the inputs to the network. Figure 32 depicts the proposed
end-to-end framework, which contains two main components: the score block
and the portfolio block.
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 111

Figure 32 Architecture of the proposed end-to-end framework.

The score block maps inputs to portfolio scores. Inputs can be any market
information that might be useful for adjusting portfolio weights. For example,
past returns up to lag p (rt−p, . . . , rt ), momentum features (MACD). More spe-
cifically, a neural network maps the input data to fitness scores for each asset.
Higher fitness scores indicate a greater likelihood of receiving larger portfolio
weights. We denote this network as fscores and the resulting fitness scores as:

st = (s1,t, . . . , sN,t )T = fscores (X t ), (103)

where fscores can be any stack of neural layers, such as convolutional, recurrent,
or attention layers.
Within the portfolio block, we transform previously derived scores st into
portfolio weights that meet the constraints of the relevant differentiable func-
tions fweights (·). We then calculate the realized portfolio return rp,t+1 from the
underlying asset returns rt+1 and derive the loss based on the chosen objective
function. In the context of Equation 101, which allows short selling and stip-
ulates that the sum of absolute weights must be one, the fitness scores require
the following transformation:
wi,t = fweights (si,t ),
esi,t (104)
= sign(si,t ) × ÍN ,
sj,t
j=1 e

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 112 Quantitative Finance

where the entire framework is differentiable so gradient ascent can be used to

optimize model parameters. In practice, investors have different risk tolerance
and face different constraints. We can optimize different objective functions
and adjust the portfolio block (fweights ) to meet these various constraints as long
as these operations are differentiable. Specifically, we examine the following
objective functions and constraints:
Objective (2) Global Minimum Variance Portfolio (GMVP):

minVar(rp.t+1 ), (105)
wt

Objective (3) Maximum Sharpe Ratio Portfolio (MSRP):

E(rp,t+1 )
max , (106)
wt Std(rp,t+1 )
Constraint (1) Long-only and ∥wt ∥1 = 1: To ensure nonnegative weights, we
apply the softmax activation function to the scores. For i = 1, . . . , n, we define:
esi,t
wi,t = Ín sj,t , (107)
j=1 e

where Listing 2 in Appendix D demonstrates how to construct this constraint

in Pytorch.
Constraint (2) Maximum Position and ∥wt ∥1 = 1: To ensure the weights auto-
matically satisfy the upper bound u, we transform the scores using a generalized
sigmoid function ϕa (x) = a + 1+e1 −x (where a ≥ 0). Upon setting a = nu−1
1−u
, we
obtain wi,t ≤ u as follows:
ϕa (|si,t |)
wi,t = sign(si,t ) × Ín , (108)
j=1 ϕa (|sj,t |)

if we set the maximum position u to 1, ϕa (x) reduces to the standard sigmoid

function.
Constraint (3) Cardinality and ∥wt ∥1 = 1: To handle cardinality, we begin by
defining a sorting operator Π(·) that takes st ∈ Rn and generates a permutation
matrix Π(st ) ∈ Rn×n . As a result, s̃t = Π(st ) is the vector st arranged in descend-
ing order. Based on this ordering, we take long positions in the top k assets and
short positions in the bottom k to form our portfolio:
1 1 {si,t >du } e |si,t | 1 1 {sj,t <dl } e |si,t |
wi,t = × Ín |sj,t |
− × Í n |sj,t |
,
2 j=1 1 {sj,t >du } e
2 j=1 1 {si,t <dl } e (109)
jnk
du = s̃t [k], dl = s̃t [n − k], k = + 1,
2
where s̃t [k] denotes the k-th entry of s̃t , that is, the k-th largest value in st . To
calculate the sorting operator, we first introduce a square matrix Λti,j derived
from the fitness score st as follows:

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 113
Õ
Λti,j = (n + 1 − 2i)sj,t − |sj,t − sm,t |. (110)
m

According to previous works (Blondel et al., 2020; Cuturi, Teboul, & Vert,
2019; Grover, Wang, Zweig, & Ermon, 2018; Ogryczak & Tamir, 2003), the
permutation matrix Π(st ) can be constructed as:

 1, if j = argmax(Λti,: ),

Π(st )i,j = (111)
 0, otherwise.

Since the argmax function is not differentiable, Grover et al. (2018) intro-
duce a NeuralSort layer that substitutes argmax with softmax, producing a
differentiable approximation of Π(st ):
št )i,: = softmax(Λt ).
Π(s (112)
i,:

Thus Equation (109) becomes differentiable, allowing for the use of standard
gradient descent.
Constraint (4) Leverage, i.e. L∥wt ∥1 = L: In line with Equation (104), we scale
the overall exposure of the positions by a factor of L:
e |si,t |
wi,t = L × sign(si,t ) × Ín |s | . (113)
j=1 e
j,t

To illustrate the performance of this deep learning framework for portfo-

lio optimization, we use daily data from 735 stocks within the Russell 3000
Index. The dataset spans from 1984/01/03 to 2021/07/06 and the testing period
is designated as 2001 to 2021. When reporting the performance, we include
evaluation metrics from previous sections, as well as Beta, to gauge the port-
folio’s correlation with the S&P 500 Index. This metric is useful to consider
because it measures the sensitivity of a portfolio’s return to that of a market
index. Ideally, we would like a portfolio that is less correlated with the market,
as this helps limit risk during market downturns.
Table 9 displays the experimental results, which are divided into four sec-
tions. The first section (Baselines) comprises four benchmark strategies: the
S&P 500 Index, an equally weighted portfolio (EWP), the maximum diversi-
fication (MD) portfolio, and the global minimum variance portfolio (GMVP)
(Theron & Van Vuuren, 2018). The second block (MSRP) of Table 9 compares
the classical mean-variance optimization approach (MPT) with the proposed
deep learning algorithms (E2E) optimizing Sharpe ratio. For MPT, we first
predict assets’ returns by minimizing the mean squared loss and then substitute
these estimates to obtain optimal portfolio weights. For deep learning models,
gradient ascent is used to directly optimize the Sharpe ratio. We test on sev-
eral models in this part including a linear model (LM), a multilayer perceptron
(MLP), an LSTM network, and a CNN model.
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
use, available at [Link] [Link]
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of

Table 9 Performance metrics.

% of
E(R) Std(R) DD MDD Sharpe Sortino Beta +Ret

Baselines
S&P 500 0.061 0.196 0.140 0.568 0.402 0.563 1.000 54.1%
EWP 0.130 0.212 0.148 0.548 0.682 0.973 1.000 54.6%
MD 0.439 0.239 0.141 0.519 1.641 2.785 0.599 54.8%
GMVP 0.080 0.081 0.059 0.408 0.992 1.360 0.257 56.4%

MSRP
MPT-LM 0.004 0.015 0.011 0.062 0.290 0.414 0.009 50.4%
MPT-MLP 0.008 0.027 0.019 0.140 0.299 0.424 0.036 51.5%
MPT-LSTM 0.014 0.017 0.011 0.043 0.858 1.259 0.014 52.0%
MPT-CNN 0.007 0.017 0.012 0.093 0.426 0.609 0.014 51.3%
E2E-LM 0.049 0.044 0.030 0.168 1.116 1.649 0.011 54.6%
E2E-MLP 0.044 0.026 0.016 0.073 1.688 2.657 0.008 55.2%
use, available at [Link] [Link]
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of

E2E-LSTM 0.060 0.023 0.013 0.017 2.604 4.448 0.017 57.8%

E2E-CNN 0.023 0.024 0.017 0.084 0.931 1.365 0.046 53.1%

Other Objective Functions

E2E-LSTM-GMVP 0.001 0.011 0.008 0.060 0.047 0.067 -0.004 50.4%
E2E-LSTM-MVPλ=10 0.064 0.317 0.207 0.878 0.349 0.534 0.342 51.8%
E2E-LSTM-MVPλ=20 0.179 0.169 0.115 0.380 1.055 1.555 0.195 54.8%
E2E-LSTM-MVPλ=30 0.168 0.116 0.076 0.187 1.394 2.149 0.060 55.2%

Constraints
E2E-LSTM-MSRP-Long 0.368 0.197 0.125 0.253 1.691 2.666 0.767 56.6%
E2E-LSTM-MSRP-LEV 0.321 0.112 0.068 0.151 2.540 4.203 0.132 57.6%
E2E-LSTM-MSRP-CAR 0.032 0.056 0.039 0.167 0.588 0.844 -0.011 52.0%
E2E-LSTM-MSRP-MAX 0.057 0.021 0.012 0.026 2.683 4.459 0.021 57.8%
 116 Quantitative Finance

The third block (other objective functions) indicates the results for the
application of deep learning to different objective functions including global
minimum variance portfolio (GMVP) and mean-variance problem in Equa-
tion 101. The final section (Constraints) explores the influence of multiple
constraints by constructing a strictly long portfolio aimed at maximizing the
Sharpe ratio (MSRP-LONG), a leveraged portfolio (LEV) with L = 5, a
cardinality-constrained strategy (CAR) that selects 20% of the instruments
thereby going long on the top decile and shorting the bottom decile, and lastly
a portfolio that imposes a 5% maximum position limit for each instrument
(MAX).
In the second block of Table 9, the end-to-end (E2E) deep learning methods
outperform both the MPT and baseline models. The third block highlights how
varying objective functions influence model performance. Specifically, GMVP
not surprisingly provides the lowest variance. Additionally, adjusting the risk
aversion parameter λ in the mean-variance approach allows users to control
their preferred risk level – raising λ increases the penalty on risk, thereby
reducing variance. The final block presents results under different constraints,
demonstrating the framework’s flexibility. Users thus have the ability to cus-
tomize these constraints to align with their individual requirements and trading
conditions.

6.6 Recent Developments on Volatility Forecasting

and Portfolio Construction
In this last part of the section, we explore some more recent developments in the
application of deep learning to volatility forecasting and portfolio optimization.

6.6.1 Graph-Based Models and LLMs for Volatility Forecasting

It can be helpful to view financial assets as network structures. Supply networks
are one helpful example, in which volatility can spill over from one company to
another in a network. Moreover, textual information, such as news can also be
incorporated in the forecasting task. There has thus been an increased inter-
est in applying graph-based methods and large language modeling to such
problems.
We will review and detail some key works that apply these techniques, focus-
ing on how these models offer a deeper understanding of market dynamics and
provide more reliable predictions.

Volatility Forecasting with Graph-Based Models Traditional models for

volatility forecasting typically focus on single-variate time-series, where the

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 117

volatility of each asset is predicted independently based on its own historical

data. While models that use vector forms provide a way to study multiple time-
series simultaneously, they still fall short in capturing the complex relationships
and interactions between assets. This is a limitation in financial markets, where
assets often exhibit strong correlations and spillover effects that influence their
volatility.
We now look at graph-based models that address this issue by explicitly rep-
resenting these relationships as a graph, where assets are connected based on
their correlations or other relevant interactions with one another. By doing so,
traders can better capture the intricate dependencies across assets, allowing for
a more accurate and holistic approach to volatility forecasting. C. Zhang, Pu,
Cucuringu, and Dong (2024) introduces Graph-HAR (GHAR), which lays out
a framework to forecast multivariate realized volatility by extending the HAR
model with graphs. Suppose we have vd,t = (v1d,t, · · · , vNd,t ) as the vector of real-
ized volatility for n assets. Recall that a HAR model incorporates lagged values
of daily, weekly, and monthly volatilities to predict the next day’s volatility
(Equation 90). We can define GHAR as:

vd,t+1 = α + βd vd,t + βw vw,t + βm vm,t

(114)
+ γd à · vd,t + γw à · vw,t + γd à · vd,t,

where à is the normalized adjacency matrix of A. The adjacency matrix A

encodes relationships between assets (A[i, j] is the weight of the edge between
node i and node j) which can be determined in several ways. For instance,
we can resort to correlation-based methods, such as computing the pairwise
correlations of returns and adding correlations to edges between assets when
their correlations exceed a certain threshold. Ã · vt represents the neighborhood
aggregation over different horizons and γ represents the effects from con-
nected neighbors. Notably, GHAR assumes a linear relationship between the
volatilities of two connected assets. However, we have the ability to introduce
nonlinearities by using graph convolutional layers.
This brings us to the GNNHAR network which is a GNN-enhanced HAR
model by C. Zhang et al. (2023) which replaces the linear neighborhood aggre-
gation in GHAR (the term à · vt ) with GNNs. Specifically, let us define a
GNNHAR with l layers:

vd,t+1 = α + V t β + H (l) γ,
(115)
H (l) = GNN(H (l−1), A) = σ(ÃH (l−1) W (l) ),

where V t = (vd,t+1, vw,t+1, vm,t+1 ) ∈ R n×3 and H (0) = V t . We define W (l) as the
learnable weights and σ denotes the ReLU activation function.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 118 Quantitative Finance

GNNs possess several advantages for forecasting volatility. One key benefit
is that a GNN can model both the direct and indirect effects of asset inter-
actions. For example, when one asset experiences a large shock, it can cause
volatility to spill over to other assets in the network, even if those assets were
not directly affected by the initial event. This phenomenon is known as the
spillover effect and can reverberate between assets that are not directly related.
In other words, it describes the transmission of financial disturbances, such
as price movements, volatility shocks, or shifts in market sentiment, as they
propagate between a network of assets. GNNs are capable of modeling these
spillover effects, as they can incorporate a broader set of market dynamics that
traditional methods may miss.
In addition, GNNs can handle high-dimensional data efficiently. By leverag-
ing the graph structure, GNNs can learn from a vast array of asset interactions
without becoming overwhelmed by the dimensionality of the data. As a result,
GNNs can learn complex dependencies from historical data, making them more
adept at forecasting future volatility in a multivariate setting. Moreover, GNNs
have the ability to adapt to the evolving relationships between assets which
allows them to respond to changing market conditions, an especially valuable
trait in the fast-moving world of financial markets.

Volatility Forecasting with Text-Based Features LLMs (Large Language

Models) facilitate the use of new data sources for generating alpha. They are
capable of processing and interpreting vast amounts of text from different
sources, including news reports, social media feeds, and earnings call tran-
scripts. As a result, LLMs can capture meaningful insights, subtle sentiment
shifts, and nuanced market signals that might elude traditional numeric data
analyses. The incorporation of these text-derived features can thus lead to more
robust and timely predictions of market volatility, especially in fast-moving or
sentiment-driven trading environments.
In the work of Rahimikia, Zohren, and Poon (2021), the authors provide a
detailed exploration of how specialized word embeddings can be harnessed to
improve realized volatility forecasts. Word embeddings assign numeric vec-
tors to words, ensuring that terms sharing similar meanings or usage patterns
lie close together in the vector space. Put differently, these embeddings trans-
form discrete text (strings) into continuous numerical representations, where
semantic and contextual affinities are captured by each word’s position and
proximity. Their methodology emphasizes building domain-specific embed-
dings tailored to financial terminology and contexts, as opposed to using
generic NLP models. By training on a large corpus of financial documents, they
are able to detect subtle differences in how words or phrases are used across

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 119

various market scenarios, such as regulatory changes, earnings surprises, or

shifts in investor sentiment.
The paper demonstrates that by integrating these carefully calibrated text-
based features with conventional numeric factors in a machine learning frame-
work, one can better capture signals and significantly enhance predictive
performance. On typical days, time-series models generally tend to outperform
purely news-based models in volatility forecasting. This is because they cap-
ture historical price patterns and market dynamics under normal conditions.
However, news-based models tend to perform better during volatility shocks,
as they can quickly incorporate real-time information from news sources that
may drive sudden market fluctuations. Given the strengths of each, the opti-
mal performance is often achieved through models that integrate both text and
price data.
Several studies have demonstrated the effectiveness of such hybrid models.
For instance, Atkins, Niranjan, and Gerding (2018) show that sentiment analy-
sis from financial news enhances volatility forecasting accuracy. Similarly, Du,
Xing, Mao, and Cambria (2024) highlight the benefits of integrating natural
language processing with time-series analysis.

6.6.2 Graph-Based Models and LLMs for Portfolio Optimization

Utilizing Graph-Based Models to Improve Portfolio Construction As dis-
cussed in Section 3 and Section 6.6.1, GNNs are naturally adept at modeling
relationships among different companies. We now introduce the application
of graph-based models to construct portfolios. Suppose we have a graph G =
(V, E) where v denotes the nodes which are the companies and E are the edges
that represent the relationships between companies. If there are N nodes and
each node (asset) has a feature vector xi ∈ Rd , we can obtain a node-feature
matrix X ∈ RN×d by stacking the individual vectors.
We can build and train a GNN with historical data. To construct a graph, we
first need to define the nodes and create the adjacency matrix A. Again, there
are many different approaches to obtain the adjacency matrices and the most
straightforward way is probably to obtain them from correlation matrices. In
Figure 33, we present such an example where we calculate an adjacency matrix
by using the correlations of fifty futures contracts. After obtaining the graph,
we can then adopt graph layers to model interested outputs. Recapping the
derivation of a graph convolutional layer (shown in Equation 69), it processes
information as:
H l+1 = σ(ÃH (l) W (l) ), (116)

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 120 Quantitative Finance

Figure 33 A graph built from adjacency matrix of futures contracts.

where H (l+1) is the matrix of the node embedding at layer l and H (0) = X, W (l)
is a trainable weight matrix, σ is the nonlinear activation function and à is a
normalized version of the adjacency matrix A. Depending on the purpose of
the task, the final output embedding can vary.
Ekmekcioğlu & Pınar (2023) extend the framework introduced in Section 6.5
with graph layers to directly learn optimal asset allocations. By treating each
asset as a node and connections between assets as edges, they outline a frame-
work to capture intricate relationships that traditional models often overlook.
In this approach, GNNs are used as the primary tool for learning these relation-
ships and aggregate signals from each node’s neighbors to form more expres-
sive embeddings of each asset. The results indicate that GNN-based models
can provide better insights into how assets co-move and how certain market
events propagate through a network of financial assets. Moreover, graph-based
approaches allow the model to dynamically learn higher-order dependencies
among clusters of assets, rather than simply relying upon pairwise correlations
or static factor models.
Another interesting work by Korangi, Mues, and Bravo (2024) seeks to
capture the evolving relationships among hundreds of assets over extended
horizons. They elect to use Graph Attention Networks (GATs) to incorporate
dynamic information about how assets co-move and influence one another.
In such a framework, each asset is a node in a time-evolving graph, and the
adjacency matrix is periodically updated using rolling windows of returns or
other market signals. At each network snapshot, the GAT layer uses attention
mechanisms to assign weights to edges, so that connections with higher rele-
vance receive proportionally more information flow. The authors demonstrate

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 121

that the GAT-based model outperforms benchmarks and delivers consistently

superior results over the long term.
Additionally, GNNs provide a powerful and flexible framework for inte-
grating alternative datasets that do not consist of purely price or return-based
signals. Traditional approaches to portfolio optimization often rely on histor-
ical price returns and standard covariance estimates. However, these methods
may fail to capture more nuanced or rapidly evolving relationships between
assets, particularly when crucial information – such as industry news, social
media sentiment, or supply chain linkages – is available. GNNs address
this gap by allowing practitioners to construct networks from diverse data
sources, where each node represents a company (or other financial entity), and
edges capture meaningful relationships derived from any number of alternative
datasets.
One concrete example is building a news network, as discussed in Wan et al.
(2021). In this setup, a connection (edge) between two companies is formed
based on co-mentions in news articles, the frequency of joint coverage, or
sentiment correlations extracted from text analytics. Figure 34 shows a graph
built from textual data instead of a return correlation matrix. This information
might highlight hidden interdependencies. For instance, two companies oper-
ating in distant industries may appear uncorrelated based upon their returns
but share a major client–supplier relationship. A fact that might be consist-
ently flagged by journalists. By embedding these news-driven relationships
into a GNN, the model can learn representations that account not only for price
co-movements but also for deeper relational structures present in textual data.
As a result, the GNN-based portfolio optimization strategy might spot risks or
growth opportunities that purely return-centric models overlook.
Incorporating alternative data sources might also enhance model robustness,
as purely price-driven correlations could weaken due to market regime shifts
or high volatility. Supplemental links derived from news or other non-price
sources may provide more stable signals. This greater diversification of infor-
mation flows can also help the model identify important patterns – whether it is
a sudden strategic alliance, regulatory development, or unforeseen supply chain
disruption. Taken together, GNN models help capture a holistic view of how
businesses truly connect and interact. Accordingly, they often uncover valuable
structural insights that lead to more informed and potentially more profitable
portfolio decisions.

Incorporating Text-Based Features and Techniques from LLMs In addition

to the application of LLMs for volatility forecasting, we can naturally extend
the same concept to portfolio construction. As discussed previously, LLMs

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 122 Quantitative Finance

Figure 34 A graph built from a news network. Colors indicate that assets are
allocated to the same group.

can be used to extract features from earning call transcripts or macroeconomic

reports. This allows them to capture different views on market conditions that
are not readily available in price/volume data. By incorporating this informa-
tion into optimization techniques, we can construct portfolios that are better
aligned with the current market environment. Specifically, we look at the work
proposed by Hwang, Kong, Lee, and Zohren (2025) where they integrate LLM-
derived embeddings directly into the portfolio optimization process. Note that
the discussion in this section is rather intuitive, so we omit some details but
focus on the ideas underlying the adoption of LLMs for portfolio optimization.
Suppose we have a sequences of asset returns rt = {r1,t, · · · , rN,t } where ri,t
indicates the return of asset i at time t, and wt ∈ R N is the corresponding port-
folio weight at time t. Hwang et al. (2025) constructs a forecasting model f that

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 123

not only forecasts future returns but also predicts the corresponding portfolio
weights:

(rt+1:t+H, wt+1:t+H ) = f(rt−L:t, xt−L:t ), (117)

where xt denotes macroeconomic features, L is the look-back window and

H is the predicted horizon. To simultaneously generate predicted returns and
portfolio weights, the authors propose a custom loss function:

LLoss = βLMSE + (1 − β)LDecision, (118)

where LMSE is the standard MSE loss that measures the discrepancy between
actual returns and predicted returns, and LDecision measures how inaccuracies
in predicted returns translate into suboptimal portfolio decisions. An input
embedding is then used to process data from multiple modalities, specifi-
cally time-series decomposition and LLM-enhanced semantic embeddings.
After that, several network layers are implemented to detect temporal patterns
with LLM-derived semantic embeddings and convert predictions into portfolio
weights. Finally, the hybrid loss function (Equation 118) is optimized to derive
forecasts and portfolio weights. The field of LLMs in finance is still in its early
stage with a limited number of published works. For a broad coverage of how
LLMs can be applied to quantitative finance, interested readers can refer to
Kong, Nie, et al. (2024).

7 Applications to Market Microstructure

and High - Frequency Data
In this section, we delve into the microcosm of the financial world and focus on
high-frequency microstructure data. This field is probably the most attractive
area for deep learning methods, given the rich and detailed view of market
dynamics. High-frequency microstructure data captures every change in market
conditions, including price fluctuations, order volumes, and transaction times.
This granular view is ideal for training sophisticated algorithms.
Traditionally, in early financial markets, stocks were traded in a format
known as pit trading. In this system, traders and brokers gathered in a des-
ignated trading floor area, known as the “pit,” to conduct transactions via open
outcry. This involved shouting and using hand signals to communicate buy and
sell orders. The chaotic environment relied heavily on the physical presence and
vocal abilities of traders to execute trades quickly and efficiently. Market prices
were determined through a process of verbal negotiation and immediate, face-
to-face interactions. Despite its seeming disorder, pit trading enabled real-time
price discovery and liquidity in an era before electronic trading. However, it had

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 124 Quantitative Finance

obvious limitations, such as restricted market access for remote participants and
slower dissemination of price information. These shortcomings eventually led
to the adoption of electronic trading systems.
The transition from traditional pit trading to electronic trading systems,
marked a pivotal transformation in financial markets, fundamentally alter-
ing the landscape of global finance and paving the way for modern trading
practices. This shift began in earnest during the late twentieth century as tech-
nological advancements made electronic trading feasible and attractive. One
of the earliest and most notable shifts occurred with the establishment of the
NASDAQ in the early 1970s. As the first electronic stock market, the NAS-
DAQ used computer and telecommunication technology to facilitate trading
without a physical trading floor. Around the same time, the New York Stock
Exchange (NYSE) introduced the Designated Order Turnaround (DOT) sys-
tem, which routed orders electronically to the trading floor, although they were
still executed via open outcry.
In the late 1980s and early 1990s, as computers became more powerful and
network technology more sophisticated, more exchanges began to explore elec-
tronic trading options. The London Stock Exchange (LSE) moved away from
face-to-face trading with the “Big Bang” deregulation of 1986, which included
the introduction of electronic screen-based trading. This shift was mirrored by
exchanges around the world, including the Toronto Stock Exchange (TSE) and
the Frankfurt Stock Exchange (FSE).
The development of Globex by the Chicago Mercantile Exchange (CME) in
1992 was another significant advancement. Globex was an electronic trading
platform intended for after-hours trading that would eventually become a 24-
hour worldwide digital trading environment. Similarly, EUREX, established
in 1998 as a result of the merger between the German and Swiss derivatives
exchanges, was among the first to go fully electronic, setting a precedent
for derivatives trading globally. The adoption of electronic trading and the
Limit Order Book (LOB) system revolutionized market dynamics. With the
ability to process high volumes of transactions at unprecedented speeds, trad-
ing became faster, more efficient, and more accessible. Moreover, electronic
trading reduced the costs associated with trading and increased transparency
by making market data widely available. It also democratized market access,
enabling more participants to engage from remote locations.
Today, nearly all major stock and derivatives exchanges operate electroni-
cally. The transition has not only altered how trades are executed but also how
markets are monitored and regulated. Advanced algorithms and high-frequency
trading strategies that rely on microsecond advantages in electronic trading
environments have become prevalent, prompting ongoing discussions about

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 125

market fairness and stability. In order to digitize trading, every verbal bid and
ask needs to be converted into digital orders that can be entered into the LOB.
Each trader’s shouts and hand signals become electronic messages that spec-
ify the quantity, price, and conditions of trades. The electronic LOB system
then aggregates these orders, organizing them by price level and quantity. The
system then continuously updates as new orders come in, orders are modified,
and trades are executed. This shift also enhances transparency by providing all
market participants with a detailed real-time view of market activity and depth,
something that was not previously possible in the chaotic environment of the
trading pit.
Modern exchanges can generate billions of such messages in a day. The
high resolution and volume of this data enable deep learning models to discern
intricate patterns and dependencies that might be invisible in lower-frequency
data. Next, we will give a detailed description of high-frequency microstruc-
ture data. This will include an exploration of the inner working mechanism of
exchanges and the aggregation of individual order messages into limit order
books, which reflect supply and demand at the microstructure level. By lever-
aging such large datasets, we have numerous opportunities across various
financial applications, such as generating predictive signals that drive algorith-
mic trading decisions, optimizing trade execution strategies, and even creating
advanced generative models that can simulate entire exchange markets. Such
simulations can be used with reinforcement learning algorithms to design better
trading strategies, accounting for market impact, fill rate, and market anoma-
lies. Consequently, high-frequency microstructure data is not just facilitating
more informed decisions but is also a key component of many innovations in
financial technologies.

7.1 The Inner Working Mechanism of an Exchange

In this section, we detail the order lifecycle and explain how a limit order
book is maintained within an exchange. Once a trader places an order, the
corresponding message traverses various intermediaries, including exchanges,
banks, brokers, and clearing firms. Exchanges typically broadcast these mes-
sages in real time through a data feed, enabling the reconstruction of the LOB.
In essence, the LOB is simply the organized collection of these order messages.
The comprehensive assembly of these messages, referred to as market by
order (MBO) data, is one of the most detailed sources of microstructure
information. Specifically, MBO data captures every market participant’s order
instructions, and activities, such as placing a new order or canceling an existing
one. The fundamental elements of MBO data include time stamps, order prices,

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 126 Quantitative Finance

order volumes (sizes), order directions (sides), order types (a market order, a
limit order, etc.), order IDs which serve as a unique and anonymous identifier
for each individual order, and actions that describe the specific instruction of
a trader (buying, selling, or canceling an order). Table 10 shows a snapshot of
sequences of MBO data that contains essential information. For simplicity, we
omit some nonessential auxiliary information.
Figure 35 presents a snapshot of the LOB at a given time t which illustrates
the collection of all currently active limit orders. When a trader places orders, a
market order is matched immediately with an existing, resting order, whereas a
limit order enables traders to specify the worst price and quantity they wish to
transact. These limit orders stay active. Once an exchange has received a limit
order, it will place the order at the appropriate position within the existing LOB.
The incoming MBO data continuously alters the LOB and a new snapshot of
the LOB is formed whenever it gets updated.
A LOB consists of two primary types of orders: bids and asks. A bid order
signifies a willingness to purchase an asset at a specified price or lower, while
an ask order indicates an intention to sell an asset at a particular price or higher.
As shown in Figure 35, bids or asks have prices P(t) and sizes (volumes) V(t).
Each rectangle in the graph represents a single order, with its size represented
by the square’s height. Therefore, each level of a LOB is an ordered queue of
all limit orders at that specific price level.
Figure 36 illustrates how a limit order book evolves and demonstrates the
impact of an MBO message on the existing LOB. For instance, at the top of
Figure 36, a new limit order (ID=46280) is added to the ask side of the order
book with a price of 70.04 and a size of 7580. This order addition updates the
order book by placing the new order at the corresponding price level. Similarly,
the LOB is altered when there is a cancellation (as shown in the middle top
figure), a partial cancellation (middle bottom figure), or when a market buy
order is executed (bottom figure).
In practice, we can obtain high-frequency microstructure data by subscrib-
ing to market exchanges. Exchanges typically offer data across three tiers:

Figure 35 A snapshot of LOB at time t.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
use, available at [Link] [Link]
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of

Table 10 An example of a sequence of market by order data.

Time stamp ID Type Side Action Price Size

2022-04-06 [Link].125873685 587984865448934894 2 1 1 58.45 1578.0

2022-04-06 [Link].875348668 587984865448937899 1 N/A 2 N/A N/A
2022-04-06 [Link].584863148 587984865448937899 2 1 0 58.50 4781.0
2022-04-06 [Link].871548935 587984865448931459 1 2 1 58.50 2141.0
2022-04-06 [Link].933314896 587984865448938794 1 N/A 2 N/A N/A
 128 Quantitative Finance

Figure 36 This illustration demonstrates how MBO data modifies a LOB.

Top: A new limit order is introduced; Middle top: An existing order is
canceled; Middle bottom: An order undergoes a partial cancellation;
Bottom: A marketable buy limit order crosses the spread.

Level 1, Level 2, and Level 3. Each tier provides progressively more detailed
information and capabilities, with corresponding subscription costs:

• Level 1 Data: This tier comprises the price and volume of the latest trade,
along with the current best bid and ask prices, which is commonly referred
to as quote data.
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 129

• Level 2 Data: This tier supplies LOB data, providing more comprehensive
information than Level 1 by displaying bid and ask prices along with their
respective volumes across multiple deeper levels of the order book.
• Level 3 Data: This tier goes beyond Level 2 by providing unaggregated
details of bids and asks placed by individual traders (MBO data), delivering
the most granular view of market activity.

The choice of which data source to use depends on the specific application
or analysis being conducted. Each tier of market data offers unique advantages
and levels of detail suitable for different purposes. LOB data, typically pro-
vided at Level 2, aggregates the total available quantities at each price level
in the market. This aggregated view gives insight into the overall demand and
supply dynamics at a microstructure level, helping analysts assess liquidity,
price stability, and potential market impact. However, LOB data lacks informa-
tion about individual orders, focusing instead on summarized market activity.
In contrast, MBO data, available at Level 3, provides granular details about
individual market participants’ behaviors. It includes unaggregated bids and
asks, along with unique order identifiers. This level of detail enables a deeper
understanding of queue positions, order prioritization, and the trading strate-
gies employed by participants. MBO data is especially valuable for applications
that require precise modeling of order flow dynamics, such as market impact
analysis, execution optimization, and algorithmic trading. By combining LOB
and MBO data, it is possible to gain both macro and micro views of the mar-
ket, allowing for more comprehensive analyses tailored to the needs of specific
trading strategies or research objectives.

7.2 Deep Learning–Based Predictive Signals

In recent years, the use of deep learning models for analyzing high-frequency
microstructure data has gained significant attention. This growing trend is
fueled by the immense volume of data generated by modern exchanges, with
billions of quotes, orders, and trades produced within a single trading day.
High-frequency microstructure data offers a rich source of information for
advanced modeling and prediction. Deep learning models are exceptionally
effective for analyzing this type of data because they can identify and interpret
intricate patterns.(Atsalakis & Valavanis, 2009).
We can take snapshots of limit order books with a look-back window and for-
mat them as an “image” that is shown in Figure 37. The topology of LOB shows
clear patterns in terms of prices and volumes and can be fed into deep learn-
ing algorithms. The work of Tsantekidis et al. (2017a) designs likely the first
deep learning model to successfully apply a CNN to predict price movement for

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 130 Quantitative Finance

Figure 37 Limit order book data across times.

limit order books. A key advancement in this study is the application of CNNs
to directly predict stock prices using LOB data. To do so, the authors adapt
CNNs, traditionally successful in image processing, to handle the structured
time-series data of LOBs. By treating the LOB as a multidimensional array,
CNNs can learn spatial hierarchies and patterns within the order book that are
predictive of future price movements. This approach leverages the depth of data
available, capturing subtle yet critical shifts in market sentiment that might be
indicative of future trends. Other studies (Z. Zhang, Zohren, & Roberts, 2019a)
have shown that CNNs can outperform classical statistical models and other
machine learning methods in predicting short-term price changes, providing
traders with a powerful tool for making more informed decisions.
Interestingly, the work of Sirignano and Cont (2018) has uncovered universal
features of price formation in limit order books. By analyzing vast amounts of
LOB data across different assets and markets, their models have identified com-
mon patterns and dynamics that govern price changes. These insights suggest
that despite the apparent complexity and noise within financial markets, there
are underlying principles and patterns that can be extracted through deep learn-
ing. The ability of deep learning models to distill these features from the data
not only enhances predictive accuracy but also provides a deeper understanding
of market mechanics.
In a more specialized context, Z. Zhang et al. (2019a) carefully designed a
deep network, termed DeepLOB, to predict price movements from LOB data
using an architecture that combines convolutional filters and LSTM modules.
Convolutional filters are utilized to capture the spatial patterns of the LOB,
while LSTM modules are employed to model longer-term temporal dependen-
cies. This proposed network continues to achieve state-of-the-art performance,
serving as a benchmark and inspiring a wide range of studies and applications
in financial modeling and trading. We implement DeepLOB for a regression
problem and attach the code script in Listing 3 in Appendix D.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 131

Figure 38 An attention model that utilizes limit order books for

multi-horizon forecasting.

However, DeepLOB can only make a single-point estimation. In practice, the

predictive horizon remains a hyperparameter that needs be carefully adjusted
as it determines holding time, trading frequency, risk, and more. To over-
come this, Z. Zhang and Zohren (2021) extends on DeepLOB and implements
Seq2Seq and Attention modules for LOB to produce multi-horizon estimates.
Instead of obtaining a single-point estimation, they obtain a forecasting path
that can be better utilized to generate trading strategies. Figure 38 depicts this
Attention structure. It shows that the Attention module places different weights
across time, with short-term predictions rolled forward to generate long-term
estimates. In addition, the work of Z. Zhang, Lim, and Zohren (2021) utilizes
MBO data to predict price movements and demonstrates that predictive signals
obtained from MBO data deliver comparable results to models trained on LOB
but are less correlated.
For a more complete review of the predictive models for LOB, readers
can refer to Briola, Turiel, and Aste (2020), where they have compared and
benchmarked several machine learning algorithms and deep networks on the
same feature space, dataset, and tasks. The later work Briola, Bartolucci, and
Aste (2024) also proposes an innovative operational framework that evaluates
predictions’ practicality. They studied instruments across various dimensions

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 132 Quantitative Finance

including tick size, predictive horizon and order book depths. Prata et al. (2024)
also carefully compare the predictive power of fifteen cutting-edge DL mod-
els based on LOB data. For more interesting works, readers can refer to Bao,
Yue, and Rao (2017); Chen, Chen, Huang, Huang, and Chen (2016); Di Persio
and Honchar (2016); Dixon (2018); Doering, Fairbank, and Markose (2017);
Fischer and Krauss (2017); Nelson, Pereira, and de Oliveira (2017); Selvin,
Vinayakumar, Gopalakrishnan, Menon, and Soman (2017); Tsantekidis et al.
(2017b, 2017a).

7.3 Deep Learning for Trade Execution

In the previous section, we introduced various predictive models. We now dis-
cuss trade execution, which heavily depends on the usage of high-frequency
microstructure data. In the fast-paced world of financial markets, the execution
of trades with precision and efficiency is paramount. Trade execution focuses
on the granular details of executing large orders in financial markets to min-
imize costs and market impact. This aspect of trade execution delves into the
strategies and techniques used to break down large orders into smaller, man-
ageable parts and to determine the optimal execution sequence. The goal is to
achieve the best possible execution price while mitigating the adverse effects
on the market, such as price slippage and increased volatility.
When executing large orders, the sheer increase in volume can influence
the market price, causing unfavorable movements against a trader’s interests.
By strategically breaking down and timing the execution of these large orders,
traders can reduce their market footprint, thus minimizing market impact and
realizing more favorable prices. Effective trade execution strategies aim to
lower the overall transaction costs. These costs include not only explicit costs
like commissions and fees but also implicit costs such as slippage and opportu-
nity costs. By optimizing the execution process, traders can significantly reduce
these costs, and increase their net returns.
The execution prices of trades have a direct influence on the success of a trad-
ing strategy. Effective trade execution ensures transactions are carried out at the
most favorable prices possible, thereby reducing slippage – the gap between
the anticipated trade price and the price at which the trade is actually executed.
This is particularly important in fast-moving markets where prices can change
rapidly. The quality of trade execution is critical for large orders. Poor exe-
cution can result in significant deviations from the expected trade price,
adversely affecting the overall trading strategy. Implementing sophisticated
execution algorithms and techniques can improve execution quality, ensuring
that transactions are executed at or close to the target price levels.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 133

One of the foundational works in trade execution is the study of Bertsimas

and Lo (1998). They introduce a framework for minimizing the cost of execut-
ing large orders by considering the trade-off between market impact and price
risk. They use dynamic programming approaches to determine the optimal
trade execution path and highlight the importance of considering both tem-
porary and permanent market impact when executing large trades. This work
lays the foundation for modern trade execution strategies by formalizing the
optimization problem faced by traders.
Another important work is Almgren and Chriss (2001). They develop a
model to optimize trade execution by balancing market impact costs against the
variance of the execution price. This work introduces the concept of an efficient
frontier in trade execution, where different strategies can be evaluated based
on their cost-risk profiles, providing a quantitative basis for the development
of execution algorithms that are used in practice. Their framework has become
a cornerstone in the field, influencing both academic research and practical
implementations of execution algorithms.
In Gatheral (2010), the authors extend previous models by incorporating
more realistic assumptions about market impact and price dynamics. Their
work provides deeper insights into the temporal evolution of market impact,
helping traders to develop more effective execution strategies over longer
time horizons. The work of Obizhaeva and Wang (2013) presents a model
that incorporates supply and demand dynamics in determining optimal trading
strategies. They also suggest optimal execution paths that adapt to changing
market conditions and liquidity.
The integration of deep learning models into trade execution leverages
high-frequency data and sophisticated algorithms to optimize execution strate-
gies further. This has led to the concept of Deep Reinforcement Learning
(DRL), a branch of machine learning that merges reinforcement learning (RL)
(Sutton & Barto, 2018) with deep learning. DRL takes advantage of deep neu-
ral networks to understand complex representations and to make decisions
based on these representations in environments where the results of actions are
both uncertain and delayed. This framework fits the problem of trade execution,
which is essentially a classical sequential decision-making process. Our goal
is therefore to find an optimal order placement strategy that aims to optimize
some evaluation metrics, such as minimizing transaction costs without causing
adverse market impact.
We now briefly introduce RL and discuss several works that apply DRL to
trade execution. RL provides a framework in which agents are trained to make a
series of decisions by interacting with their environment (shown in Figure 39).
Specifically, at any time t, an agent receives some representations (St ) of current

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 134 Quantitative Finance

Figure 39 A schematic description of RL.

environments and takes an action (At ) based on the observed information. This
action either leads to a reward (Rt ) or a penalty that indicates the goodness of the
chosen action. The agent then moves to the next state (St+1 ), and this procedure
continues until the environment concludes. Throughout, the agent’s objective
Í
is to maximize the expected total rewards (E( Rt )).
DRL combines the components of RL with deep neural networks to learn
complex state spaces and effective policies from high-dimensional inputs.
There are a range of DRL algorithms. Deep Q-Networks (DQNs) mark a
major advancement in reinforcement learning by integrating Q-learning princi-
ples with the robust function approximation abilities of deep neural networks.
Traditional Q-learning, which is a model-free reinforcement learning method,
depends on a Q-table to record and update Q-values for every state-action
combination. A Q-value represents the network’s estimate of the expected
discounted sum of future rewards when taking a specific action in a given
state according to the current optimal policy. However, in its traditional form,
this technique becomes unmanageable in environments with extensive or con-
tinuous state spaces because the memory and computational demands grow
exponentially.
DQNs address this challenge by using deep neural networks to approximate
the Q-value function, enabling them to process high-dimensional inputs such as
images or intricate market data. A key advancement in DQNs is the implemen-
tation of experience replay. This method involves retaining past interactions
in a replay buffer and randomly re-selecting mini-batches of these experiences
during training. By disrupting the temporal sequences in the data, experience
replay helps to stabilize the learning process and enhance the algorithm’s over-
all performance. Furthermore, DQNs incorporate a target network, which is
an intermittently updated replica of the Q-network. This target network pro-
vides consistent target values for training, thereby improving the stability and

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 135

convergence of the learning process. By incorporating these techniques, DQNs

have achieved remarkable success across various domains, such as playing
Atari games at superhuman levels. Their capabilities also hold great promise
for applications in fields like finance, robotics, and beyond.
Policy Gradient methods constitute another category of reinforcement learn-
ing algorithms that directly optimize the policy by modifying the policy func-
tion’s parameters to maximize the expected reward. Value-based approaches
create policies by estimating value functions. In contrast, policy gradient tech-
niques parameterize the policy itself – commonly with a neural network – and
then adjust these parameters in a manner that increases the expected reward.
This straightforward method of policy optimization provides several bene-
fits, including more efficient management of high-dimensional and continuous
action spaces.
One key method within policy gradient techniques is the REINFORCE
algorithm. This algorithm utilizes Monte Carlo sampling to approximate the
gradient of the expected reward relative to the policy parameters and then
updates these parameters through gradient ascent. More sophisticated strate-
gies, like Actor-Critic algorithms, integrate policy gradient methods with value
function approximation. This combination helps to lower the variance in gradi-
ent estimates, resulting in more stable and efficient learning processes. Policy
gradient methods are widely applied to complex tasks like robotic control, game
playing, and financial trading, where the ability to directly optimize the policy
offers significant advantages in terms of flexibility and performance.
DQNs and policy gradient methods form the basis of DRL. More advanced
algorithms tend to be extensions of these two approaches. Some well-known
techniques, such as Deep Deterministic Policy Gradient (DDPO) and Proxi-
mal Policy Optimization (PPO), offer unique advantages and address different
challenges in DRL. DDPG is built to handle environments with continuous
action spaces and utilizes an actor-critic architecture. In this framework, the
actor network is responsible for selecting actions, while the critic network eval-
uates their performance. To stabilize learning, DDPG uses experience replay
and target networks, which help reduce correlations in the training data and
smoothen the update process. PPO, on the other hand, aims to simplify the pol-
icy optimization process while ensuring stability. It strikes a balance between
exploration and exploitation by clipping the probability ratios between the new
and old policies during updates. This prevents excessively large updates that
can destabilize learning, making PPO robust and widely applicable to various
problems.
In the study by Nagy, Calliess, and Zohren (2023), DRL combined with
experience replay is utilized to train a trading agent with the objective of

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 136 Quantitative Finance

Figure 40 A brief comparison between the baseline strategy and RL policy

for AAPL on 2012-06-14. New limit orders that are not immediately executed
are represented by circles, executed trades by crosses, and order cancellations
by triangles. Lines connect open orders to their corresponding cancellations
or executions.

maximizing trading returns. The findings reveal that the RL agent formulates an
effective strategy for inventory management and order placement, surpassing a
heuristic benchmark trading strategy that employs the same signals. Figure 40
illustrates a 17-second segment from the testing period, comparing the baseline
strategy with the RL approach. The first two panels show the highest bid, low-
est ask, and mid-prices, alongside trading activities for buy orders (highlighted
in green) and sell orders (highlighted in red). Since the simulation encom-
passes the entire LOB, the influence of trading actions on bid and ask prices is
observable. The third panel depicts the progression of inventory positions for
both strategies, and the final panel displays the trading profits in USD over the
duration of the period.
The findings indicate that both strategies impact the prices within the LOB
by introducing new order flows into the market. These new orders interact
with existing ones, thereby influencing liquidity at the top bid and ask levels.
Throughout the examined timeframe, the baseline strategy experiences minor
losses attributed to frequent changes in its signals which alternate between
anticipating declining and rising future prices. This behavior results in aggres-
sive trading, causing the strategy to incur the spread cost with each transaction.
On the other hand, the RL strategy outperforms by employing a more subdued
approach. This minimizes the effects of market volatility while allowing the RL
strategy to effectively manage its positions. It trades prudently when exiting
long positions and makes strategic decisions when establishing new ones. In

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 137

Figure 41 The distribution of executed volume per time step, with the
horizontal axis representing the time step, vertical axis indicating the volume,
and columns corresponding to different execution strategies. The box plots
show the interquartile ranges, medians (marked by orange lines), means
(indicated by blue triangles), and the 10th and 90th percentiles (represented
by whiskers).

the latter part of the observed period, the RL strategy notably increases its pas-
sive buy orders (depicted as green circles in the second panel of Figure 40).
These orders are connected by green lines to their respective executions or
cancellations, with some actions occurring beyond the timeframe shown in the
figure.
To further present how different DRL algorithms affect execution paths,
we take an example from Schnaubelt (2022) that optimizes order place-
ments on cryptocurrency exchanges. Figure 41 illustrates the executed volume
across various time steps for four different strategies: submit-and-leave (S&L),
backwards-induction Q-learning (BQL), deep double Q-networks (DDQN),
and proximal policy optimization (PPO). Several consistent patterns are
observed in the average executed volume fractions. Firstly, a substantial portion
of the volume is typically executed in the final time step, which usually involves
completing any remaining volume through a market order. Secondly, when
analyzing the volume fractions within the first three time steps, the majority of

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 138 Quantitative Finance

the execution generally occurs in the initial step. Thirdly, as the initial volume
v0 increases, the volume executed in the final time step also rises, while the
volume fraction executed in the earlier steps tends to decrease. These trends
can be attributed to the limited liquidity available during the initial time steps.
When comparing various execution strategies, it becomes apparent that
the S&L method handles a smaller portion of the volume within the first
three time steps compared to the deep reinforcement learning approaches PPO
and DDQN. Although the S&L strategy maintains a positive average volume
fraction, its median fraction is zero across all three initial time steps. In con-
trast, both DDQN and PPO agents exhibit similar execution patterns, with the
majority of the volume being carried out in the first time step.

7.4 Generative Models for Limit Order Books

In the section’s final part, we look at generative models, an expanding area
within machine learning that has recently attracted considerable interest.
Generative models are statistical frameworks designed to produce new data
instances that closely mimic the distribution of an existing dataset. Unlike dis-
criminative models, which aim to classify or predict outcomes based on input
data, generative models focus on learning the joint probability distribution
of the data. This capability allows them to generate realistic and innovative
data samples that align with the inherent patterns and structures present in the
training dataset.
Generative models encompass a broad range of applications and are ben-
eficial for enhancing data availability and quality across multiple fields. In
numerous disciplines, acquiring extensive, high-quality datasets for training
machine learning models is often challenging due to factors such as pri-
vacy issues, substantial costs, and data access restrictions. Generative models
address these obstacles by producing realistic synthetic data that accurately
reflects the statistical properties of the original datasets. This synthetic data can
supplement existing datasets, resulting in the development of more robust and
precise machine learning models. For instance, in the healthcare sector, gen-
erative models can generate synthetic patient records that preserve the critical
patterns found in real data while ensuring patient privacy is maintained.
In the creative arts, generative models can produce innovative images,
music, and artwork, expanding the possibilities of digital creativity. In sci-
entific research, such models can simulate experiments, predict molecular
structures, and generate new hypotheses, helping to accelerate discovery and
innovation. Furthermore, generative models enable the creation of person-
alized experiences across various applications. In recommendation systems,

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 139

they can simulate user preferences and generate tailored content suggestions,
improving user satisfaction and engagement. In gaming, generative models can
craft personalized environments and narratives suited to each player’s pref-
erences. By utilizing the power of generative models, developers can create
more customized and engaging experiences, boosting user satisfaction and
retention.
For high-frequency microstructure data, we can use generative models to
enhance simulations by generating realistic, high-fidelity data that is accu-
rately representative of complex financial markets. This is particularly useful
for modeling market impact as such interactions are difficult to simulate with
static historical data. Furthermore, we can use high-quality synthetic data to
study the problem of regime shift, a notorious problem for financial time-series
that often leads to overfitting and poor generalization. By improving the model-
ing of market dynamics, generative models enhance decision-making processes
and improve risk management.
The roots of generative modeling lie in traditional statistical methods, which
focus on modeling the underlying distributions of data. Some of the foun-
dational approaches include Gaussian Mixture Models (GMMs) and Hidden
Market Models (HMMs). GMMs represent data as a mixture of multiple Gauss-
ian distributions, each capturing a different aspect of the data distribution.
GMMs are effective for clustering and density estimation but struggle with
high-dimensional data. HMMs are used to model sequential data, where the
data-generating process is assumed to follow a Markov process with hid-
den states. They are frequently applied in speech recognition and time-series
analysis, but they struggle to capture complex dependencies.
Advancements in deep learning algorithms have profoundly transformed the
generative modeling landscape over the past several years, shifting it from con-
ventional statistical approaches to advanced deep learning frameworks. This
progression has been fueled by the demand for models that are more pre-
cise, efficient, and capable of generating complex data. Neural networks, with
their proficiency in learning intricate representations, have been instrumental
in developing more robust and adaptable generative models.
There are several remarkable works that leverage the power of deep networks
to provide a new paradigm for generative modeling. Variational Autoencoders
(VAEs) introduced by Kingma and Welling (2013) combine principles from
Bayesian inference and neural networks. They use an encoder-decoder archi-
tecture to learn a probabilistic representation of data, enabling efficient gener-
ation of new samples. VAEs marked a significant step forward in generating
realistic data while providing a solid theoretical foundation.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 140 Quantitative Finance

Figure 42 Price trajectories and the associated percentiles of terminal prices

for both real and generated data.

Generative Adversarial Networks (GANs) are a pivotal development

that has significantly advanced the field of generative modeling (Goodfel-
low et al., 2014). A GAN is composed of two distinct neural networks: the
generator and the discriminator. These networks engage in a competitive min-
imax game, in which the generator creates synthetic data samples and the
discriminator evaluates their authenticity. Through this adversarial training
process, GANs are capable of producing highly realistic and convincing data.
One of the most significant applications of generative models has been to
study LOB market dynamics, a task that is generally assumed to be very diffi-
cult. Understanding and modeling market dynamics is important for studying
market impact and avoiding adverse price movements. In Cont, Cucuringu,
Kochems, and Prenzel (2023), the authors introduce a nonparametric method
for modeling the dynamics of a limit order book by utilizing a GAN. Given
time-series data obtained from the order book, this GAN is trained to learn the
conditional distribution of the LOB’s future state based on its current state.
Figure 42 presents an example of simulated LOBs. In it, we can see that both
distributions of generated and real price paths align closely. The right side of the
figure illustrates the percentiles of the terminal prices, where the distribution
of price changes over 200 transitions is generally well-matched, especially for
the middle percentiles (from 5% to 95%). However, noticeable discrepancies
appear at the extreme points, corresponding to the 1% and 99% quantiles, rep-
resenting the minimum and maximum values. This suggests that the generated
paths do not capture the same tail characteristics as the real data. Although
the 0% and 100% quantiles are often noisy, this deviation remains a consistent
observation.
Nagy, Frey, et al. (2023) introduces an alternative approach to simulate
LOB with an end-to-end auto-regressive generative model that directly gen-
erates tokenized LOB messages. Figure 43 juxtaposes the return distributions

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 141

Figure 43 The distributions of mid-price returns for generated (blue) and

realized (red) data with the mean (solid lines) and 95% confidence intervals
(shaded regions). Left: Google; Right: Intel.

of the generated data with those of the actual realized data over the span of 100
future messages. The findings demonstrate that the model effectively mirrors
the mid-price return distributions, even though these were not directly included
in the training loss function. The average returns exhibit no significant drift or
trend, and the shaded areas, representing the 95% confidence intervals of the
distributions, align closely.
To further test the authenticity of the generated data, returns are sampled
from the generative model, and correlation is calculated between the gener-
g r for 100 future messages (s ∈
ated returns rt+s and the realized returns rt+s
[1, · · · , 100]). As shown in the top of Figure 44, there exists a consistently pos-
itive correlation for both Google (ρ ≈ 0.1) and Intel (ρ ≈ 0.2). The lower
panel displays the corresponding p-values from t-tests evaluating the alterna-
tive hypothesis H1 : ρ > 0 against the null hypothesis H0 : ρ = 0. The dotted
line represents the 5% significance level. For the Google model, the p-values
remain at or near the 5% threshold for up to 80 future messages, whereas for
Intel, the correlations stay statistically significant for at least 100 messages. The
sustained positive correlation indicates directional forecasting power which
suggests new possibilities for alpha.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 142 Quantitative Finance

Figure 44 Top: Pearson correlation coefficient ρ between the generated and

actual returns, reflecting the performance of directional forecasting; Bottom:
the corresponding p-values. Left: Google; Right: Intel.

By generating realistic LOB data, researchers and industry professionals can

obtain meaningful insights into the behaviors of market participants and the
dynamics of order flow. In turn, this helps them study the execution of large
orders and analyze their impact on the market. In addition, we can use gen-
erative models as environments for DRL algorithms. Historical data is static
whereas the real market is dynamic and reacts to actions. Generative models
create realistic synthetic data, allowing DRL agents to train in a simulated
environment that mirrors real market conditions. This approach not only accel-
erates the development and testing of DRL-based trading algorithms but also
ensures that these algorithms are robust and adaptable to a wide range of market
scenarios.
The prospects for generative models in the financial sector are bright, with
promising avenues for future research and implementation. One other key area
of interest is utilizing generative models for stress testing and scenario analysis.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 143

This application enables financial institutions to better evaluate their ability to

withstand extreme market conditions and build more resilient risk management
systems. Additionally, improving the interpretability of generative models and
developing standardized evaluation metrics will enhance their practical utility
and facilitate their adoption in the financial industry. As these models advance
and become increasingly refined, they are certain to facilitate cutting-edge
research initiatives and play a crucial role in shaping the future of trading and
market analysis.

8 Conclusions
This final section concludes our exploration of the applications of deep learn-
ing to quantitative finance. It aims to summarize key insights from the Element
and discuss future opportunities and challenges in integrating these fields,
providing a foundation for future work.

8.1 Summary and Key Takeaways

Deep learning is revolutionizing contemporary quantitative trading and reshap-
ing the world of financial markets. This Element provides an in-depth analysis
of the methods and models underpinning this development. It also highlights
the capacity of deep learning models to automatically extract complex features,
uncover hidden patterns within extensive financial datasets, and facilitate the
development of more precise and effective trading strategies.
This Element is aimed at quantitative researchers in academia and indus-
try, as well as data scientists and developers interested in the field. It blends
foundational concepts with real-world applications and practical use cases
to demonstrate how these models can be used to automate decision-making,
enhance predictive accuracy, and improve trading performance in dynamic and
high-stakes market environments. We provide a dedicated GitHub repository9
to demonstrate examples included in the Element.
This Element is divided into two main sections: Foundations and Applica-
tions. The first part focuses on the fundamental aspects of financial time-series,
covering topics such as statistical analysis, hypothesis testing, and related
concepts. Financial datasets possess unique characteristics, and a solid under-
standing of their statistical properties is important for conducting meaningful
financial analysis. Following this, we introduce the concept of supervised learn-
ing, along with an overview of deep learning models. The covered concepts

9 See [Link] or [Link]

Trading.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 144 Quantitative Finance

range from simple fully connected layers to the more advanced attention mech-
anism, which is particularly effective in capturing long-range dependencies
within structured datasets.
Although deep learning has achieved significant advancements, deep net-
works frequently face issues such as overfitting, where models perform excep-
tionally well on training data but have difficulty generalizing to new, unseen
datasets. To mitigate this issue, this Element outlines a complete workflow
for implementing deep learning algorithms in quantitative trading. The work-
flow covers crucial stages, including data collection, exploratory data analysis,
and cross-validation methods specifically adapted for financial datasets. These
stages address key aspects like data distribution, stationarity, and the distinc-
tive characteristics of financial time-series. These considerations are critical
for creating models that achieve not only high accuracy but also robustness
and reliability for implementation in real-world trading environments.
The second part of the Element is dedicated to the application of deep learn-
ing algorithms to various financial contexts. It places a key focus on one of the
core tasks in quantitative trading: generating predictive signals. We explore a
range of deep learning architectures designed for this purpose, demonstrating
how these models can effectively forecast market movements. On top of this,
we delve into advanced applications, such as improving momentum trading
and cross-sectional momentum strategies. Additionally, we address portfolio
optimization by introducing methods that enable the direct optimization of port-
folio weights from market data. This end-to-end approach eliminates the need
for intermediate steps, such as estimating returns and working with covari-
ance matrices of returns, which are often difficult to implement in practical
scenarios.
We provide an in-depth examination of the operational dynamics of mod-
ern securities exchanges, illustrating the processes behind financial transactions
and the generation of high-frequency microstructure data, including order book
updates and trade executions. Furthermore, we analyze the unique attributes of
different asset classes, such as equities, bonds, commodities, and cryptocurren-
cies, highlighting the specific challenges and opportunities for applying deep
learning techniques effectively to each.

8.2 Future Possibilities and Challenges

As the convergence of deep learning and quantitative trading progresses, the
field presents immense opportunities alongside significant challenges. Next,
we discuss some areas that are worth future exploration.
In this Element, our primary focus is on time-series data, including prices and
trading volumes. However, we also explore the inclusion of alternative data,
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Deep Learning in Quantitative Trading 145

such as text, and techniques specific to those data types as potential sources
of additional alpha. In Section 3, we look at how recent advances in NLP,
such as transformer-based models like BERT and GPT, have made it feasible
to extract nuanced information from unstructured textual data. Such methods
could be used to evaluate data from news articles, social media, and earnings
call transcripts to inform sentiment analysis and event prediction. Similarly,
computer vision models can be used to analyze visual patterns in images. Prac-
titioners could thus use satellite data, product shelves, or even weather imagery
to provide insights into supply chain activity or predict market trends.
Another interesting area of further research is the explainability of deep net-
works. As deep learning models become increasingly sophisticated, the lack of
interpretability poses challenges to understanding why a model makes specific
decisions. In quantitative trading, where financial stakes and regulatory scru-
tiny are high, explainable algorithms are essential for building trust in model
outputs and avoiding unintended biases. For trading strategies, explainability
should encompass not only technical factors but also ethical considerations. It
is important to ensure that algorithms do not exploit market inefficiencies in
ways that harm retail investors or contribute to systemic risks. For instance, on
May 6, 2010, the U.S. stock market underwent the Flash Crash, during which
the Dow Jones Industrial Average plummeted by nearly 1,000 points within
minutes before swiftly rebounding. This sudden decline was initiated by a sub-
stantial sell order executed by a mutual fund employing a trading algorithm
intended to reduce market impact. The algorithm indiscriminately offloaded a
large volume of E-mini S&P 500 futures contracts, ignoring prevailing prices
and market conditions. HFT algorithms quickly picked up on this activity,
starting a cascade of rapid-fire selling that spread across markets.
Interpretability has already been studied in academia, and methods like
SHAP (Shapley Additive Explanations), Integrated Gradients (IG), and LIME
(Local Interpretable Model-Agnostic Explanations) can be used to provide
insights into model behavior. SHAP assigns each feature a contribution score
for each prediction that indicates that feature’s importance. Differently, IG is an
attribution-based method and assesses the impact of each input feature on the
predicted output by summing the gradients along a path from a baseline to the
input. Similarly, LIME takes an approximation method that adopts a simpler
model to explain individual predictions. Despite their utility, these methods still
face significant challenges that limit their effectiveness in certain contexts. For
example, SHAP can be computationally expensive and LIME relies on local
approximations that may not accurately capture global model behavior. Addi-
tionally, these methods can struggle with capturing interactions among features
in time-series or nonlinear domains, leading to incomplete interpretations.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 146 Quantitative Finance

Accordingly, the interpretability of models still remains a promising research

direction, offering opportunities to develop more robust, efficient, and domain-
specific tools that bridge the gap between complex predictions and actionable
insights.
Quantum computing is poised to revolutionize many fields, and quantitative
trading is no exception. In theory, quantum computers can address specific
types of problems exponentially faster than classical machines, offering effi-
cient parallel processing and the ability to solve high-dimensional challenges.
This can be particularly valuable for tasks such as optimizing portfolio alloca-
tions or identifying high-dimensional nonlinear relationships in market data. By
leveraging quantum-enhanced deep learning algorithms, we have the potential
to optimize model training and explore complex patterns that classical systems
can not handle. However, quantum computing technology is still in its infancy
and access to scalable, fault-tolerant quantum systems is limited. Moreover,
a wide gap between quantum algorithms and deep learning frameworks still
remains. This research area will require interdisciplinary expertise to address
the open questions regarding the practical applicability and cost-efficiency of
quantum systems in trading.
Deep learning’s potential for quantitative trading is vast, offering transform-
ative possibilities for the financial industry. To harness the full power of these
advanced techniques, sustained and focused research is essential. This com-
mitment to ongoing research will allow financial institutions to refine trading
strategies, enhance performance, and adopt these innovations responsibly. It
is equally important that the use of such technologies upholds market integrity
and operates within an ethical framework. It is our hope that this Element serves
as a foundational resource in advancing this shared vision, fostering progress
while contributing to market stability and fairness.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Acronyms
ACF Auto Correlation Function.
AR Autoregressive Model.
ARMA Autoregressive Moving Average Model.
BERT Bidirectional Encoder Representations from Transformers.
BTC Bitcoin.
CAPM Capital Asset Pricing Model.
CBOE Chicago Board Options Exchange.
CDS Credit Default Swaps.
CME Chicago Mercantile Exchange.
CNNs Convolutional Neural Networks.
DDPO Deep Deterministic Policy Gradient.
DeFi Decentralized Finance.
DMNs Deep Momentum Networks.
DOT Designated Order Turnaround.
DQNs Deep Q-Networks.
DRL Deep Reinforcement Learning.
ETF Exchange-Traded Fund.
ETH Ethereum.
FCNs Fully Connected Networks.
FPR False Positive Rates.
FSE Frankfurt Stock Exchange.
FX Foreign Exchange Market.
GANs Generative Adversarial Networks.
GBM Geometric Brownian Motion.
GCNs Graph Convolutional Neural Networks.
GED Generalized Error Distribution.
GMMs Gaussian Mixture Models.
GNNs Graph Neural Networks.
GP Gaussian Process.
GRUs Gated Recurrent Units.
HAR Heterogeneous Autoregressive.
HEAVY High-Frequency Based Auto-regressive and Volatility.
HL Huber Loss.
HMMs Hidden Market Models.
IG Integrated Gradients.
IPOs Initial Public Offerings.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 148 Acronyms

IRD Interest Rate Differential.

Leaky-ReLU Leaky Rectified Linear Units.
LIME Local Interpretable Model-Agnostic Explanations.
LLMs Large Language Models.
LOBSTER Limit Order Book System.
LSE London Stock Exchange.
LSTM Long Short-Term Memory.
MA Moving Average Model.
MACD Moving Average Crossover Divergence.
MAE Mean Absolute Error.
MedAE Median Absolute Error.
MSE Mean Squared Error.
NBBO National Best Bid and Offer.
NYSE New York Stock Exchange.
OTC Over-the-Counter.
PACF Partial Autocorrelation Function.
PDF Probability Density Function.
PMF Probability Mass Function.
PPO Proximal Policy Optimization.
REITs Real Estate Investment Trusts.
ReLU Rectified Linear Units.
RL Reinforcement Learning.
RNNs Recurrent Neural Networks.
ROC Receiver Operating Characteristics.
Seq2Seq Sequence to Sequence Learning.
SHAP Shapley Additive Explanations.
SMA Simple Moving-Average Crossover.
SMBO Sequential Model-Based Optimization.
S&P500 Standard & Poor’s 500.
TPR True Positive Rates.
TSE Toronto Stock Exchange.
VAEs Variational Autoencoders.
VaR Value at Risk.
WRDS Wharton Research Data Services.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Appendix A
Different Asset Classes
We here introduce several key asset classes that are particularly relevant to the
topics discussed in this Element and are widely traded in financial markets.
These asset classes include equities, bonds, foreign exchange (FX), futures,
options, exchange-traded funds (ETFs), and cryptocurrencies, each of which
presents unique characteristics and potential applications of deep learning.
While our focus is on these prominent categories, it is important to note that
this list is by no means exhaustive. Financial markets encompass a broad range
of additional asset classes, such as real estate investment trusts (REITs) and
derivatives like spread-betting, each offering distinct challenges and appli-
cations. Our goal is to establish a basic comprehension of these key asset
classes, allowing readers to better understand the deep learning methods pre-
sented in the Element. Future exploration of other asset classes can further
enrich the contextual knowledge base and expand the practical scope of these
methodologies.

A.1 Equity Markets

Equities are among the most well-known financial securities, representing
ownership stakes in companies. As of December 2024, global equity markets
have reached a total market capitalization exceeding 123 trillion USD, sur-
passing the previous record set in October 2021. When companies decide to
go public through Initial Public Offerings (IPOs), their shares are traded on
major exchanges, such as the New York Stock Exchange (NYSE) or NAS-
DAQ, making them available to a wide spectrum of investors. While equities
are exchange-traded, equity markets are highly fragmented, consisting of pri-
mary exchanges, secondary exchanges, and alternative trading venues like dark
pools. This fragmentation can lead to complexities in price discovery and exe-
cution. Liquidity in these markets is typically facilitated by market makers, who
play the crucial role of ensuring that there is always a counterparty to trades. A
key feature of equity trading in the United States is the National Best Bid and
Offer (NBBO), which requires brokers to execute trades at the best available
bid and offer prices across all venues. This regulation aims to protect investors,
but its implementation can create challenges, particularly in high-frequency
trading environments or during volatile market conditions.
Investors need to be mindful of several issues unique to equities. Delistings,
for example, can significantly impact a portfolio and often lead to survivorship
bias in historical datasets. This bias arises because delisted or failed companies

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 150 Appendix A

are removed from indices and databases, skewing performance analyses. Addi-
tionally, corporate actions such as dividends, stock splits, and mergers must
be properly accounted for in price series to avoid misinterpreting historical
data. Adjusting prices for these actions ensures that analyses and back-tests
accurately reflect the financial realities of investing in equities. Beyond single
stocks, indices such as the S&P 500 and Dow Jones Industrial Average aggre-
gate the prices of multiple equities to monitor the performance of specific areas
within equity markets. These indices are constructed and weighted in various
ways, for example price-weighted, equal-weighted, or market-cap-weighted,
depending upon the index-specific methodology. Indices serve multiple pur-
poses: They provide benchmarks for fund performance, offer insights into
market trends, and serve as tradable instruments themselves.

A.2 Bonds
Bonds are fixed-income instruments that represent loans provided by lenders
to borrowers, typically corporations or government entities. They offer predict-
able income and a range of maturities to suit diverse investment goals. Compa-
nies issue corporate bonds to obtain capital for purposes such as expansion,
operational needs, or refinancing existing debts. Corporate bonds typically
provide higher yields compared to government bonds to compensate for the
increased credit risk taken on by lenders. The risks of these bonds are evaluated
and classified by agencies like Moody’s and Standard & Poor’s. Conversely,
government bonds are issued by national, state, or municipal authorities to
finance public expenditures and infrastructure projects. Bonds from stable
governments, such as U.S. Treasuries, are considered some of the safest invest-
ments, while sovereign debt from emerging markets may carry higher yields but
also greater risk due to economic and political volatility.
The bond market operates primarily in an over-the-counter (OTC) format,
where trades are negotiated directly between buyers and sellers rather than on
centralized exchanges. This OTC structure allows for flexibility in terms of
transactions but often results in lower transparency compared to equity markets.
Despite this, the bond market is enormous, with an estimated global market size
exceeding 130 trillion USD. This valuation underscores its importance along-
side equities as a cornerstone of financial systems worldwide. Instead of trading
OTC, investors can gain exposure to bonds in several other ways. ETFs pro-
vide a simple and efficient method for individuals to access a diversified basket
of bonds. These ETFs track indices composed of corporate, government, or
municipal bonds and allow investors to trade bond exposure on stock exchanges
with ease. Bond futures provide an alternative means for investors to protect

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Appendix A 151

against interest rate fluctuations or to engage in speculative trading based on

price changes. These instruments are widely used by institutional investors
because of their liquidity and leverage. Mutual funds and index funds focused
on bonds also provide a way to access professionally managed portfolios,
offering diversification and, in the case of index funds, cost efficiency.
When investing in bonds, several factors must be considered. Interest rate
sensitivity is a crucial aspect, as bond prices have an inverse relationship
with interest rates. Rising rates can thus lead to capital losses, especially for
bonds with longer maturities. Credit risk is another key consideration, as the
probability of an issuer failing to meet its obligations changes significantly
between corporate and government bonds. Liquidity can also be a concern,
while government bonds are generally liquid, corporate bonds might encoun-
ter liquidity issues, especially during times of market turmoil. Overall, bonds
are essential components of a well-diversified portfolio, especially given their
relatively low correlation. Bonds are generally favored by investors seeking
stable income, capital preservation, and reduced portfolio volatility. Whether
accessed through OTC trading, ETFs, futures, or mutual funds, bonds provide
versatile tools to meet a range of investment objectives. Their importance in
financial markets cannot be overstated, as they continue to serve as a foundation
of income and stability for both individual and institutional investors.

A.3 Foreign Exchange Market

The foreign exchange market, often referred to as Forex or FX, is the largest and
most liquid financial market globally, with over 7 trillion USD in daily trades.
Forex market participants trade currencies for purposes of international trade,
investment flows, and speculative interests. Unlike equities and bonds, Forex
does not trade on centralized exchanges. Instead, trading occurs OTC, with par-
ticipants ranging from large institutions and governments to retail traders. This
structure ensures the market is active 24 hours a day, spanning key financial
centers including London, New York, Tokyo, and Sydney.
Forex trading entails simultaneously purchasing one currency while selling
another, forming currency pairs such as EUR/USD or GBP/JPY. These pairs are
divided into multiple categories, notably including majors, minors, and exotics.
Majors include the most traded currencies globally, such as the U.S. Dollar
(USD), Euro (EUR), and Japanese Yen (JPY). Minors exclude the USD but
include other major currencies. Lastly, exotics consist of less liquid and more
volatile currencies from emerging markets.
Market participants engage in Forex for different reasons. Corporations and
governments trade currencies to manage and hedge their exposure to currency

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 152 Appendix A

price fluctuations that might affect international trade or the value of their
reserves. Financial institutions, hedge funds, and retail traders often engage
in Forex for speculative purposes, seeking to profit from changes in exchange
rates. Exchange rate movements are influenced by multiple elements such as
interest rate disparities, geopolitical incidents, economic indicators, and central
bank strategies. Leverage plays an important role in Forex trading by allowing
traders to hold positions substantially larger than the amount of capital they
commit as collateral. This magnifies both possible gains and associated risks.
Accordingly, the high levels of leverage available in Forex can lead to signifi-
cant losses, particularly for inexperienced traders. Moreover, the decentralized
and largely unregulated nature of the market means participants should choose
brokers carefully to ensure transparency and fair dealing.
Access to the Forex market has been democratized significantly through
technology, allowing retail traders to participate via online platforms. These
platforms provide traders with exposure to Forex markets through spot trading,
forward contracts, and options. Beyond direct trading, investors can also gain
exposure to currency movements through ETFs that track the performance of
currency indices or specific currency pairs. Futures contracts on major curren-
cies offer yet another way to speculate or hedge currency exposure, providing
a regulated alternative to OTC Forex trading.

A.4 Futures
Futures are standardized financial agreements that require a buyer to purchase,
or a seller to sell, an underlying asset at a set price on a designated future
date. Futures are essential instruments in global financial markets, used both
for speculation and hedging against price movements. The total market for
futures is vast, spanning financial instruments, commodities, and more. The
Chicago Mercantile Exchange (CME) is the most prominent and liquid futures
exchange globally. Futures contracts are intrinsically tied to “future deliver-
ables,” meaning a contract specifies the terms for the delivery of the respective
underlying asset at expiry. However, in practice, most futures contracts are
either cash-settled or closed out prior to delivery, particularly for financial
futures where physical delivery is less common. The ability to settle contracts in
cash adds flexibility for traders and investors, reducing the logistical challenges
associated with taking physical delivery of assets, such as oil or agricultural
products.
Each futures contract has a specific expiry date, and traders often need to
“roll” contracts if they wish to maintain their position beyond that expiration
date. Rolling consists of closing the position in the nearing expiration contract

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Appendix A 153

and simultaneously establishing a position in a future-dated contract. This proc-

ess is subject to price differences between contracts, particularly in markets
with contango or backwardation. Contango and backwardation refer to how
current futures prices compare to the spot (current) price of an underlying asset.
Contango means that the futures price is higher than the current spot price,
and backwardation is the opposite. Thus, to maintain a consistent price series
for analysis, adjustments are often required. Backward adjustments, as dis-
cussed in Section 5, modify historical prices to account for differences between
successive contracts, ensuring continuity in time-series data. However, these
adjustments distort historical price levels. Consequently, traders and investors
must carefully manage their positions, particularly around contract expiry,
and be aware of the implications of rolling contracts and adjusting historical
prices.
Futures are particularly popular due to their high liquidity, low trading fees,
and the leverage that can be exercised through margin accounts. Using margin,
traders can manage large positions with only a small portion of their capital,
thereby increasing both potential profits and risks. This leverage, combined
with the standardized nature of contracts, makes futures a preferred choice for
both retail and institutional traders. Among the most traded futures contracts
globally are the S&P 500 futures on the CME, which allow investors to gain
or hedge exposure to the performance of the broader U.S. equity market. Addi-
tionally, futures are not limited to financial instruments and they are widely
used in commodities markets, covering assets such as crude oil, gold, agricul-
tural products, and natural gas. These contracts enable producers, consumers,
and traders to hedge against price volatility and secure future prices. However,
the popularity of certain commodity futures can decline over time and lead to
their delisting, as similarly occurs in equity markets.

A.5 Options
Options are financial derivatives that provide the holder with the right, but
not the obligation, to buy or sell an underlying asset at a predetermined price,
known as the strike price, on or before a specific expiration date. Similar to
futures, options are extensively utilized for hedging, speculative activities, and
income generation. They are traded on centralized exchanges such as the Chi-
cago Board Options Exchange (CBOE) and OTC markets. The options traded
on these exchanges are standardized to ensure greater liquidity and transpar-
ency. There are two primary types of options: call options and put options. A
call option grants the holder the right to purchase the underlying asset, while a
put option allows the holder to sell it. Each option contract requires the payment

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 154 Appendix A

of a premium, which is the cost the buyer pays to the seller for the rights the
option provides. The value of an option is influenced by various factors, includ-
ing the price and expected volatility of the underlying asset, the time remaining
until expiration, and prevailing interest rates.
Unlike futures, which impose mandatory obligations, options offer greater
flexibility. The purchaser of an option has the discretion to decide whether to
exercise the contract, whereas the seller (or writer) must adhere to the con-
tract terms if the buyer chooses to exercise it. Options are available on a wide
range of underlying assets, such as stocks, indices, commodities, currencies,
and even interest rates. For example, an investor with a stock portfolio might
buy put options to protect against a potential drop in stock prices. Similarly, a
business exposed to fluctuating commodity prices might purchase call options
to guarantee maximum costs for raw materials. Speculators use options to profit
from anticipated price movements, benefiting from the contracts’ relatively low
upfront cost compared to that of the underlying asset to gain leverage to price
movements.
However, the flexibility of options comes with complexity. A key character-
istic of options is their expiration date, after which a contract expires worthless
if not exercised. This creates the need for strategic decision-making around
whether and when to exercise an option. Rolling options or closing a posi-
tion in a near expiry option and simultaneously opening a new position in a
longer-dated contract is a common practice to maintain exposure beyond an
approaching expiration date. Another key feature of options is the leverage
they offer. A slight fluctuation in the price of the underlying asset can cause
large percentage changes in an option’s value. This leverage can amplify both
profits and losses, requiring careful position sizing and risk management.
Options trading has grown significantly in popularity, driven by technolog-
ical advancements and the rise of retail trading platforms. Exchange-traded
options, such as those on major indices like the S&P 500, are among the most
actively traded due to their high liquidity and broad appeal. Meanwhile, OTC
options allow for customized contracts tailored to specific needs, but they come
with less transparency and higher counterparty risk.

A.6 Exchange-Traded Funds

Exchange-traded funds (ETFs) are investment instruments that blend the diver-
sification benefits of mutual funds with the trading flexibility of individual
stocks. They are structured to mirror the performance of a specific index,
sector, commodity, or asset class, providing investors with an easy and cost-
efficient method to access a broad array of markets. Over the years, ETFs have

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Appendix A 155

revolutionized the world of passive investing, contributing to a significant shift

in how capital is allocated across global markets. The rise of passive investing
through ETFs has been remarkable. From their inception in the early 1990s,
ETFs have grown exponentially, with global assets under management now
exceeding 10 trillion USD. This increase highlights the trend that investors are
increasingly favoring investment strategies that are cost-effective, transparent,
and efficient, rather than traditional active management. ETFs make it possi-
ble for investors to gain exposure to broad market indices like the S&P 500
or MSCI World Index without the need for individual stock selection, making
them ideal for those seeking diversification with minimal effort.
Similar to individual stocks, ETFs are traded on exchanges, providing
investors with the benefits of liquidity and flexibility. Unlike mutual funds
which are priced solely at the end of the trading day, ETFs can be purchased
and sold throughout the trading session. Market makers critically ensure liq-
uidity and facilitate smooth intraday trading in ETF markets. They do so by
buying and selling both ETF shares and the underlying securities that an ETF
tracks. These market makers are also responsible for delivering the underlying
“basket” of securities at the end of each trading day. The arbitrage mech-
anism maintains a tight alignment between ETF prices and their net asset
value (NAV). Beyond traditional market-cap-weighted ETFs, the industry has
seen significant innovation, particularly in the realm of smart beta (a better
risk-adjusted beta). Smart beta ETFs deviate from conventional indices by
weighting their components based on alternative criteria such as value, momen-
tum, or volatility. These methods are designed to take advantage of certain risk
premiums or enhance returns while retaining the benefits of transparency and
cost-effectiveness.
The ETF market has transformed the investment landscape, making it eas-
ier than ever for individuals to access a wide array of assets and strategies.
Whether used for passive exposure to global markets or tactical allocations
through smart beta, ETFs continue to shape modern portfolio management.
Their combination of simplicity, efficiency, and adaptability suggests that they
will remain a cornerstone of investment portfolios worldwide.

A.7 Cryptocurrency
Cryptocurrencies are digital or virtual currencies that are built upon decen-
tralized blockchain technologies and leverage cryptography to ensure their
security. They possess several features that distinguish them significantly
from traditional major asset classes. Primarily, cryptocurrencies are usually
not governed by any central authority, making them resistant to government

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 156 Appendix A

interference or manipulation. Most decentralized digital assets do not provide

ownership in any underlying entity or physical asset, whereas stocks, for
example, represent fractional ownership in a company that provides share-
holders with a portion of the company’s profits along with voting privileges.
Additionally, the transactions of cryptocurrencies are generally recorded on a
public ledger that is accessible to anyone. Thus, while many equity transac-
tions are not public and corporations only intermittently disclose information,
blockchain-based currencies offer much greater, real-time transparency
Cryptocurrency markets operate 24/7 on both centralized and decentralized
exchanges. Centralized exchanges such as Binance, Coinbase, and Kraken offer
spot, derivative, and staking markets. However, centralized exchanges come
with inherent counterparty risks as the collapse of FTX (one of the largest
cryptocurrency exchanges). The fraud and mismanagement at FTX led to cat-
astrophic losses for its users. Such incidents underscore the importance of
due diligence and risk management when dealing with centralized platforms.
Compounding these challenges is the fragmentation of cryptocurrency markets
across thousands of cryptocurrencies and a multitude of exchanges. This disper-
sion can lead to liquidity issues, pricing discrepancies, and security concerns,
making the cryptocurrency ecosystem both vibrant and complex.
Decentralized Finance (DeFi) offers an alternative to centralized exchanges
by using blockchain-based smart contracts to enable financial transactions
without intermediaries. Platforms like Uniswap and Aave demonstrate how
DeFi can facilitate peer-to-peer trading and lending with enhanced transparency
and reduced reliance on centralized entities. This approach mitigates counter-
party risk while expanding access to financial services, though it introduces
new considerations such as smart contract vulnerabilities.
Spot trading forms the backbone of cryptocurrency markets, where digital
assets like Bitcoin (BTC) and Ethereum (ETH) are bought and sold directly
at current market prices, with immediate settlement. Beyond spot trading, the
market offers a range of derivative instruments, including futures contracts
and perpetual futures. Perpetual futures, in particular, have gained widespread
popularity due to their unique design, allowing traders to maintain leveraged
positions indefinitely without an expiration date, as long as margin require-
ments are met. These products are heavily traded on platforms like Binance
and Bybit, offering significant liquidity and trading opportunities.
For investors who prefer not to directly hold cryptocurrencies, there are sev-
eral other ways to gain exposure to their price movements. Spot ETFs provide
a regulated avenue for tracking the value of assets like BTC and ETH with-
out the need for direct ownership. Futures-based products, which track Bitcoin
or Ethereum, also offer a way to participate in the market, though they carry

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Appendix A 157

risks associated with rolling contracts. These financial products bridge the gap
between traditional investment frameworks and the digital asset space, making
it easier for institutional and retail investors to enter the market.

A.8 Others
Besides the major asset classes, there are numerous other products that pro-
vide investors with diverse opportunities to achieve their financial goals. We
list a few here. Commodities, although often traded via futures, also exist as
a standalone asset class. This category encompasses physical goods such as
gold, silver, crude oil, natural gas, and agricultural items like wheat and corn.
Investors can participate in the commodities market by owning these assets
directly, engaging in futures contracts, or investing in ETFs that track com-
modities. Commodities are particularly valued for their role as inflation hedges
and their historically low correlation with traditional financial assets, making
them useful for portfolio diversification.
While futures and options are the most frequently traded derivatives, other
types of derivatives also hold significant importance in financial markets. For
instance, swaps are extensively utilized in the interest rate and currency sec-
tors. Interest rate swaps enable parties to exchange fixed-rate payments for
floating-rate payments, or the other way around, allowing them to manage their
exposure to interest rate variability. Currency swaps involve the exchange of
principal and interest payments in different currencies, serving as essential tools
for multinational corporations and governments to manage foreign exchange
risks. Additionally, credit default swaps (CDS) function as a type of insurance
against a borrower’s default, playing a crucial role in credit markets and risk
management strategies.
The real estate market is another prominent market. Often, the cost of buy-
ing or selling a property is high and the process is time-consuming. However,
real estate investment trusts (REITs) offer investors a way to access the real
estate market without directly owning physical properties. By pooling funds
from multiple investors, these investment entities can purchase, oversee, and
finance income-producing real estate assets such as office buildings, retail cen-
ters, apartment complexes, and industrial facilities. Publicly traded REITs are
listed on stock exchanges, with liquidity and ease of access similar to that of
equities. Private non-traded REITs are also available to accredited investors and
often focus on niche markets. REITs attract income-oriented investors due to
legal requirements that compel them to pay out a significant share of their earn-
ings as dividends, typically offering higher returns than conventional equities.
Nonetheless, their success can be impacted by factors like interest rate changes,

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 158 Appendix A

trends in the property market, and economic cycles, rendering them sensitive
to macroeconomic shifts. Note that some dividends from REITs are qualified
as capital gains rather than income which can get more favorable taxation.
Private equity and venture capital represent another distinct asset class, pro-
viding opportunities to invest in companies that do not trade on public markets.
Venture capital targets early-stage, high-growth startups, while private equity
focuses on mature companies, often involving buyouts or growth investments.
These investments usually involve committing capital over an extended period
and bearing higher risks, but they also provide the potential for considerable
returns.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Appendix B
Access to Market Data
B.1 Professional
For professionals, there are a variety of established third-party providers that
deliver high-quality market data, tailored to the needs of institutional investors,
traders, and financial analysts. Providers such as Bloomberg, Refinitiv, and
S&P Global offer comprehensive datasets spanning multiple asset classes,
along with advanced analytical tools and integration options. These plat-
forms have become industry staples, ensuring reliable and timely access to
financial information critical for decision-making. In addition to third-party
providers, many professional market participants access direct market feeds
from exchanges. These feeds deliver raw, real-time data, including order book
details, trade executions, and price updates, providing the low-latency access
required for high-frequency trading and algorithmic strategies. Beyond tra-
ditional market data, there is also a growing demand for alternative data,
non-conventional datasets that provide unique insights into market trends and
behavior. This can include information from annual reports, social media sen-
timent, credit card transactions, and forum discussions. Alternative data has
become a critical tool for gaining a competitive edge, offering perspectives not
available from standard financial datasets. Together, these resources constitute
the standard data sources accessed by professionals.

B.2 Academic
Academics often have access to subsidized data sources formatted specifi-
cally for research and academic use. These resources can be tailored to meet
the requests of universities and researchers. Such data is often used to study
financial markets, corporate behavior, and economic trends.
One of the primary resources for academics is Wharton Research Data Ser-
vices (WRDS), a global data platform that hosts a huge amount of data (more
than 350 TB) aggregated from global data vendors. WRDS encompasses a
range of databases including Compustat, CRSP, TFN (THOMSON), TAQ and
many others. WRDS not only covers historical financial time-series data but
also provides access to corporate fundamentals, macroeconomic indicators and
more. Most academics can gain access to the WRDS platform through a sub-
scription provided by their universities. As a result, WRDS is widely used in
academic research.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 160 Appendix B

Another notable resource is LOBSTER (Limit Order Book System), which

was established in 2013 and offers high-frequency limit order book data. Since
its inception, the focus of LOBSTER has shifted to serving the academic com-
munity by providing reconstructed limit order book data for all the stocks traded
on NASDAQ. Accordingly, the LOBSTER dataset is derived from NASDAQ’s
Historical Total View ITCH files and encompasses the complete depth of the
order books. Each historical snapshot includes the bid and ask data for up to
200 price levels, with the number of price levels varying based on the specific
security. This data source is particularly suited to researchers who are interested
in market microstructure, trading dynamics, order flow and market efficiency.
In addition to these platforms, academics may also access resources like
Quandl for economic and financial data, S&P Capital IQ for corporate insights,
and DataStream for macroeconomic and time-series data. Many universities
partner with these providers to offer discounted or free access to their students
and faculty. For niche research needs, open-source data repositories, govern-
ment databases, and exchange-specific resources can also provide valuable
datasets.

B.3 Personal Enthusiast

For personal enthusiasts, accessing market data is easier than ever, with a vari-
ety of platforms and tools offering low-cost or even free data solutions. These
resources cater to hobbyists, retail investors, and independent developers who
want to explore financial markets, test strategies, or simply stay informed.
Cryptocurrency market data is likely a good starting point for enthusiasts since
it is easy to obtain high-quality market data from cryptocurrency exchanges like
Binance that deliver real-time data directly to users through REST, WebSocket,
and FIX APIs. Additionally, other platforms like CoinGecko and CoinMarket-
Cap offer extensive information on prices, trading volumes, and blockchain
statistics at no cost. This easily accessible high-quality market data places retail
investors on a more level playing field with professionals.
For other asset classes, such as equities, fixed-income securities, and ETFs,
other popular free platforms like Yahoo Finance, Google Finance, and Trad-
ingView provide access to live prices, historical charts, and basic financial
metrics. These platforms are user-friendly, making them ideal for beginners
who want to track markets and learn about investing without significant
upfront costs. For enthusiasts looking for more advanced data, platforms like
Alpha Vantage, [Link], and Finnhub offer affordable APIs that deliver
market data in customizable formats, including live prices and fundamental

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Appendix B 161

information. These tools are especially popular among developers and quanti-
tative enthusiasts who want to integrate financial data into their own projects
or build custom trading algorithms.
Finally, for those interested in more niche or alternative datasets, open-
source repositories and public APIs from organizations like AlphaQuery and
Kaggle can offer unique insights and opportunities for experimentation. The
wide availability of these tools ensures that personal enthusiasts have plenty of
options to explore financial markets, regardless of their respective experience
levels or budgets.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Appendix C
Investment Performance Metrics
Here, we introduce various metrics that are used to gauge the performance of a
portfolio or a trading strategy. We denote the daily trade returns from a strategy
as Rt :

Rt = wt−1 rt − C|wt − wt−1 |,

pt − pt−1 (C.1)
rt = ,
pt−1
where pt denotes daily price series and wt is our trading positions. We here use
a fixed commission cost C. Note that this is a linear cost model and we do not
consider market impact. In practice, trading a large volume could lead to higher
costs as there might be not enough volume on the opposite side and we have to
accept worse prices to liquidate our positions.
Annualized Expected Return:

E(RAnnual ) = E(Rt ) × AF, (C.2)

where AF is the annualization factor. If we do not consider the compounding

factor, we can set AF equal to 252 to get a quick view of the annualized expected
return.
Annualized Standard Deviation:
√
Std(RAnnual ) = Std(Rt ) × AF. (C.3)

Annualized Sharpe Ratio:

E(RAnnual − Rf )
Sharpe = , (C.4)
Std(RAnnual )
where Rf is the risk-free rate and we can set it as the treasury yields or LIBOR.
In this Element, we set it to 0 for simplicity. Sharpe ratio and Sortino ratio are
risk-adjusted returns that measure the return per unit risk.
Annualized Downside Deviation:
√
DD(RAnnual ) = Std(Rt < 0) × AF. (C.5)

Annualized Sortino Ratio:

E(RAnnual − Rf )
Sortino = , (C.6)
DD(RAnnual )
where Rf is again the risk-free rate.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Appendix C 163

Maximum Drawdown refers to the largest decline from an investment’s

peak value to its lowest point over a specific timeframe. We can calculate the
maximum drawdown (MD) as:
High
pt − pLow
t
MD = High
, (C.7)
pt
High
where pt is the peak value before the largest price drop and pLow
t is the
subsequent lowest price before new high is established.
Percentage between Positive and Negative Returns:
E(Rt > 0)
% + Ret = . (C.8)
E(Rt < 0)

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Appendix D
Code Scripts
1 def rolling_forward_cv (X_torch , y_torch , train_end_fractions =
[0.7, 0.8, 0.9], val_fraction = 0.1):
2 """ Perform 3-fold rolling forward CV. Each validation is 10%
of data """
3 N = len(X_torch)
4 all_data []
5 for frac in train_end_fractions :
6 train_end = int(frac * N)
7 val_end = int(( frac + val_fraction) * N)
8 # Safety check if val_end exceeds dataset size
9 if val_end > N:
10 break # no more folds possible if we run off the
end
11 # Create train/val splits
12 X_train , y_train = X_torch [: train_end], y_torch [:
train_end]
13 X_val , y_val = X_torch[train_end:val_end], y_torch[
train_end:val_end]
14 all_data.append (( X_train , y_train , X_Val , y_val))
15 return all_data

Listing 1: A rolling forward cross-validation approach.

1 import [Link] as nn
2

3 class MLP([Link]):
4 def __init__(self , seq_length , n_features , y_dim):
5 super ().__init__ ()
6
7 [Link] = [Link](
8 [Link] (),
9 [Link](seq_length*n_features , 4),
10 [Link] (),
11 [Link] (4, y_dim))
12
13 def forward(self , x):
14 x = [Link](x, start_dim =1)
15 x = [Link](x)
16 y = [Link](x, dim =1)
17 return y

Listing 2: A MLP network for deep portfolio optimization with a long-only

constraint.

1 import [Link] as nn
2

3 class deeplob([Link]):
4 def __init__(self , device):
5 super ().__init__ ()

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Appendix D 165

6 [Link] = device
7
8 self.conv1 = [Link](
9 nn.Conv2d(in_channels =1, out_channels =32,
kernel_size =(1 ,2), stride =(1 ,2)),
10 [Link](negative_slope =0.01) ,
11 nn.BatchNorm2d (32) ,
12 nn.Conv2d(in_channels =32, out_channels =32,
kernel_size =(4 ,1)),
13 [Link](negative_slope =0.01) ,
14 nn.BatchNorm2d (32) ,
15 nn.Conv2d(in_channels =32, out_channels =32,
kernel_size =(4 ,1)),
16 [Link](negative_slope =0.01) ,
17 nn.BatchNorm2d (32) ,
18 )
19 self.conv2 = [Link](
20 nn.Conv2d(in_channels =32, out_channels =32,
kernel_size =(1 ,2), stride =(1 ,2)),
21 [Link] (),
22 nn.BatchNorm2d (32) ,
23 nn.Conv2d(in_channels =32, out_channels =32,
kernel_size =(4 ,1)),
24 [Link] (),
25 nn.BatchNorm2d (32) ,
26 nn.Conv2d(in_channels =32, out_channels =32,
kernel_size =(4 ,1)),
27 [Link] (),
28 nn.BatchNorm2d (32) ,
29 )
30 self.conv3 = [Link](
31 nn.Conv2d(in_channels =32, out_channels =32,
kernel_size =(1 ,10)),
32 [Link](negative_slope =0.01) ,
33 nn.BatchNorm2d (32) ,
34 nn.Conv2d(in_channels =32, out_channels =32,
kernel_size =(4 ,1)),
35 [Link](negative_slope =0.01) ,
36 nn.BatchNorm2d (32) ,
37 nn.Conv2d(in_channels =32, out_channels =32,
kernel_size =(4 ,1)),
38 [Link](negative_slope =0.01) ,
39 nn.BatchNorm2d (32) ,
40 )
41

42 self.inp1 = [Link](
43 nn.Conv2d(in_channels =32, out_channels =64,
kernel_size =(1 ,1), padding='same '),
44 [Link](negative_slope =0.01) ,
45 nn.BatchNorm2d (64) ,
46 nn.Conv2d(in_channels =64, out_channels =64,
kernel_size =(3 ,1), padding='same '),
47 [Link](negative_slope =0.01) ,
48 nn.BatchNorm2d (64) ,
49 )

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 166 Appendix D

50 self.inp2 = [Link](
51 nn.Conv2d(in_channels =32, out_channels =64,
kernel_size =(1 ,1), padding='same '),
52 [Link](negative_slope =0.01) ,
53 nn.BatchNorm2d (64) ,
54 nn.Conv2d(in_channels =64, out_channels =64,
kernel_size =(5 ,1), padding='same '),
55 [Link](negative_slope =0.01) ,
56 nn.BatchNorm2d (64) ,
57 )
58 self.inp3 = [Link](
59 nn.MaxPool2d ((3, 1), stride =(1, 1), padding =(1, 0)),
60 nn.Conv2d(in_channels =32, out_channels =64,
kernel_size =(1 ,1), padding='same '),
61 [Link](negative_slope =0.01) ,
62 nn.BatchNorm2d (64) ,
63 )
64
65 # lstm layers
66 [Link] = [Link](input_size =192, hidden_size =64,
num_layers =1, batch_first=True)
67 self.fc1 = [Link] (64, 1)
68

69 def forward(self , x):

70 # h0: (number of hidden layers , batch size , hidden size)
71 h0 = [Link](1, [Link] (0), 64).to([Link])
72 c0 = [Link](1, [Link] (0), 64).to([Link])
73 x = [Link] (1)
74 x = self.conv1(x)
75 x = self.conv2(x)
76 x = self.conv3(x)
77 x_inp1 = self.inp1(x)
78 x_inp2 = self.inp2(x)
79 x_inp3 = self.inp3(x)
80 x = [Link]((x_inp1 , x_inp2 , x_inp3), dim =1)
81 x = [Link] (0, 2, 1, 3)
82 x = [Link](x, (-1, [Link] [1], [Link] [2]))
83 x, _ = [Link](x, (h0 , c0))
84 x = x[:, -1, :]
85 x = self.fc1(x)[:,0]
86 return x

Listing 3: The network architecture for DeepLOB for a regression problem.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 References
Akiba, T., Sano, S., Yanase, T., Ohta, T., & Koyama, M. (2019). Optuna: A
next-generation hyperparameter optimization framework. In Proceedings of
the 25th acm sigkdd international conference on knowledge discovery & data
mining (pp. 2623–2631).
Almgren, R., & Chriss, N. (2001). Optimal execution of portfolio transactions.
Journal of Risk, 3, 5–40.
Atkins, A., Niranjan, M., & Gerding, E. (2018). Financial news predicts stock
market volatility better than close price. The Journal of Finance and Data
Science, 4(2), 120–137.
Atsalakis, G. S., & Valavanis, K. P. (2009). Surveying stock market fore-
casting techniques–Part II: Soft computing methods. Expert Systems with
Applications, 36(3), 5932–5941.
Bachelier, L. (1900). Théorie de la spéculation. In Annales scientifiques de
l’école normale supérieure (Vol. 17, pp. 21–86). Elsevier
Bahdanau, D., Cho, K., & Bengio, Y. (2014). Neural machine translation by
jointly learning to align and translate. arXiv preprint arXiv:1409.0473.
Bao, W., Yue, J., & Rao, Y. (2017). A deep learning framework for financial
time series using stacked autoencoders and long-short term memory. PloS
one, 12(7), e0180944.
Beck, M., Pöppel, K., Spanring, M., et al. (2024). xlstm: Extended long short-
term memory. arXiv preprint arXiv:2405.04517.
Bengio, Y., Simard, P., & Frasconi, P. (1994). Learning long-term dependen-
cies with gradient descent is difficult. IEEE transactions on neural networks,
5(2), 157–166.
Bertsimas, D., & Lo, A. W. (1998). Optimal control of execution costs. Journal
of financial markets, 1(1), 1–50.
Blondel, M., Teboul, O., Berthet, Q., & Djolonga, J. (2020). Fast differentia-
ble sorting and ranking. In Hal Daumé & Singh, Aarti (eds), International
conference on machine learning (pp. 950–959). PMLR
Borovykh, A., Bohte, S., & Oosterlee, C. W. (2017). Conditional time
series forecasting with convolutional neural networks. arXiv preprint
arXiv:1703.04691.
Boureau, Y., Ponce, J., & LeCun, Y. (2010). A theoretical analysis of feature
pooling in vision algorithms. In Proceedings of international conference on
machine learning (icml’10) (Vol. 28, p. 3).

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 168 References

Briola, A., Bartolucci, S., & Aste, T. (2024). Deep limit order book forecasting.
arXiv preprint arXiv:2403.09267.
Briola, A., Turiel, J., & Aste, T. (2020). Deep learning modeling of limit order
book: A comparative perspective. arXiv preprint arXiv:2007.07319.
Cesa, M. (2017). A brief history of quantitative finance. Probability, Uncer-
tainty and Quantitative Risk, 2(1), 1–16.
Chen, J.- F., Chen, W.- L., Huang, C.- P., Huang, S.- H., & Chen, A.- P.
(2016). Financial time-series data analysis using deep convolutional neural
networks. In Cloud computing and big data (ccbd), 2016 7th international
conference on (pp. 87–92).
Cho, K., Van Merriënboer, B., Gulcehre, C., et al. (2014). Learning phrase rep-
resentations using rnn encoder-decoder for statistical machine translation.
arXiv preprint arXiv:1406.1078.
Cont, R., Cucuringu, M., Kochems, J., & Prenzel, F. (2023). Limit order book
simulation with generative adversarial networks. SSRN 4512356.
Cuturi, M., Teboul, O., & Vert, J.- P. (2019). Differentiable ranking and sort-
ing using optimal transport. Advances in Neural Information Processing
Systems, 32.
Devlin, J., Chang, M.- W., Lee, K., & Toutanova, K. (2018). Bert: Pre-
training of deep bidirectional transformers for language understanding. arXiv
preprint arXiv:1810.04805.
Di Persio, L., & Honchar, O. (2016). Artificial neural networks architectures for
stock price prediction: Comparisons and applications. International Journal
of Circuits, Systems and Signal Processing, 10, 403–413.
Dixon, M. (2018). Sequence classification of the limit order book using
recurrent neural networks. Journal of Computational Science, 24, 277–286.
Doering, J., Fairbank, M., & Markose, S. (2017). Convolutional neural net-
works applied to high-frequency market microstructure forecasting. In
Computer science and electronic engineering (ceec), 2017 (pp. 31–36).
Du, K., Xing, F., Mao, R., & Cambria, E. (2024). Financial sentiment analysis:
Techniques and applications. ACM Computing Surveys, 56(9), 1–42.
Ekmekcioğlu, Ö., & Pınar, M. Ç. (2023). Graph neural networks for deep
portfolio optimization. Neural Computing and Applications, 35(28), 20663–
20674.
Fischer, T., & Krauss, C. (2017). Deep learning with long short-term memory
networks for financial market predictions. European Journal of Operational
Research, 270(2), 654–669.
Frazier, P. I. (2018). Bayesian optimization. In Recent advances in optimization
and modeling of contemporary problems (pp. 255–278). Informs.
Gatheral, J. (2010). No-dynamic-arbitrage and market impact. Quantitative
finance, 10(7), 749–759.
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 References 169

Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep learning. MIT Press.
([Link])
Goodfellow, I., Pouget-Abadie, J., Mirza, M., et al. (2014). Generative adver-
sarial nets. Advances in neural information processing systems, 27.
Grover, A., Wang, E., Zweig, A., & Ermon, S. (2018). Stochastic optimization
of sorting networks via continuous relaxations. In International conference
on learning representations.
Gu, A., & Dao, T. (2023). Mamba: Linear-time sequence modeling with
selective state spaces. arXiv preprint arXiv:2312.00752.
He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image
recognition. In Proceedings of the ieee conference on computer vision and
pattern recognition (pp. 770–778).
Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural
computation, 9(8), 1735–1780.
Hornik, K., Stinchcombe, M., & White, H. (1989). Multilayer feedforward
networks are universal approximators. Neural networks, 2(5), 359–366.
Hwang, Y., Kong, Y., Lee, Y., & Zohren, S. (2025). Decision-informed neural
networks with large language model integration for portfolio optimization.
Jin, M., Wang, S., Ma, L., et al. (2023). Time-LLM: Time series forecasting by
reprogramming large language models. arXiv preprint arXiv:2310.01728.
Kalman, R. E. (1960). A new approach to linear filtering and prediction
problems. Journal of Basic Engineering, Transactions of the ASME, 82(1),
35–45.
Kingma, D. P., & Welling, M. (2013). Auto-encoding variational Bayes. arXiv
preprint arXiv:1312.6114.
Kipf, T. N., & Welling, M. (2016). Semi-supervised classification with graph
convolutional networks. arXiv preprint arXiv:1609.02907.
Kong, Y., Nie, Y., Dong, X., et al. (2024). Large language models for finan-
cial and investment management: Applications and benchmarks. Journal of
Portfolio Management, 51(2) 162–210.
Kong, Y., Wang, Z., Nie, Y., et al. (2024). Unlocking the power of lstm for long
term time series forecasting. arXiv preprint arXiv:2408.10006.
Korangi, K., Mues, C., & Bravo, C. (2024). Large-scale time-varying portfo-
lio optimisation using graph attention networks. arXiv preprint arXiv:2407
.15532.
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). Imagenet classification
with deep convolutional neural networks. Advances in neural information
processing systems, 25.
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017). Imagenet classifica-
tion with deep convolutional neural networks. Communications of the ACM,
60(6), 84–90.
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 170 References

Lim, B., Arık, S. Ö., Loeff, N., & Pfister, T. (2021). Temporal fusion trans-
formers for interpretable multi-horizon time series forecasting. International
Journal of Forecasting, 37(4), 1748–1764.
Lim, B., & Zohren, S. (2021). Time-series forecasting with deep learning:
A survey. Philosophical Transactions of the Royal Society A, 379(2194),
20200209.
Lim, B., Zohren, S., & Roberts, S. (2019). Enhancing time-series momen-
tum strategies using deep neural networks. The Journal of Financial Data
Science, 1(4), 19–38.
Liu, T.- Y. (2009). Learning to rank for information retrieval. Foundations and
Trends® in Information Retrieval, 3(3), 225–331.
Liu, Y., Hu, T., Zhang, H., et al. (2023). itransformer: Inverted transformers are
effective for time series forecasting. arXiv preprint arXiv:2310.06625.
Luong, M.- T., Pham, H., & Manning, C. D. (2015). Effective approaches
to attention-based neural machine translation. arXiv preprint arXiv:1508.
04025.
Maas, A. L., Hannun, A. Y., Ng, A. Y., et al. (2013). Rectifier nonlinearities
improve neural network acoustic models. In Proc. icml (Vol. 30, p. 3).
Markowitz, H. (1952). Portfolio selection. The Journal of Finance, 7(1), 77–91.
Mhaskar, H. N., & Micchelli, C. A. (1993). How to choose an activation
function. Advances in neural information processing systems, 6.
Moreno-Pino, F., & Zohren, S. (2024). Deepvol: Volatility forecasting from
high-frequency data with dilated causal convolutions. Quantitative Finance,
24(8), 1105–1127.
Moskowitz, T. J., Ooi, Y. H., & Pedersen, L. H. (2012). Time series momentum.
Journal of Financial Economics, 104(2), 228–250.
Nagy, P., Calliess, J.- P., & Zohren, S. (2023). Asynchronous deep double
dueling q-learning for trading-signal execution in limit order book markets.
Frontiers in Artificial Intelligence, 6 1151003.
Nagy, P., Frey, S., Sapora, S., et al. (2023). Generative AI for end-to-end limit
order book modelling: A token-level autoregressive generative model of
message flow using a deep state space network. In Proceedings of the fourth
ACM international conference on AI in finance (pp. 91–99).
Nair, V., & Hinton, G. E. (2010). Rectified linear units improve restricted
Boltzmann machines. In Icml (pp. 807–814).
Nelson, D. M., Pereira, A. C., & de Oliveira, R. A. (2017). Stock market’s
price movement prediction with LSTM neural networks. In Neural networks
(ijcnn), 2017 international joint conference on (pp. 1419–1426).

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 References 171

Nie, Y., Nguyen, N. H., Sinthong, P., & Kalagnanam, J. (2022). A time series
is worth 64 words: Long-term forecasting with transformers. arXiv preprint
arXiv:2211.14730.
Obizhaeva, A. A., & Wang, J. (2013). Optimal trading strategy and supply/de-
mand dynamics. Journal of Financial markets, 16(1), 1–32.
Ogryczak, W., & Tamir, A. (2003). Minimizing the sum of the k largest
functions in linear time. Information Processing Letters, 85(3), 117–122.
Poh, D., Lim, B., Zohren, S., & Roberts, S. (2021a). Building cross-sectional
systematic strategies by learning to rank. The Journal of Financial Data
Science, 3(2), 70–86.
Poh, D., Lim, B., Zohren, S., & Roberts, S. (2021b). Enhancing cross-sectional
currency strategies by context-aware learning to rank with self-attention.
arXiv preprint arXiv:2105.10019.
Poh, D., Lim, B., Zohren, S., & Roberts, S. (2021c). Enhancing cross-sectional
currency strategies by ranking refinement with transformer-based architec-
tures. arXiv preprint arXiv:2105.10019.
Poh, D., Roberts, S., & Zohren, S. (2022). Transfer ranking in finance: appli-
cations to cross-sectional momentum with data scarcity. arXiv preprint
arXiv:2208.09968.
Prata, M., Masi, G., Berti, L., et al. (2024). Lob-based deep learning models
for stock price trend prediction: A benchmark study. Artificial Intelligence
Review, 57(5), 1–45.
Pu, X. S., Roberts, S., Dong, X., & Zohren, S. (2023). Network momen-
tum across asset classes. Stephen and Dong, Xiaowen and Zohren, Stefan,
Network Momentum across Asset Classes (August 7, 2023).
Rahimikia, E., Zohren, S., & Poon, S.- H. (2021). Realised volatility fore-
casting: Machine learning via financial word embedding. arXiv preprint
arXiv:2108.00480.
Reisenhofer, R., Bayer, X., & Hautsch, N. (2022). Harnet: A convolu-
tional neural network for realized volatility forecasting. arXiv preprint
arXiv:2205.07719.
Schnaubelt, M. (2022). Deep reinforcement learning for the optimal placement
of cryptocurrency limit orders. European Journal of Operational Research,
296(3), 993–1006.
Selvin, S., Vinayakumar, R., Gopalakrishnan, E., Menon, V. K., & Soman, K.
(2017). Stock price prediction using LSTM, RNN and CNN-sliding window
model. In Advances in computing, communications and informatics (icacci),
2017 international conference on (pp. 1643–1647).

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 172 References

Simonyan, K., & Zisserman, A. (2014). Very deep convolutional networks for
large-scale image recognition. arXiv preprint arXiv:1409.1556.
Sirignano, J., & Cont, R. (2018). Universal features of price formation in finan-
cial markets: Perspectives from deep learning. arXiv preprint arXiv:1803
.06917.
Soleymani, F., & Paquet, E. (2021). Deep graph convolutional reinforcement
learning for financial portfolio management–deeppocket. Expert Systems
with Applications, 182, 115127.
Sun, Q., Wei, X., & Yang, X. (2024). Graphsage with deep reinforcement learn-
ing for financial portfolio optimization. Expert Systems with Applications,
238, 122027.
Sutskever, I., Vinyals, O., & Le, Q. V. (2014). Sequence to sequence learning
with neural networks. In Advances in neural information processing systems
(pp. 3104–3112).
Sutton, R. S., & Barto, A. G. (2018). Reinforcement learning: An introduction.
MIT press.
Theron, L., & Van Vuuren, G. (2018). The maximum diversification invest-
ment strategy: A portfolio performance comparison. Cogent Economics &
Finance, 6(1), 1427533.
Tsantekidis, A., Passalis, N., Tefas, A., et al. (2017a). Forecasting stock prices
from the limit order book using convolutional neural networks. In Business
informatics (cbi), 2017 ieee 19th conference on (Vol. 1, pp. 7–12). IEEE.
Tsantekidis, A., Passalis, N., Tefas, A., et al. (2017b). Using deep learning to
detect price change indications in financial markets. In Signal processing
conference (eusipco), 2017 25th european (pp. 2511–2515).
Van Den Oord, A., Dieleman, S., Zen, H., et al. (2016). Wavenet: A generative
model for raw audio. arXiv preprint arXiv:1609.03499, 12.
Vaswani, A., Shazeer, N., Parmar, N., et al. (2017). Attention is all you need.
Advances in neural information processing systems, 30.
Vergara, J. R., & Estévez, P. A. (2014). A review of feature selection methods
based on mutual information. Neural Computing and Applications, 24, 175–
186.
Wan, X., Yang, J., Marinov, S., et al. (2021). Sentiment correlation in financial
news networks and associated market movements. Scientific Reports, 11(1),
3062.
Wang, J., Zhang, S., Xiao, Y., & Song, R. (2021). A review on graph neural
network methods in financial applications. arXiv preprint arXiv:2111.15367.
Wang, Y., Wu, H., Dong, J., et al. (2024). Deep time series models: A
comprehensive survey and benchmark. arXiv preprint arXiv:2407.13278.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 References 173

Wood, K., Giegerich, S., Roberts, S., & Zohren, S. (2021). Trading with the
momentum transformer: An intelligent and interpretable architecture. arXiv
preprint arXiv:2112.08534.
Wood, K., Kessler, S., Roberts, S. J., & Zohren, S. (2023). Few-shot learning
patterns in financial time-series for trend-following strategies. arXiv preprint
arXiv:2310.10500.
Wood, K., Roberts, S., & Zohren, S. (2021). Slow momentum with fast rever-
sion: A trading strategy using deep learning and changepoint detection. arXiv
preprint arXiv:2105.13727.
Wu, H., Xu, J., Wang, J., & Long, M. (2021). Autoformer: Decomposi-
tion transformers with auto-correlation for long-term series forecasting.
Advances in Neural Information Processing Systems, 34, 22419–22430.
Wu, Z., Pan, S., Chen, F., et al. (2020). A comprehensive survey on graph neu-
ral networks. IEEE transactions on neural networks and learning systems,
32(1), 4–24.
Zhang, C., Pu, X., Cucuringu, M., & Dong, X. (2023). Graph neural networks
for forecasting multivariate realized volatility with spillover effects. arXiv
preprint arXiv:2308.01419.
Zhang, C., Pu, X., Cucuringu, M., & Dong, X. (2024). Graph-based meth-
ods for forecasting realized covariances. Journal of Financial Econometrics,
nbae026.
Zhang, C., Zhang, Z., Cucuringu, M., & Zohren, S. (2021). A universal end-
to-end approach to portfolio optimization via deep learning. arXiv preprint
arXiv:2111.09170.
Zhang, X., Chowdhury, R. R., Gupta, R. K., & Shang, J. (2024). Large language
models for time series: A survey. arXiv preprint arXiv:2402.01801.
Zhang, Y., & Yan, J. (2023). Crossformer: Transformer utilizing cross-
dimension dependency for multivariate time series forecasting. In The
Eleventh International Conference on Learning Representations.
Zhang, Z., Lim, B., & Zohren, S. (2021). Deep learning for market by order
data. Applied Mathematical Finance, 28(1), 79–95.
Zhang, Z., & Zohren, S. (2021). Multi-horizon forecasting for limit order
books: Novel deep learning approaches and hardware acceleration using
intelligent processing units. arXiv preprint arXiv:2105.10430.
Zhang, Z., Zohren, S., & Roberts, S. (2019). Deep convolutional neural net-
works for limit order books. IEEE Transactions on Signal Processing, 67(11),
3001–3012.
Zhang, Z., Zohren, S., & Roberts, S. (2019a). Deeplob: Deep convolutional
neural networks for limit order books. IEEE Transactions on Signal Pro-
cessing, 67(11), 3001–3012.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 174 References

Zhang, Z., Zohren, S., & Roberts, S. (2019b). Extending deep learning mod-
els for limit order books to quantile regression. Proceedings of Time Series
Workshop of the 36 th International Conference on Machine Learning, Long
Beach, California, PMLR 97, 2019.
Zhang, Z., Zohren, S., & Roberts, S. (2020). Deep learning for portfolio
optimization. The Journal of Financial Data Science, 2(4), 8–20.
Zhou, H., Zhang, S., Peng, J., et al. (2021). Informer: Beyond efficient trans-
former for long sequence time-series forecasting. In Proceedings of aaai (pp.
11106–11115).
Zhou, Y.- T., & Chellappa, R. (1988). Computation of optical flow using a
neural network. In Icnn (pp. 71–78).

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Acknowledgments
We owe our profound gratitude to everyone who contributed to the successful
completion of this book. To begin with, we wish to acknowledge our fami-
lies for their unfailing support. We also want to recognize the assistance of our
colleagues and friends, whose insightful conversations, constructive criticism,
and fresh perspectives helped us to shape the content of this book. In particular,
we would like to thank our senior colleagues, Steve Roberts, Xiaowen Dong,
Jan Calliess, Mihai Cucuringu, Alex Shestopaloff, Janet Pierrehumbert, Jakob
Foerster, Ani Calinescu, Nick Firoozye, Chao Ye, Xiaoqing Wu, and Yongjae
Lee, as well as research students and postdocs whose work was featured
here, Bryan Lim, Kieran Wood, Daniel Pho, Will Tan, Fernando Moreno-Pino,
Chao Zhang, Vincent Tan, Xingyue Pu, Yaxuan Kong, Yoontae Hwang, Felix
Drinkall, Dragos Gorduza, Peer Nagy, Xingchen Wan, Binxin Ru, and Sasha
Frey. Special thanks also to Samson Donick for proofreading the entire man-
uscript, as well as several of the above students for proofreading individual
sections. Additional thanks goes to George Nigmatulin, Yaxuan Kong and
Yonntae Hwang for helping with didactic materials around the book. More-
over, we would like to thank Bank of America for hosting a short lecture series
based on the book attended by 200 quants. In particular, special thanks go to
Robert De Witt, Ilya Sheynzon and Shih-Hau Tan for organising the event, as
well as to Leif Andersen for carefully reading the manuscript and providing
additional comments.
Our thanks extend as well to the editorial and publishing team, in particular
our editor Riccardo Rebonato for his insightful feedback and patience along the
process. We are deeply thankful to the Machine Learning Research Group and
the Oxford-Man Institute, at the University of Oxford for providing us with a
supportive research environment. We would also like to thank Man Group for
sponsoring the institute and their engagement through their academic liaisons
Anthony Ledford and Slavi Marinov. Without all your support, this book would
never have come to fruition.

To our families

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Quantitative Finance

Riccardo Rebonato
EDHEC Business School
Editor Riccardo Rebonato is Professor of Finance at EDHEC Business School and holds
the PIMCO Research Chair for the EDHEC Risk Institute. He has previously held academic
positions at Imperial College, London, and Oxford University and has been Global Head
of Fixed Income and FX Analytics at PIMCO, and Head of Research, Risk Management
and Derivatives Trading at several major international banks. He has previously been on
the Board of Directors for ISDA and GARP, and he is currently on the Board of the Nine
Dot Prize. He is the author of several books and articles in finance and risk management,
including Bond Pricing and Yield Curve Modelling (2017, Cambridge University Press).

About the Series

Cambridge Elements in Quantitative Finance aims for broad coverage of all major topics
within the field. Written at a level appropriate for advanced undergraduate or graduate
students and practitioners, Elements combines reports on original research covering an
author’s personal area of expertise, tutorials and masterclasses on emerging
methodologies, and reviews of the most important literature.

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]
 Quantitative Finance

Elements in the Series

Machine Learning for Asset Managers
Marcos M. López de Prado
Advances in Retirement Investing
Lionel Martellini and Vincent Milhau
A Practitioner’s Guide to Discrete-Time Yield Curve Modelling: With Empirical
Illustrations and MATLAB Examples
Ken Nyholm
Girsanov, Numeraires, and All That
Patrick S. Hagan and Andrew Lesniewski
Causal Factor Investing: Can Factor Investing Become Scientific?
Marcos M. López de Prado
The Behavioral Economics and Politics of Global Warming: Unsettling Behaviors
Hersh Shefrin
Hydrodynamics of Markets: Hidden Links Between Physics and Finance
Alexander Lipton
Deep Learning in Quantitative Trading
Zihao Zhang and Stefan Zohren

A full series listing is available at: [Link]/EQF

Downloaded from [Link] IP address: [Link], on 03 Oct 2025 at [Link], subject to the Cambridge Core terms of
use, available at [Link] [Link]