Machine Learning With Applications: Markus Vogl, Peter Gordon Rötzel (LL.M), Stefan Homes

Machine Learning with Applications 8 (2022) 100302
Contents lists available at ScienceDirect
Machine Learning with Applications

journal homepage: www.elsevier.com/locate/mlwa
Forecasting performance of wavelet neural networks and other neural

network topologies: A comparative study based on financial market data sets
Markus Vogl a ,∗, Peter Gordon Rötzel (LL.M) b,a , Stefan Homes a
a University of Applied Sciences Aschaffenburg, Würzburger Stra"ze 45, 63743 Aschaffenburg (DE), Germany
b
University of Stuttgart, Keplerstra"ze 17, 70174 Stuttgart (DE), Germany
ARTICLE INFO ABSTRACT

JEL classification: In this study, we analyse the advantageous effects of neural networks in combination with wavelet functions on
C22 the performance of financial market predictions. We implement different approaches in multiple experiments
C32 and test their predictive abilities with different financial time series. We demonstrate experimentally that
C52
both wavelet neural networks and neural networks with data pre-processed by wavelets outperform classical
C45
network topologies. However, the precision of conducted forecasts implementing neural network algorithms
C53
G17
still propose potential for further refinement and enhancement. Hence, we discuss our findings, comparisons
with ‘‘buy-and-hold’’ strategies and ethical considerations critically and elaborate on future prospects.
Keywords:
Wavelet neural networks
Wavelet decomposition
Financial forecasting
Forecast evaluation
Neural network topology
Financial markets
1. Introduction (such as time series), NNs exhibit a higher potential for producing more
accurate predictions than conventional statistical methods (e.g. expo-
Forecasting models tailored for financial time series are discussed nential smoothing, as shown by Hill et al., 1996) (Paliwal & Kumar,
frequently in business and science (Sezer et al., 2020). The development 2009). Applications to the field of financial market and risk manage-
of computer-based methods has witnessed significant progress, which ment predictions are given in Petneházi (2021), among others, stating
is well illustrated by, for example, Renaissance Technologies, led by convolutional neural networks (CNNs) for value-at-risk predictions.
James Simons. Renaissance Technologies has systematically outper- However, according to Makridakis et al. (2018), NNs have not revealed
formed market growth over many years through the execution and their full potential, which can be seen in Peng et al. (2021), analysing
advanced analysis of algorithms and signals (Burton, 2016). However, several deep neural networks to assess empirical performance in terms
the majority of actively managed funds1 based on statistical anal- of technical indicators. Furthermore, Peng et al. (2021) state that
ysis display severe underperformance owing to lower yields earned
no strategy under analysis was able to outperform simple buy-and-
compared to the respective benchmark (market) indices (Otuteye &
hold strategies. Contrastingly, following, for example. Chalvatzis and
Siddiquee, 2019). In particular, said underperformance renders itself
Hristu-Varsakelis (2020) or Nobre and Neves (2019) state respective
visible once trading and management fees are considered, which are
outperformance of buy-and-hold strategies already. The advantages
compared with passive investments, such as buy-and-hold strategies
of NNs as predictive models are owing to the significantly increased
(Otuteye & Siddiquee, 2019).
The rapid growth in the research field of artificial intelligence (AI) amount of data availability, as well as higher computational capacity
since 2010 highlights neural networks (NNs), particularly in terms (Jordan & Mitchell, 2015). Therefore, the methods and experiments
of computer-aided methods (Sezer et al., 2020). Based on machine discussed in this study examine, for example, the conjecture that larger
learning and pattern recognition algorithms in large amounts of data and better data sets lead to a more accurate prediction of stock and
index prices. The latter conjecture follows the well-known remark of
∗ Corresponding author.
E-mail addresses: markus.vogl@vogl-datascience.de (M. Vogl), peter.roetzel@th-ab.de (P.G. Rötzel), s150554@th-ab.de (S. Homes).
1
Personal decisions are involved in the active process of investing, whereas passive management of funds is based solely on index allocation, for example, on
market capitalisation (Gruber, 1996).
2
Non-periodic, localised wave function, which integral yields exactly zero value (Daubechies, 1990).
https://doi.org/10.1016/j.mlwa.2022.100302
Received 4 December 2021; Received in revised form 10 March 2022; Accepted 11 April 2022
Available online 18 April 2022
2666-8270/© 2022 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
M. Vogl, P.G. Rötzel and S. Homes Machine Learning with Applications 8 (2022) 100302
Peter Norvig, namely, that ‘‘more data beats clever algorithms but in terms of financial market prediction. Further, to substantiate our
better data beats more data’’. Furthermore, the concept of wavelets,2 findings regarding both the research questions, we perform a respective
taken from the field of signal processing, provides some interesting back-test.
application possibilities in individual tests regarding the analysis of Moreover, we state other studies taken out of the academic liter-
financial time series (Crowley, 2007). In particular, for the processing ature (e.g. Kumbure et al., 2022 or Yang & Wang, 2021), proposing
of time series with different periodicity, as well as for short- and long- comparisons of WNNs or otherly NN comparisons to provide a holistic
term cycles, wavelet transforms offer advantages (e.g. for business picture of stated groundwork determinations. Finally, we elaborate on
cycle analysis) (Crowley, 2007). Although Sezer et al. (2020) state that the underlying topology that is best suited for this kind of methodology
novel methodologies in NN designs are well researched, it remains and discuss the results in terms of performance comparability with
unresolved, whether a combination of NNs with signal processing buy-and-hold strategies and ethical considerations critically.
techniques, such as previously mentioned wavelet functions, indicates a
(positive) effect on the predictive performance in financial forecasting 2. Literature review
(Alexandridis & Zapranis, 2014).
Therefore, this gap warrants further investigation, leading to our In the 1990s, many NN-projects3 were carried out in the field
first research question, namely, the elucidation of potential relations of time-series analysis, and most of the latter projects present the
between the amount of input data and predictive time horizons. Fur- basis for corresponding future research endeavours (Vellido et al.,
thermore, regarding prediction accuracy, it appears that various net- 1999). Nonetheless, according to Sezer et al. (2020), the area of NN
work topologies perform differently at variable time horizons (Tsan- research was neglected until 2011. During this period, many publica-
tekidis et al., 2017). Our second research question elucidates the pro- tions (e.g. Chen, 1994 or Vellido et al., 1999) reported problems with
cedure of proportionally replacing the neurons of established network NNs, mainly owing to computers in the said period to provide very
topologies with wavelons (i.e. a wavelet function as replacement of a little computational power or respective memory (Lim et al., 2004).
sigmoidal activation function, refer to Alexandridis & Zapranis, 2014), Furthermore, problems with successful implementation existed; hence,
thus, generating wavelet neural networks (WNNs) (Zhang & Benveniste, more simplistic models were considered sufficient at the time (Adya &
1992). Alexandridis and Zapranis (2014), as well as other analysed Collopy, 1998). With the progression of the 2010s and driven by the
publications (e.g. Billings & Wei, 2005; Zhang, 1997), reveal that cer- extensive availability of data and inexpensive computational capacity,
tain wavelet functions are more suitable than other respective wavelet the field of NN-based AI experienced a rediscovery, visible in the vast
equations. A wavelet activation function (e.g. a Morlet-wavelet) can dissemination of academic research (Jordan & Mitchell, 2015). The
outperform a logistic function; however, additional analysis of further results show that NNs are suitable for producing acceptable predictions
wavelet functions is required to achieve a respective generalisation in the case of financial time series, even if only to a limited extent
of the latter superiority presupposition (Anjoy & Paul, 2017). The (Adya & Collopy, 1998). Nevertheless, the potential for improvement
fundamental demand for combined models (i.e. of hybrid models), was recognised, including the application of hybrid models (e.g. Vui
such as WNNs (see also Yang & Wang, 2021), is derived from the fact et al., 2013; Zhou et al., 2019). Hybrid models generally combine
that (financial) time series contain various information components, two or more well-established methods (e.g. classical NNs with signal
such as time and frequency information (Chakrabarty et al., 2015). processing techniques or stochastic elements) into a novel application
Frequency components can be extracted by applying transformations (Dong et al., 2016). However, hybrid models should be tested carefully
(e.g. Fourier transformation [FT]) to the data sets (Chakrabarty et al., with different time series before deployment to obtain a reasonable
2015). Unfortunately, once the FT is applied, the required time in- understanding of their inherent functionality (Guresen et al., 2011;
formation components are lost, whereas, the wavelet transformation Yang & Wang, 2021).
preserves said time information (Chakrabarty et al., 2015). Addition- First, following Alexandridis and Zapranis (2013, 2014), referring to
ally, NNs are advantageous over statistical models (e.g. exponential the intermediate research stream mentioned earlier, we apply a hybrid
smoothing) regarding the processing capabilities of nonlinear functions model combined with a respective wavelet transform to transfer a time
(Alexandridis & Zapranis, 2013; Hill et al., 1996). series into the frequency domain. Subsequently, based on Taspinar
Therefore, within the academic literature, corresponding (2018), the signal is processed and subdued to a respective back-
approaches are elaborated on from two different research streams, transformation to obtain data, serving as input for a classical NN. The
namely, intermediate and generalised. Moreover, a combination of NNs second, more innovative strategy consists of partially replacing the
and wavelets is possible, applicable and reasonable for various scientific neurons of the referred network with wavelons (Zhang & Benveniste,
disciplines (Alexandridis & Zapranis, 2014). The intermediate research 1992). In addition, the resulting WNN is constructed from a three-
stream deals with the pre-processing of time series by applying wavelet layer MLP (Alexandridis & Zapranis, 2014). Henceforth, we refer to
decompositions (Alexandridis & Zapranis, 2014). Therefore, we discuss the previously mentioned basic structure applying wavelons as part of
the effect of the additional input of time series processed with wavelet the generalised research stream, which differs from the intermediate
decomposition on a NN instead of applying only unprocessed time research stream in terms of activation functions (Alexandridis & Zapra-
series. The experiment is focused on comparing the predictive success nis, 2014). Wavelet activation functions are applied to the underlying
to well established NNs, such as multilayer perceptron (MLP) and long network topology instead of the commonly implemented sigmoidal
short-term memory (LSTM) networks without pre-processing the data. activation functions (Zhang et al., 1995).
In addition, we aim to determine, whether performance deviations To the best of our knowledge, some publications discuss the men-
between different referring prediction periods exist. In response to the tioned approaches; however, elaborate exclusively on their exploration
previously mentioned implication by Norvig, several publications elab- in general or aim directly at specialised applications (e.g. see Yang
orate on the ‘‘more data’’ hypothesis critically. For example, Walczak & Wang, 2021 for a comparison in terms of energy series) (Alexan-
(2001) concludes that considering more than two years of data has no dridis & Zapranis, 2014). This is due to increased design and testing
significant effect on the accuracy of forecasting models. We explore requirements for WNNs (Oussar & Dreyfus, 2000). Nevertheless, ac-
a WNN approach in which the activation functions of an NN (e.g. in cording to Alexandridis and Zapranis (2013), a framework for building
the hidden layers) are proportionally replaced by wavelet functions WNNs is available, which discusses construction and training algo-
(Zhang & Benveniste, 1992). Therefore, we investigate the change in rithms in detail. However, most publications with an economically
the forecasting performance of an NN, in which a wavelet is applied as
an activation function. More specifically, we intend to determine the 3
For example, see Hill et al. (1996), Kaastra and Boyd (1996), and Zhang
most appropriate wavelet function in combination with NN topologies et al. (1995) for further reference.
2
relevant alignment focus solely on the intermediate research stream. In Table 1

The most commonly implemented network topologies as depicted in our
particular, Bao et al. (2017), which combines accuracy and profitability
sampled literature. The results are in line with the findings of Sezer et al.
measurements, can be considered in this regard. Further, Bao et al. (2020) as a point of reference.
(2017) present reasonable and reproducible values for three different Network topology Percentage of analysed publications
network topologies (including LSTM). The referring data sets that
LSTM 31%
Bao et al. (2017) investigate are various economic indices, including MLP 21%
S&P500 and DJIA100, as well as four other Asian indices. Hybrid 8%
Even if the literature provides some successful predictions with NNs, Other specifications 40%
as mentioned before, successful forecasts are not guaranteed because Notes: LSTM: long short-term memory; MLP: multilayer perceptron.
of the stochastic functionality of NNs, which is based on the learning
algorithm implemented in each NN (Kaastra & Boyd, 1996). The latter
learning algorithm performs a gradient descent operation as part of the and quality of the input data are crucial for the precision of prediction
training process and can compute an acceptable solution (i.e. local error capability and the quality of an NN. We tested this experimentally by
minimum). Unfortunately, the results do not necessarily represent the developing several NNs according to the intermediate approach and
optimal solution (i.e. the global error minimum) of the approximation comparing the respective results based on the quantity of input data as
of the respective training function (Kaastra & Boyd, 1996). Further- well as with two common NN topologies, namely, MLP and LSTM. The
more, some wavelet functions in the field of application with NNs are hypothesis of the second question is that the involvement of a WNN by
more commonly implemented than others (Alexandridis & Zapranis, a generalisation approach fundamentally improves the success of pre-
2014). Depending on the application, different wavelet functions are dictions. Furthermore, we demonstrate that not all wavelet functions
recommended, mainly the family of Gaussian wavelets (e.g. Billings & serve equally well as activation functions. Therefore, further classical
Wei, 2005 or Zhang, 1997) (Alexandridis & Zapranis, 2014). Moreover, NNs are constructed with different wavelets as activation functions.
wavelets can be custom-designed if required, i.e., specifically tailoring
said functions to the respective application and its objectives (Misiti 4. Methodology, experimental design, and data description
et al., 2007). This study focuses on two Gaussian wavelets, namely, the
first and second derivatives of the Gaussian bell curve as depicted in For the experiments, we create eight different NNs in accordance
Barlow (1983). The latter is also labelled as the Mexican Hat function, with the criteria defined in the research questions (see Section 3)
owing to its shape. and examine the latter for respective predictive capabilities. There-
Different application frequencies are also observable in the choice of fore, we initially examine and implement two well-established network
topology for NNs. Therefore, we present the three most applied topolo- topologies, then, based on the results, we implement the other six ex-
gies from our sampled literature, as shown in Table 1. This selection is perimental NNs. This procedure aims to create the same foundation for
also favoured to set up benchmark networks in the experiment with both research questions, thus, minimising the implementation effort,
MLP and LSTM as bases. Since all other topologies are built on or as well as the risk of potential errors, while concurrently improving
derived from MLPs, the MLP is the most widely researched network comparability. MLP-based NNs can only handle one time series simulta-
topology in the field of financial market prediction and is repeatedly neously and can only predict a single value. By contrast, the other LSTM
taken as a referring benchmark (Sezer et al., 2020). According to topologies are capable of an n-step-ahead-forecast and accept multiple
De Faria et al. (2009), this simple NN predicts the positive or negative time series as input. Therefore, LSTMs allow predictions over longer
sign for yields in the Brazilian stock market with 60% accuracy. The periods, which exceed just one-day-ahead forecasts.
aforementioned second and ongoing wave of AI research has shown In addition, we intend to derive statements about whether more
significant progress since the 2010s, especially, with newly developed input data leads to better results, referring to the initially stated and
NN topologies (Jordan & Mitchell, 2015). Furthermore, it would be of academically critically discussed proposition of Walczak (2001) (see
interest to elaborate on subsequently the aforementioned hybrid models Section 1). To study the research stream of WNNs, we first implement
(Sezer et al., 2020; Vui et al., 2013). However, such an experiment does four topologies followed by two more topologies to study subsequently
not necessarily lead to predictive success, as demonstrated through the the effects of wavelet decomposition.
combination of generalised autoregressive conditional heteroscedastic-
ity models and MLP, which are highly influenced by existing noise 4.1. Wavelet neural networks
effects in the data under analysis by Guresen et al. (2011).
By contrast, Tsaih et al. (1998) find the hybrid extended futures Like classical NNs, a WNN generally consists of three different
forecast model to be promising in terms of trading S&P500 futures. In layers: input, hidden, and output layers (Alexandridis & Zapranis,
addition, a similar methodology can be found in Zhang et al. (2001), 2014). Owing to theoretical construction possibility of feedforward NNs
who also analyse futures trading and show that hybrid models con- in terms wavelet decompositions (see Pati & Krishnaprasad, 1993 or
sisting of wavelets and NNs lead to more profitable trading results Zhang & Benveniste, 1992 for early reference), WNNs represent an
than otherwise implemented MLPs, although they are dependent on alternative to cope with NN weaknesses (such as randomised starting
outliers in different market situations. Therefore, we will also elabo- values in training algorithms) (Alexandridis & Zapranis, 2013). Fur-
rate on the general performance of NN solutions in comparison with thermore, WNNs represent a generalisation of radial basis function net-
simplistic ‘‘buy-and-hold’’ strategies and elucidate for which market works (Alexandridis & Zapranis, 2013). In contrast to well-established
actors under which dynamical presupposition these methodologies can networks, the hidden layer of a WNN does not contain neurons, but
be favourable (see Section 6.2). wavelons that fulfil the same task as neurons. The only difference lies
in the respective activation function specifications, namely, incorpo-
3. Research hypotheses rating a wavelet function instead of a sigmoidal function representa-
tion (Alexandridis & Zapranis, 2014). To be more detailed, the nodes
We further elaborate on hybrid models by focusing on the appli- (i.e. the wavelons) of a WNN are the wavelet coefficients of the func-
cation of different NNs with wavelet components, since the correct tion expansion, yielding a significant value (Alexandridis & Zapranis,
specification of network topology as well as wavelet-function is a key 2013). Moreover, multidimensional wavelets preserve the ‘‘universal
problem according to Chen et al. (2006). Both of the aforementioned approximation’’ property that is characteristic for NNs (Alexandridis &
research questions aim to improve the performance of NNs in financial Zapranis, 2014). Furthermore, reasons to conduct said alterations lay
market predictions. The first research question claims that the quantity within the characteristics of wavelets (refer to Bernard et al., 1998),
3
Fig. 1. Structure of a three-layer wavelet neural network (WNN), consisting of input, hidden, and output layers. Note that the activation function of a WNN (right) differs from
those of standardised neural networks (left), namely, the WNN incorporates wavelons, which represent themselves via respective wavelet functions.
Fig. 2. Wavelet functions applied as wavelon activation functions, namely, the Gaussian wavelet activation functions: Gauss1 (left) and Gauss2 (right).
namely, high compression capabilities in computing the value at a Table 2

Comparison of well-established neural networks and experimen-
single point or updating the functional estimate from a novel local
tal wavelet neural networks elaborated on during the empirical
measure, which involves only a small subset of all given coefficients, investigation of this study.
respectively (Alexandridis & Zapranis, 2013, 2014). Thus, WNNs al- Established network topology Experimental network
low for constructive procedures, which efficiently initialise network
WNN (MLP & Gauss1)
parameters, i.e. convergence to the global minimum of the referring MLP
WNN (MLP & Gauss2)
cost function (Alexandridis & Zapranis, 2013). To be more detailed, the
WNN (LSTM & Gauss1)
initial weight vector of a WNN4 is present into close proximity of the LSTM
WNN (LSTM & Gauss2)
global minimum and, therefore, drastically reduces training times, as
Notes: MLP: multilayer perceptron; LSTM: long short-term
also stated in Alexandridis and Zapranis (2013) and Yang and Wang memory.
(2021). Thus, the general idea of a WNN is given by the aim to adapt
the corresponding wavelet basis to the respective training data (Alexan-
dridis & Zapranis, 2014). In case of multi-wavelet NNs the according
is common to conduct an initial decomposition of the signal (e.g. of a
activation function is given by a linear combination of wavelet bases
respective stock price series) by applying a wavelet transform (Mallat,
and combinable with discrete wavelet transforms (DWTs) or principal
1989). Second, following Taspinar (2018), the results are processed
component analysis (PCA) algorithms (Alexandridis & Zapranis, 2013).
in the respective frequency domain. Finally, the signal is transformed
A conceptual illustration of this is shown in Fig. 1. The only significant
back to the time domain by executing a back-transformation (Mallat,
change in the implementation of a WNN compared with other NNs is
1989). One disadvantage is that signal decomposition by transforma-
the inclusion of the manually defined activation function, namely, the
tion leads to overlaps at the edges of the respective time–frequency
resulting and previously noted wavelons. Two wavelet functions are
series and is amplified by larger choices of the decomposition level
applied, namely, the first (Gauss1) and second (Gauss2) derivatives of
(Williams & Amaratunga, 1997). Unfortunately, said edge effects can-
the Gaussian bell curve, as illustrated in Fig. 2. The two standard net-
work topologies, MLP and LSTM, incorporate the previously mentioned not be completely avoided due to the finite nature of real-world time-
Gaussian wavelets, resulting in four test networks, which are depicted series data (Torrence & Compo, 1998). Therefore, in accordance with
in Table 2. Anjoy and Paul (2017), we select decomposition levels, which do
not exceed the value of five. The additional data to be analysed
4.2. Wavelet decomposition neural networks within the WDNNs are pre-processed by applying the previously men-
tioned wavelet decomposition and being fed into the LSTM topology
In contrast to WNNs, the NN itself is not changed within the accordingly.
wavelet decomposition neural network (WDNN) research; however, The wavelet transform displays the property of a low-pass filter,5
more time series are added to the input amount (Alexandridis & Za- which we intend to exploit (Taspinar, 2018). Furthermore, we focus
pranis, 2014). First, regarding the research field of signal processing, it on two NNs with different levels of decomposition. The first WDNN
performs one level of pre-processing and the second performs five,
4
For an in-depth mathematical treatise refer to Alexandridis and Zapranis
5
(2013, 2014) and Zapranis and Alexandridis (2008, 2009). Denoising of the input data is conducted (Torrence & Compo, 1998).
4
Table 3 in the case of hybrid models, such as WNNs, it is not certain that
Comparison of well-established neural networks and experi-
the accuracy represents a performance measure that evaluates the
mental wavelet decomposition neural networks (WDNNs). The
WDNNs decompose respective signals or series through wavelet
predictive power of (financial) time series in a meaningful and correct
transformation within the frequency domain before conducting manner (Alexandridis & Zapranis, 2014; Kumbure et al., 2022). Fur-
a back-transformation into the time domain, respectively. thermore, we determine a potential gap within the academic literature;
Established network topology Experimental network namely, only a few studies address this difficulty properly. Therefore,
WDNN (Level 1) we disregard accuracy as a measure and apply another measure that is
LSTM
WDNN (Level 5) more practicable and more descriptive in terms of evaluating forecasts,
Notes: LSTM: long short-term memory; WDNN: wavelet as described in the following (Alexandridis & Zapranis, 2014). In the
decomposition neural network. final back-test described in Section 4.3, we performed several runs and
statistically evaluate individual predictions. Subsequently, we present
the results as a 95% confidence interval around the true value to
which results in an increase in the number of input time series by be predicted and as percentage values, following Altman and Bland
one and five, respectively. Both WDNNs are based on a basic LSTM (2005). Therefore, the individual prediction results are averaged over a
topology. Hence, following the recommendation of Lahmiri (2014), time horizon. Further, twice the standard deviation of the mean value
Daubechies wavelet 4 (DB4) is the most feasible for decomposition is subtracted and then added to obtain the accuracy interval, which we
purposes. Hereinafter, we generate the experimental NNs, as shown in apply within our empirical setting (Altman & Bland, 2005).
Table 3.
4.4. Data description
4.3. Basic network structure, fitting procedure and accuracy intervals
Following Halevy et al. (2009), one of the most important factors for
The aforementioned network structure of the three layers is applied solving both research questions is represented by the data under inves-
for all NNs built on an MLP basis. Further, regarding networks with tigation. Therefore, we extract all relevant data sets from the renowned
an LSTM basis, additional layers are added to enable the function of Refinitiv Eikon Datastream (formerly Thomson Reuters), which are
n-step-ahead forecasts, which are not available for the MLPs under publicly available as adjusted-values, that is, adjusted for corporate
consideration. The output layer corresponds to the number of features actions, such as stock splits (Refinitiv Limited, 2019). Another factor
(i.e. the number of time series) being fed into the NN. Nevertheless, our to be considered in the choice of data is the respective stock exchange
experiment intends only to predict the daily (-adjusted) closing prices. from which the time series are obtained. At this point, we favour the
In terms of network structure, the hidden layer remains sufficiently New York Stock Exchange (NYSE) because of its high trading volume
variable such that the number of neurons can be adjusted, if necessary. (Statista Research Department, 2019). If the NYSE is not capable of
Following the guidelines proposed by Guresen et al. (2011), the number providing sufficient data quality, the trading venue with the largest
of learning epochs (see Rashid, 2016) of an NN should be larger than trading volume in the home country of the respective corporation is
the number of weights and, thus, of the neuron connections in the selected as the relevant source. All time series are provided in the
network (Guresen et al., 2011). However, respecting this guideline related national currency of the stock exchange.6 The recorded datasets
leads to overfitting in the present experiment (Guresen et al., 2011). were available until the end of February 2020. We examine 20 datasets,
The application of excessive amounts of learning epochs ensures that of which 15 are public companies and five are indices, as displayed in
the memory capacity of the network approximates the training data Table 4. The stocks analysed are based on eight of the largest companies
perfectly, which negatively affects the prediction performance of un- within the MSCI World index in terms of market capitalisation. As
known data (Lawrence et al., 1998). Therefore, for all topologies, it these positions in the MSCI World index represent American stocks
is important to determine the number of learning epochs that lead almost exclusively, we supplement the analysis with six of the largest
to the best forecasting results before overfitting sets in. To simplify positions within the Euro STOXX50 index (Ticker SX5T), in a manner
this step, we implement an early stopping mechanism, as proposed that European stocks are also included in our experiment (STOXX,
in Caruana et al. (2000). This ensures that the training process is 2021). These datasets are supplemented ultimately by the stock of the
stopped, as soon as no further significant improvement in fitting can Asian Alibaba Group.
be achieved (Kaastra & Boyd, 1996). The mean squared error (MSE) The indices analysed are the NASDAQ100, DJIA100, DAX30, EURO
can be calculated as a measure of goodness of fit (GoF) for a given STOXX50 and CSI300. If required, we clean the datasets of recording
(stock) time series, which is approximated by the referring network errors, i.e., in some cases, missing or erroneous values are averaged
(Alexandridis & Zapranis, 2014). The MSE indicator describes the av- from the previous and subsequent values (Kaastra & Boyd, 1996).
erage of the squared deviations at each given value in the time series Refinitiv datasets are available on a daily basis and are provided as
of the approximated curve to the real value of the training data curve ‘‘.csv’’-files. The procedure for time-series data originating from indices
(Alexandridis & Zapranis, 2014). The MSE is supposed to be minimised is handled analogously. The first date recorded in each case is the day
in the NN without causing overfitting (Alexandridis & Zapranis, 2014). when a share has been tradable on the referring stock exchange and,
The final value to be predicted must not be trained as an input data therefore, can differ from the respective initial public offering date. For
point in the NN (Adya & Collopy, 1998; Kaastra & Boyd, 1996). To test all financial time series under consideration, the opening and closing
the repeated predictions of the NNs, we apply an 80–20 split of the data prices, as well as the highest and lowest values within the daily data
as proposed by Kaastra and Boyd (1996). frequency, are recorded. Each dataset is split into training and test sets
Therefore, all network topologies are based on only 80% of the (refer to Section 4.3). The common division of 80% for the training set
data and subsequently predict values from the isolated 20% test dataset and 20% for the test set is applied (Kaastra & Boyd, 1996).
(Kaastra & Boyd, 1996). Finally, the specification of the results and In this experiment, we implement NNs as regressors to output the
their precision measures requires a formal discussion. Even if the prices of stocks and, consequently, are, in principle, not obligated
well-accepted accuracy is a measure representing the GoF in terms to scale the input data (Albon, 2018). However, as observed in the
of time-series approximation, it does not necessarily imply anything datasets under analysis, many stock prices rise sharply over the years
about how well an NN can predict future value developments outside considered. To consider the latter insight, we apply a scaling function
the training data set (Cristea et al., 2000). Assuming overfitting, for
instance, produces very high accuracy based on the training data, yet, 6
US-Dollar for American companies and Alibaba Group, Euro for European
displays no predictive capability (Caruana et al., 2000). Especially companies, as well as index points for indices.
5
Fig. 3. Overview of the predictability logic of our empirical analysis, namely, iterative counts per prediction, per data set, as illustrated.
Table 4 or index and LSTM topology, less for shorter data sets,8 respectively.
Selected publicly data sets for this study, consisting of the company or
Furthermore, up to 20 predictions are added with NNs on MLP basis
respective index name, the ticker symbol, and the country of origin.
The data sets are extracted from Refinitiv Eikon Datastream (formerly
for each data set, thus, we obtain a total number of 58,446 predic-
Thomson Reuters) on a daily frequency until February 2020. tions, which are documented and evaluated according to the back-test
Name Symbol Country described previously. As suspected, the examination of the prediction
Stocks
results provides a differentiated picture of the quality of predictions
regarding the different selected NNs. Therefore, it is favourable to
Alibaba Group BABA.K China
address different perspectives within the respective evaluations. The
L’Oréal S.A. OREP.PA
France first perspective envisages the results with a focus on the general
LVMH SE LVMH.PA
performance of an NN over the time horizon from one to 30 days.
Allianz SE ALV
Linde plc LIN
Therefore, we consider all available results of each network, separated
Germany per time horizon but averaged across all data sets. The second perspec-
SAP SE SAP.N
Siemens SIEGn.DE tive elucidates the performance within the predicted stocks and indices.
Apple Inc. AAPL.O For each stock or index, the best network is determined for each time
Alphabet Inc. C GOOGL.O horizon. Further, it can be deduced whether an NN performs better than
Amazon Inc. AMZN.O others regarding certain data sets and time-horizons.
Facebook Inc. FB.O
Hence, we present the results in more detail. The various result
Johnson & Johnson JNJ USA
Johnson & Johnson JNJ tables display the prediction accuracy around the true value by cal-
JP Morgan Chase JPM culating the confidence interval described in Section 4.3, resulting in
Microsoft Corp. MSFT.O more narrow intervals for better predictions than wide intervals. The
Visa Inc. V intervals are given at 95% confidence; hence, individual predictions of
Indices individual networks are much more accurate. Because two standard
China Securities Index 300 CSI300 China deviations are selected, the intervals can also yield negative values.
EURO STOXX 50 STOXX50 Europe Nevertheless, due to the stochastic nature of NNs, more accurate out-
DAX 30 GDAXI Germany comes are not guaranteed in every prediction. The variation ranges in
the results across all stocks and indices examined are, in the best case,
Dow Jones Industrial Average DJI
USA approximately 9% around the true value and about 44% in the worst
NASDAQ 100 NDX
case. The best case is presented as a one-day-ahead forecast stated in
Table 5, indicating the best NNs per prediction period. The worst-case
interval is approximately 5% wider than the best network displayed
that transfers the input variables into a range between zero and one. for a 30 day-ahead prediction (approximately 39% interval width) and,
Additionally, we note the benefits of the efficiency of the respective therefore, is not shown in Table 5. Hence, there are large differences
NN by employing a respective scaling operation (Kaastra & Boyd, in the respective fluctuation ranges of the accuracy intervals, as stated
1996). Moreover, we perform back-tests over a period of two years, in Table 5. A possible explanation is provided by the predicted time
for example, a training and test set split of 80–20 results in a training horizon.
data period of eight years for a test period of two years, assuming
Following Nguyen and Chan (2004) it is found that the longer
that data for exactly 10 years are available. For stocks that are listed
the time horizon, the wider the interval around the true value and
for a shorter period, the test period is reduced, whereas the 80–20
the worse the long-term forecast. The best case mentioned above is
ratio remains constant. The NNs predict 30 days of future stock prices
represented by the forecast for the next day, while the worst case is
for each test iteration, which repeatedly records predictions from 20
a 30 day forecast. Regarding the evaluation of individual stocks, the
successive inputs, as illustrated in Fig. 3. In the case of shorter datasets,
referring intervals are sometimes more precise, namely, up to 3.21%
we retain the 30 day prediction periods, yet, with a shorter iteration
for the Johnson & Johnson stock, yet, mostly within the range of the
count than 20, such as Facebook Inc. (13 samples) or Alibaba Group (9
values stated in Table 5. In individual cases, such as within the Apple
samples).
Inc. dataset, however, these are significantly less accurate, resulting
5. Empirical results in extreme cases in terms of fluctuation ranges of up to 89.60%,
rendering predictions futile. To provide a holistic representation of all
Before elaborating on respective results, we note the stochastic results, the complete overview of our empirical findings are given in
nature of NNs to impose a great influence on the magnitude deviations the supplementary material of this study.
between the predictions (Kaastra & Boyd, 1996). The before mentioned In general, many cases of the researched networks (i.e. WNNs and
implication originates from the existence of many predictions, which WDNNs) indicate better performance than the basic topologies, despite
enable the prediction of only small fractions of a percent from the a few limitations, as shown in Tables 6 and 7. These tables provide a
intended value. In addition, however, a similar number of predictions ranking of the best NNs with respect to individual datasets. The first
deviate from the true value by several percent, according to non- ranking (see Table 6) displays results for a period of one day and the
optimal gradient descent algorithms, as stated in Kaastra and Boyd second ranking (see Table 7) for longer-term predictions. For example,
(1996). In total, we record up to 600 predictions7 from each stock
8
For example, Facebook Inc. (13 times 30 days) or Alibaba Group (9 times
7
20 times 30 days. 30 days).
6
the MLP continues to be the best NN for one-day-ahead forecasts Table 5

Evaluation table referring to all data sets, stating the best network topology per period
for 40% of the datasets examined. The WNN with the first Gaussian
as well as respective accuracy intervals.
derivative as the activation function ranks as the second best in terms of
Prediction period in days Accuracy interval Best topology per period
predictability performance and is the best NN on more than one-third of
1 [−4.94%; 4.05%] MLP
the datasets. Thus, the basic MLP and WNN (MLP & Gauss1) are almost
2 [−6.28%; 6.02%] LSTM
equivalent. Even though the performance is slightly lower than that of 3 [−7.35%; 6.65%] WDNN (Level1)
the MLP, the performance is better than the LSTM topology. This over- 4 [−10.11%; 5.46%]
performance can still be seen in long-term forecasts (see Table 7). The 5 [−10.59%; 5.85%]
WNN (LSTM & Gauss1) is the best predictor for 8 out of 20 datasets and, 6 [−11.27%; 6.34%]
WNN (LSTM & Gauss1)
7 [−11.56%; 6.69%]
thus, is superior to LSTM. By contrast, the WNN (LSTM & Gauss2) does
8 [−11.80%; 7.01%]
not represent the best performer for any dataset, but displays a similar 9 [−12.29%; 7.50%]
performance to the established LSTM. Even if the WDNNs provide the 10 [−11.76%; 8.89%]
best results with 15% and 10% for a quarter of all data sets, only a 11 [−11.87%; 8.78%]
WNN (LSTM & Gauss2)
12 [−12.17%; 9.11%]
closer examination reveals further important indications.
13 [−12.75%; 9.61%]
Moreover, adding time series to the input, which is pre-processed 14 [−13.05%; 9.53%]
with wavelet decomposition, does not always produce favourable re- 15 [−13.39%; 9.88%]
sults and does not lead to an improvement in the performance of the 16 [−13.58%; 9.86%]
predictions. 17 [−13.82%; 9.73%]
18 [−14.05%; 10.00%]
For example, regarding the topology WDNN (Level 5), some predic-
19 [−14.10%; 9.85%]
tions perform worse than all other NNs. However, few cases exist in 20 [−14.55%; 10.16%]
which the WDNN (Level 5) reveals over-performance that is, actually 21 [−14.98%; 10.65%]
representing the best results within a dataset. Compared to the other 22 [−15.49%; 10.99%] WNN (LSTM & Gauss2)
23 [−15.73%; 11.14%]
topologies, the WDNN (Level 1) provides the best performance over all
24 [−15.67%; 10.99%]
data sets for a period of three days. Furthermore, the accuracy interval 25 [−16.02%; 11.31%]
of [−7.35%; 6.65%] displays a width of 14% around the true stock 26 [−16.36%; 11.89%]
price. Regardless of the comparisons, the best prediction of the WDNN 27 [−16.55%; 12.00%]
(Level 1) is found with a two-day prediction period and a fluctuation 28 [−16.72%; 12.12%]
29 [−16.70%; 12.22%]
interval of 13.4%. For single datasets, the WDNN (Level 1) is best suited 30 [−16.78%; 12.22%]
for short-term prediction, namely, for Microsoft Corp., Apple Inc., JP
Notes: MLP: multilayer perceptron; LSTM: long short-term memory; WDNN: wavelet
Morgan Chase, Allianz SE, and the CSI300 Index. An example is shown
decomposition neural network; WNN: wavelet neural network.
in Fig. 4. Please note that due to the high volume of forecasts, only
selective graphical displays are possible since a holistic display would
render itself unsuitable. The aforementioned WDNN (Level 5) causes
Table 6
said topology to be at a disadvantage over the total set of predictions Ranking of the most successful network topologies for a predictive period of one day
in most cases. In addition, the WDNN also provides a systematic under- (pred. = 1 day).
estimation of the stock price at the 95% significance level of about 20% Network topology [%] — Share of data sets best
on average, which is reflected in an accuracy interval width of 40%. predicted (pred. = 1 day)
However, an exceptional perspective on the WDNN (Level 5) MLP 40%
emerges while regarding individual stock datasets separately. The WNN (MLP & Gauss1) 35%
WDNN (Level 5) is the best topology for long-term predictions of the LSTM 10%
WNN (MLP & Gauss2) 5%
stocks of Apple Inc., Facebook Inc., and Alibaba Group, as well as the WDNN (Level 1) 5%
STOXX50 Index. The WNNs implemented in this experiment provide WNN (LSTM & Gauss2) 5%
the most accurate predictions for long-term time horizons out of all
Notes: MLP: multilayer perceptron; LSTM: long short-term memory; WDNN: wavelet
topologies and datasets under consideration. With a prediction period decomposition neural network; WNN: wavelet neural network.
ranging from four to nine days, the WNN with the first Gaussian deriva-
tive as an activation function provides the best predictions. Further,
from the 10th to the 30th prediction, the WNN with the Mexican Hat
with one-day-ahead predictions. Of all predictions, 5% lie within a
activation function is more advantageous. This answers the question of
tenth of a percent frame around the value to be achieved, and two-
which wavelet of the two is better suited from this perspective. The
thirds of all predictions are between −2% and +2%. Further, the three
Mexican Hat wavelet displays better results, but with little difference
experimental NNs achieve a more accurate ‘‘hit rate’’9 than the basic
compared to the first Gaussian derivative. The accuracy intervals for
LSTM model.
both WNNs range from approximately 15% (four days) to 29% (30
Furthermore, we note that the studied MLP-based NNs offer a more
days) around the true value. Following Fig. 5, we present an example
accurate performance than those based on LSTM across all networks
of a 30 day forecast employing the WNN (LSTM & Gauss2). The basic
with WDNN (Level 1) as the only exception. A combination of the
topology, which is more adequate for WNNs, is represented by the MLP
results regarding the accuracy interval and precision suggests the ap-
for one-day-ahead price forecasts. The first Gaussian derivative as a
wavelet is slightly better suited (interval [−6.12%; 3.71%]) than the plication of wavelets as an activation function to reduce the scatter
Mexican Hat function (interval [−4.89%; 5.21%]). For longer periods, of the predictions, but it does not necessarily improve the accuracy
only LSTM is favourable because of the possibility of n-step-ahead itself. However, if the wavelet functions are operationalised for input
forecasts. Contrary to the one-day-ahead forecast, the Mexican Hat processing, a larger variation is observed, with an equally inconsistent
wavelet reveals a better performance than the first Gaussian derivative, increase in flawlessness. Furthermore, we present a complete overview
with respect to time horizons of at least 10 days. of the one-day versus multiple-day forecasting topologies for each
Finally, the third perspective elucidates the number of almost flaw- dataset in Table 9.
less predictions. Now we examine the amount of predictions of each
network topology, which lie within a specified interval. Table 8, shows 9
Describes the number of predictions meeting the target accuracy interval
that the MLP as an established topology also presents the best precision (in line with Gupta & Lam, 1996).
7
Fig. 4. Example of 30 day prediction with experimental wavelet decomposition neural network (Level 1) based on the Microsoft stock data set. Notes: WDNN: wavelet decomposition
neural network.
Fig. 5. Example of 30 day prediction with experimental wavelet neural network (LSTM & Gauss2) based on the Johnson & Johnson stock data set. Notes: WNN: wavelet neural
network.
6. Discussion and implications 6.1. Empirical implications
The empirical results of this study confirm that hybrid models are To the best of our knowledge and following the implications of these
potentially more advantageous than classical NNs (e.g. Bao et al., 2017;
results, the premise that simply increasing the amount of data will lead
Cristea et al., 2000; Yang & Wang, 2021; Zhang et al., 2001). In
to enhanced predictive results is not confirmed with respect to the
addition, regarding the present experiments, the MLP is one of the most
experimental setting of this study. Even before the final experiment,
reliable predictors, which lends credence to its continued application
in research and practice (Guresen et al., 2011; Sezer et al., 2020). To the time series for logarithmic returns and trading volumes per day
be more detailed, we present empirical implications in Section 6.1, tested as additional inputs led to inadequate predictions of all rele-
while subsequently elaborating on the performance comparison with vant network topologies. Therefore, we discard the latter approach.
buy-and-hold investment strategies in Section 6.2, elucidating ethical Regarding WDNN (Level 5), adding additional time series merely to
considerations in Section 6.3, propose limitations in Section 6.4 and increase data volume is not favourable. These findings align with the
state future avenues of research in 6.5. conclusions of Walczak (2001), who states that larger amounts of data
8
Table 7 to the data period leads to an increase in performance in general is

Ranking of the most successful network topologies for a predictive period of multiple yet to be discussed. We exclude the Apple Inc. data set, which is
days (pred. > 1 day).
the least predictable across all NNs. Nevertheless, for the other stocks
Network topology [%] — Share of data sets best
and indices, no clear relation between the length of a data set and
predicted (pred. = 1 day)
the quality of the prediction is visible. Predominantly, one-day-ahead
WNN (LSTM & Gauss1) 40%
LSTM 35%
forecasts range within 6% and 8% fluctuations around the true value.
WDNN (Level 5) 15% Exceptions can be determined for three data sets, namely, Amazon
WDNN (Level 1) 10% Inc. with 10% interval width, Johnson & Johnson with 3%, and EURO
WNN (LSTM & Gauss2) 0% STOXX50 with 4%. Some of the longer price time series, such as
Notes: MLP: multilayer perceptron; LSTM: long short-term memory; WDNN: wavelet Microsoft Corp., result in accuracy intervals around 5% to 6%. Data
decomposition neural network; WNN: wavelet neural network. sets that consider 30 to 40 years of data show an average fluctuation of
around 7%, while those displaying only less than 20 years of data points
Table 8 also present the same fluctuation. This finding leads to the conclusion
Precision of different topologies displayed as ‘‘hit rate’’, which describes the number that a longer data period is not necessarily advantageous compared to
of predictions meeting the target accuracy interval (in line with Gupta & Lam, 1996). shorter data sets. Moreover, Walczak’s (2001) claim that larger data
One-day prediction [%] — Share of forecasts in the interval amounts do not enhance NN forecasts is substantiated. Referring to the
Topology [−0.1%; 0.1%] [−1.0%; 1.0%] [−2.0%; 2.0%] most similar publication to this present study, namely, the forecasting
MLP 5% 36% 66% results of Bao et al. (2017), showing NNs with wavelet pre-processed
WNN (MLP & Gauss1) 3% 31% 54% input to be at an advantage against other networks, for example, LSTM,
WDNN (Level 1) 3% 24% 43% are proven.
WNN (MLP & Gauss2) 2% 31% 58%
Agreeing with the presupposition of Bao et al. (2017) that NNs with
LSTM 2% 21% 42%
WNN (LSTM & Gauss1) 1% 18% 37% wavelet components can outperform classical topologies, our study
WNN (LSTM & Gauss2) 1% 21% 39% reveals that not only intermediate procedures outperform classical
WDNN (Level 5) 1% 2% 5% topologies; the presented WNNs developed from the generalisation
Notes: MLP: multilayer perceptron; LSTM: long short-term memory; WDNN: wavelet methodology perform at an even higher precision compared to the
decomposition neural network; WNN: wavelet neural network. respective intermediate results. Therefore, more innovative WNNs (as
proposed and elaborated on Yang & Wang, 2021) are preferred over
WDNNs in the majority of time horizons and data sets. The discussion
do not consistently produce better forecasting results. Although WDNN envisages the attempt to understand and render the functioning of NNs
(Level 5) achieves good results in a few cases, the time required by comprehensible by proven methods to reach its limits in many cases.
the network for training is increased significantly to the point that Therefore, we cannot confirm some detailed aspects (e.g. the outperfor-
the respective cost–benefit ratio no longer matches that of the other mance of WNN over MLP as stated by Zhang et al., 2001 derived from
topologies (e.g. MLP or LSTM). the literature in this study). Consequently, it is particularly difficult
Further, considering Norvig’s remark, namely, that better data is to grasp the performance of NNs from a big picture perspective. The
more important than simply more data, we must discuss whether methods in the research area of explainable AI, that is, the investigation
the data that a WDNN additionally receives is really better data. By of how NNs achieve results as stated in Adadi and Berrada (2018),
performing low-pass filtering during decomposition, smoothing10 (see should begin at this point and gain further importance and necessity
Fig. 6) of the input data sets occurs, as illustrated in Fig. 5. However, in the future, while employing an increasing number of applications of
this is a correction of the data by high-frequency price fluctuations. AI.
Yet, it contrastingly represents a falsification of the actual true dataset. We evaluate the precision of the NNs under analysis, as far as exact
Thus, the possibility of mutually offsetting effects must be considered. predictions are concerned, as is not completely practical so far. As
only a single digit percentage of the prediction results can be found
Therefore, we assume that the increase in quantity per se does not
within 0.1% of the exact value, we neglect the execution of these
ensure better results if the added data do not increase the overall
models as a sole prediction tool. As a workaround, we suggest training
data quality. Furthermore, predictions of MLPs, which receive only a
multiple NNs of a topology on the same data set and then consider and
single time series as input, are not significantly worse in comparison
evaluate multiple predictions in the same period, which may optionally
with those that process four or more time series. By contrast, the
average the latter to achieve higher accuracy. This may mitigate the
performance in many cases is even better, as can be seen, for example,
stochastic aspect to which NNs are fundamentally subject, even though
in the analysis of Alphabet Inc. and Microsoft Corp.
this increases the already high expenditure on computational resources
Moreover, we assess whether a larger dataset leads to better long-
even further. Alternatively, the presented networks are suitable as a
term predictions. We explicate a former presupposition by the two
complement to the tools applied in fund management practices so far.
shorter time series such as Facebook Inc. (since 2012) and Alibaba
Reducing the black-box problem with the aforementioned explainable
Group (since 2014). Complementing the latter comparison, the longest
AI is expected to help in terms of understanding and acceptance (Remus
data sets, namely Johnson & Johnson and JP Morgan Chase (both since
& O’Connor, 2001). This development can already be seen in Zhou
1980), are noted. Considering the longest prediction periods of 25 to
et al. (2019), implementing NNs in terms of predictive algorithms,
30 days, Facebook Inc. and Alibaba Group show interval widths in
while Li and Kuo (2008) generally show that wavelet algorithms help
the range of 30% to 40%; hence, they are significantly less accurate
to maximise returns on given timescales for financial institutions. More-
predictions than the other datasets. However, smaller intervals only
over, we state Puchalsky et al. (2018) proposing several optimisation
occur for one of the previously mentioned stocks with a long data
algorithms for WNNs, which can be seen as direct enhancement of our
period, namely, Johnson & Johnson. The inaccuracy of the Johnson & findings in terms algorithmic improvement possibilities. In addition,
Johnson predictions lies within the range of less than 20%. However, Kanarachos et al. (2017) applies WNNs to detect successfully real-time
the JP Morgan Chase predictions resemble the short data sets at ap- anomalies within markets, which our findings are capable of displaying
proximately 30%. Thus, we cannot currently conclude that longer input further improvement capabilities due to the comparability character
periods also lead to more accurate long-term forecasts. Simultaneously, of our study. Regarding the generalisability of our findings, WNNs are
regarding the referred data sets, whether greater input with respect also applied in different scientific domains facing the same optimisation
problems. For example, Alexandridis and Zapranis (2013, 2014) imple-
10
That is, denoising the time series (Torrence & Compo, 1998). ment the discussed WNN solutions for financial, chaotic, wind (refer to
9
Table 9
Best network topology for a one-day predictive period (pred. = 1 day) and for longer periods (pred. > 1
day) for each selected data set within our respective analysis.
Name Best network (pred. = 1 day) Best network (pred. > 1 day)
Stocks
Alibaba Group MLP WDNN (Level 5)
L’Oréal S.A. MLP WNN (LSTM & Gauss1)
LVMH SE MLP LSTM
Allianz SE WDNN (Level 1) WNN (LSTM & Gauss1)
Linde plc WNN (MLP & Gauss1) WNN (LSTM & Gauss1)
SAP SE WNN (MLP & Gauss1) LSTM
Siemens WNN (MLP & Gauss1) LSTM
Apple Inc. MLP WDNN (Level 5)
Alphabet Inc. C WNN (MLP & Gauss2) WNN (LSTM & Gauss1)
Amazon Inc. WNN (MLP & Gauss1) WNN (LSTM & Gauss1)
Facebook Inc. WNN (LSTM & Gauss2) LSTM
Johnson & Johnson LSTM LSTM
Johnson & Johnson LSTM LSTM
JP Morgan Chase LSTM WDNN (Level 1)
Microsoft Corp. WNN (MLP & Gauss1) WNN (LSTM & Gauss1)
Visa Inc. WNN (MLP & Gauss1) WNN (LSTM & Gauss1)
Indices
China Securities Index 300 WNN (MLP & Gauss1) WDNN (Level 1)
EURO STOXX 50 MLP WDNN (Level 5)
DAX 30 MLP WNN (LSTM & Gauss1)
Dow Jones Industrial Average MLP LSTM
NASDAQ 100 MLP LSTM
Notes: MLP: multilayer perceptron; LSTM: long short-term memory; WDNN: wavelet decomposition neural
network; WNN: wavelet neural network.
Fig. 6. Example for the pre-processed input sequence into wavelet decomposition neural networks, demonstrating the low-pass smoothing characteristics.
Doucoure et al., 2016; Liu et al., 2013) or breast cancer datasets and existence of stylised facts (i.e. empirical observations) such as dynam-
can further be applied to image processing, signal denoising, density ical nonlinearity (see Alexandridis et al., 2017), volatility dynamics
estimates and time-scale decomposition (see Berghorn, 2015). (see Adams et al., 2017) or momentum-induced (multi-)fractal trends
(e.g. Berghorn, 2015 or Daniel & Moskowitz, 2016). Even if the EMH
6.2. Performance comparison ‘‘buy-and-hold’’ is questioned further in its entirety due to stated deterministic chaotic
characteristics (see Vogl & Rötzel, 2022), the implications and base-
Especially in the financial domain, the discussion of the necessity of line comparisons of respective back-tests (see López de Prado, 2018)
advanced algorithms and neural network solutions in comparison with still level around beating ‘‘buy-and-hold’’ scenarios, although not all
‘‘buy-and-hold’’ investment strategies is ongoing (Couillard & Davison, institutional trading facilities propagate such passive investments. Fur-
2005; Vogl, 2021). The reason of the prominence of ‘‘buy-and-hold’’ or thermore, following Berghorn (2015) states the theoretical possibility
passive investment strategies, as introductory mentioned, lays within of outperformance, contrasting the EMH. Comparing the performance
the still broadly stated efficient market hypothesis (EMH), which in of NNs or advanced machine learning algorithms with ‘‘buy-and-hold’’
short states that financial market data is obliged to randomness and strategies in the literature can be seen, for example, in Nobre and Neves
forecasting attempts, thus, are futile in their nature (Fama, 1965a, (2019) proposing hybrid AI models for trade signal generation for
1965b; Fama & French, 1989). The EMH is vastly critiqued by Man- intra-day trading funds. In addition, Chalvatzis and Hristu-Varsakelis
delbrot (1963) and Mandelbrot and Taylor (1967) and faced with the (2020) state NN outperformance over ‘‘buy-and-hold’’ strategies in
10
automated asset trading scenarios, which our results may propose a management decision problem as given. Furthermore, black-box sys-
solid groundwork for further return maximisation. Moreover, Sezer tems, such as NN solutions, render a full ‘‘understanding’’ merely
et al. (2017) state outperformance of machine learning algorithms, impossible. Autonomous trading systems are prone to outliers as well
while Mitra (2009) shows higher risk-adjusted returns by ANNs than as possible to investments in morally discussable assets and market
via ‘‘buy-and-hold’’, which also is of interest to institutional entities. regimes. Biases in investment decisions are reflected potentially in AI
Finally, referring to the encompassing review of Kumbure et al. (2022), system once past trading results (originally conducted by a human)
show advanced algorithms to be capable to deal with complex system is applied for training purposes. Presupposing AI systems to become
dynamics such as financial markets and propose detailed comparisons more autonomous or even sentient in the future, the authors see the
of methodologies and evaluations, while conclusively stating NNs to be potential implementation of investment guidelines (e.g. CFA code of
capable of further prospering. conduct) into such AI systems conducted for trading and forecasting as
favourable to prevent misuse or fraud. Nonetheless, a formal framework
6.3. Ethical considerations for AI forecasts in the financial domain is still lacking.
As AI and advanced technology is incorporating more space in our 6.4. Limitations and future research
reality, a short debrief on ethical considerations in terms of its imple-
mentation is deemed relevant. AI technologies in general represent dual This study has several shortcomings. First, we focus only on a few
use, namely, the incorporation in peaceful civilian as well as military- indices and stocks as a singular asset class and neglect a broader pool
driven systems, which is mostly ignored (de Ágreda, 2020). To cope of available financial market data (e.g. commodities and bonds). Fur-
with such a realisation, a ban of research activity is neither favourable thermore, we only analyse daily-frequented data sets and disregard an
nor is it possible to stop respective activities, thus, the need for ethical elaboration on higher frequencies (e.g. intraday data sets). Moreover,
principles guiding such endeavours in imminent (de Ágreda, 2020). we do not discuss in detail the implications stemming from the choice
Said principles should follow two characteristics, namely, the algorithm of price versus different returns on the research questions at hand.
functionality is understood and humans retain enough system control to Furthermore, we note that in this study, we focus mainly on plain or
intervene if required, since nefarious deployment of data is possible and standard NN implementations in terms of topology selection (e.g. LSTM
AI is developed at a rapid growth rate (de Ágreda, 2020). Therefore, a or MLP) and do not discuss or test our wavelet decomposition approach
vast discussion on ethical frameworks for advanced technology exists with more complex topologies (as proposed in Yang & Wang, 2021)
(e.g. AI4PEOPLE, EAD2, COMEST or DEEPMIND), revelling around that are currently discussed in AI research. In addition, we do not
the dimensions of beneficence, human dignity, privacy, human au- elaborate on the potential of customised wavelet functions, which we
tonomy, fairness and explainability (de Ágreda, 2020; Parasuraman deem to be interesting in combination with more sophisticated network
et al., 2000). Building upon these insights, AI provides high benefits topologies. One aspect that has hardly been addressed recently is the
for humans in general, yet, can be jeopardised without implementation real-time application and usability of NNs in the field under discussion
of formerly denoted ethical codes and security measures (de Ágreda, (refer to Kanarachos et al., 2017). Thus, we note that the training of
2020). Thus, following Yu et al. (2018), AI systems render themselves all implemented NN topologies requires a high amount of time and
increasingly ubiquitous, which in the public experience, AI governance computational effort (Bao et al., 2017). Further, practical research
incorporating ethical standards become more relevant. Yu et al. (2018), should focus on improving efficiency and simplifying handling in terms
therefore, propose a taxonomy incorporating the dimensions of ethical of NNs (Remus & O’Connor, 2001). Making AI accessible to groups
dilemmas, individuality in ethical decisions, given frameworks and without an affinity for computer science should be brought into focus to
human AI interaction. Moreover, owing to AI structuring, historical promote more widespread adoption of its advantageous applicability.
data application in social life applications often result in unjust biases In addition, the dissemination and execution of AI methods in criti-
and discrimination by machine learning algorithms, raising discussions cal research fields (e.g. quantitative finance) pose further challenges
about fairness and debiasing (Birhane, 2021). Especially, social sys- in risk management and associated legal consequences, which must
tems are vulnerable due to the increase in mathematicalisation and be addressed (Adadi & Berrada, 2018). In particular, regarding the
formalisation of social issues owed to the advanced propositions of AI described fluctuations in the quality of the results of the discussed
systems, leading to operations characterisable as value-free, neutral or network topologies, risk management should not be neglected. Regard-
even amoral (Birhane, 2021). Further following Birhane (2021) state a ing methodological restrictions, the application of WNNs is limited
historical discussion about the roots of these developments and offers to applications of small input dimensions owing to computationally
guidance on the ‘‘correctness’’ of biases in terms of their definition. expensiveness if facing high dimensional input vectors, even if capable
Due to the lack of a moral agent in machines, Etzione and Etzione of handling nonlinear and non-stationary datasets (Alexandridis & Za-
(2017) propose reasons for their incorporation in AI systems, yet, stat- pranis, 2013). Finally, training of WNNs with backpropagation requires
ing the risk of the latter drawing on extreme outliers, which would lead storage and inversion of some matrices, which in case of larger datasets,
to fatal errors, labelling this occurrence as outlier fallacy. Therefore, grow fondly large and, thus, computationally expensive (Alexandridis
Banerjee (2020) suggests a computational framework for engineering & Zapranis, 2014).
intelligence to understand the concept of machine consciousness better.
Moreover, Gruson et al. (2019) discuss the generating process of AI 6.5. Concluding remarks
systems in the dimensions of ethics, legal predicaments, privacy and
financing in terms of tailor-made versus off-the-shelf AI solutions, In this study, we demonstrate experimentally that hybrid models
while discussing its implications regarding augmented reality. Gal et al. (e.g. WNNs and WDNNs) have advantages over classical topologies
(2020) show the discourse of AI in people analytics, especially pointing (e.g. LSTM) regarding financial market predictions. However, adding
out its impact in management support systems, stressing bias-free and more data does not necessarily improve prediction performance be-
lack of ethics. Building upon the moral agent in AI systems, Nath and cause the increased data quantity in the given case is accompanied by a
Sahu (2020) state its inexistence due to the lack of the answer to the loss of data quality and leads to cancellation effects. The implemented
questions of ‘‘why being moral matters’’ in terms of a moral agent to wavelet functions may be partially recommended as an alternative
be functional in AI systems. Ashok et al. (2022), therefore, propose to the sigmoid activation function; however, the choice is dependent
generalised digital ethical frameworks for AI and digital technologies. on the respective data set. Hypotheses that both approaches, namely,
Referring to the implementation of WNNs and advanced algorithms intermediate and generalisation, lead to an increase in forecasting
in financial disciplines or practical processes, the authors state the performance compared to classical NNs, can be reinforced, apart from
11
the MLP, which is the best predicting NN with regard to one-day-ahead Billings, Stephen, & Wei, Hua-Liang (2005). A new class of wavelet networks for
forecasts. Therefore, WNNs are preferable for longer-term forecasts. nonlinear system identification. IEEE Transactions on Neural Networks, 16, 862–874.
http://dx.doi.org/10.1109/TNN.2005.849842.
To apply the presented concepts in economic practice and better as-
Birhane, A. (2021). Algorithmic injustice: A relational ethics approach. Patterns, 2(2),
sess the risks for investment fund practice, additional tests on further Article 100205.
data sets and a significant increase in precision are required, which Burton, K. (2016). Inside a moneymaking machine like no other. Bloomberg Markets.
reflects the state-of-the-art goal within our sampled literature since Caruana, R., Lawrence, S., & Giles, C. L. (2000). Overfitting in neural nets:
Backpropagation, conjugate gradient, and early stopping. In NIPS.
outperformance of ‘‘buy-and-hold’’ strategies are already partially ap-
Chakrabarty, A., De, A., Gunasekaran, A., & Dubey, R. (2015). Investment horizon
plicable. Nevertheless, following the conclusions in academic literature, heterogeneity and wavelet: Overview and further research directions. Physica A:
the proof-of-concept concerning NNs with wavelet components is fully Statistical Mechanics and its Applications, 429(C), 45–61.
substantiated for the presented models. Therefore, we propose our Chalvatzis, C., & Hristu-Varsakelis, D. (2020). High-performance stock index trading
findings as generalised basis and solid groundwork for future studies, via neural networks and trees. Applied Soft Computing, 96, Article 106567.
Chen, C. H. (1994). Neural networks for financial market prediction. In Proceedings of
as help to select an optimised combination of topology versus wavelet
1994 IEEE international conference on neural networks (pp. 1201).
as well as within other optimisation procedures. Chen, Y., Yang, B., & Dong, J. (2006). Time-series prediction using a local linear wavelet
neural network. Neurocomputing, 69, 449–465.
CRediT authorship contribution statement Couillard, M., & Davison, M. (2005). A comment on measuring the Hurst exponent
of financial time series. Physica A. Statistical Mechanics and its Applications, 348,
404–418.
Markus Vogl: Conceptualisation, Methodology, Validation, Investi- Cristea, P. D., Tuduce, R., & Cristea, A. I. (2000). Time prediction with wavelet neural
gation, Resources, Writing – review & editing, Visualisation, Supervi- networks. In Proceedings of the 5th seminar on neural network applications in electrical
sion. Peter Gordon Rötzel: Supervision, Project administration, Fund- engineering (IEEE Cat. No. 00EX287) (pp. 5–10).
ing acquisition. Stefan Homes: Software, Formal analysis, Data cura- Crowley, P. M. (2007). A guide to wavelets for economists. Journal of Economic Surveys,
21, 207–267. http://dx.doi.org/10.1111/j.1467-6419.2006.00502.x.
tion, Writing – original draft. Daniel, K., & Moskowitz, T. J. (2016). Momentum crashes. Journal of Financial
Economics, 221–247.
Declaration of competing interest Daubechies, I. (1990). The wavelet transform, time-frequency localization and signal
analysis. IEEE Transaction on Information Theory, 36, 961–1005.
De Faria, E. L., Albuquerque, M. P., Gonzalez, J. L., Cavalcante, J. T. P., &
The authors declare that they have no known competing finan-
Albuquerque, M. P. (2009). Predicting the Brazilian stock market through neu-
cial interests or personal relationships that could have appeared to ral networks and adaptive exponential smoothing methods. Expert Systems with
influence the work reported in this paper. Applications, 36(10), 12506–12509.
Dong, B., Li, Z., Rahman, S. M., & Vega, R. E. (2016). A hybrid model approach for
forecasting future residential electricity consumption. Energy and Buildings, 117,
Appendix A. Supplementary data
341–351.
Doucoure, B., Agbossou, K., & Cardenas, A. (2016). Time series prediction using
Supplementary material related to this article can be found online artificial wavelet neural network and multi-resolution analysis: Application to wind
at https://doi.org/10.1016/j.mlwa.2022.100302. speed data. Renewable Energy, 92, 202–211.
Etzione, A., & Etzione, O. (2017). Incorporating ethics into artificial intelligence. The
Journal of Ethics, 21, 403–418.
References Fama, E. F. (1965a). The behaviour of stock-market prices. Journal of Business, 38(1),
34–105.
Adadi, A., & Berrada, M. (2018). Peeking inside the black-box: a survey on explainable Fama, E. F. (1965b). Portfolio analysis in a stable paretian market. Management Science,
artificial intelligence (XAI). IEEE, 6, 52138–52156, In press. 11(3), 404–419.
Adams, Z., Füss, R., & Glück, T. (2017). Are correlations constant? Empirical and Fama, E., & French, K. (1989). Business conditions and expected returns on stocks and
theoretical results on popular correlation models in finance. Journal of Banking bonds. Journal of Financial Economics, 25(1), 23–49.
& Finance, 84, 9–24. Gal, U., Jensen, T., & Stein, M. -K. (2020). Breaking the vicious cycle of algorith-
Adya, M., & Collopy, F. (1998). How effective are neural networks at fore- mic management: A virtue ethics approach to people analytics. Information and
casting and prediction? A review and evaluation. Journal of Forecasting, 17, Organization, 30, Article 100301.
481–495. http://dx.doi.org/10.1002/(SICI)1099-131X(1998090)17:5/6<481::AID- Gruber, M. J. (1996). Another puzzle: The growth in actively managed mutual funds.
FOR709>3.0.CO;2-Q. The Journal of Finance, 51, 783–810. http://dx.doi.org/10.1111/j.1540-6261.1996.
de Ágreda, Á. G. (2020). Ethics of autonomous weapons systems and its applicability tb02707.x.
to any AI systems. Telecommunications Policy, 44(6), Article 101953. Gruson, D., Helleputte, T., Rousseau, P., & Gruson, D. (2019). Data science, artificial
Albon, C. (2018). Machine learning kochbuch. 1. Auflage. Heidelberg: O’Reilly Media Inc. intelligence, and machine learning: Opportunities for laboratory medicine and the
Alexandridis, A. K., Kampouridis, M., & Cramer, S. (2017). A comparison of wavelet value of positive regulations. Clinical Biochemistry, 69, 1–7.
networks and genetic programming in the context of temperature derivatives. Gupta, A., & Lam, M. S. (1996). Estimating missing values using neural networks.
International Journal of Forecasting, 33, 21–47. Journal of the Operational Research Society, 47(2), 229–238. http://dx.doi.org/10.
Alexandridis, A. K., & Zapranis, A. D. (2013). Wavelet neural networks: A practical 2307/2584344.
guide. Neural Networks, 42, 1–27. http://dx.doi.org/10.1016/j.neunet.2013.01.008. Guresen, E., Kayakutlu, G., & Daim, T. U. (2011). Using artificial neural network models
Alexandridis, A. K., & Zapranis, A. (2014). Wavelet neural networks: With applications in stock market index prediction. Expert Systems with Applications, 38, 10389–10397.
in financial engineering, chaos, and classification. http://dx.doi.org/10.1016/j.eswa.2011.02.068.
Altman, D. G., & Bland, J. M. (2005). Standard deviations and standard errors. BMJ Halevy, A., Norvig, P., & Pereira, F. (2009). The unreasonable effectiveness of data.
(Clinical Research Ed.), 331(7521), 903. http://dx.doi.org/10.1136/bmj.331.7521. IEEE Intelligent Systems, 24, 8–12. http://dx.doi.org/10.1109/MIS.2009.36.
903. Hill, T., O’Connor, M., & Remus, W. (1996). Neural network models for time series
Anjoy, P., & Paul, R. K. (2017). Comparative performance of wavelet-based neural forecasts. Management Science, 42(7), 1082–1092, http://www.jstor.org/stable/
network approaches. Neural Computing and Applications, 1–11. 2634369.
Ashok, M., Madan, R., Joha, A., & Sivarajah, U. (2022). Ethical framework for Jordan, M. I., & Mitchell, T. (2015). Machine learning: Trends, perspectives, and
artificial intelligence and digital technologies. International Journal of Information prospects. Science, 349, 255–260.
Management, 62, Article 102433. Kaastra, I., & Boyd, M. S. (1996). Designing a neural network for forecasting financial
Banerjee, S. (2020). A framework for designing compassionate and ethical artificial and economic time series. Neurocomputing, 10, 215–236.
intelligence and artificial conciousness. Interdisciplinary Description of Complex Kanarachos, S., Christopoulos, S. -R. G., & Chroneos, A. (2017). Detecting anomalies in
Systems, 18(2-A), 85–95. time series data via a deep learning algorithm combining wavelets, neural networks
Bao, W., Yue, J., & Rao, Y. (2017). A deep learning framework for financial time series and Hilbert transform. Expert Systems with Applications, 85, 292–304.
using stacked autoencoders and long-short term memory. PLoS One, 12. Kumbure, M., Lohrmann, C., Luuka, P., & Porras, J. (2022). Machine learning tech-
Barlow, H. B. (1983). Vision: A computational investigation into the human repre- niques and data for stock market forecasting: A literature review. Expert Systems
sentation and processing of visual information: David Marr. San francisco: W. H. with Applications, 197, Article 116659.
Freeman, 1982. pp. xvi + 397. Journal of Mathematical Psychology, 27, 107–110. Lahmiri, S. (2014). Wavelet low-and high-frequency components as features for pre-
Berghorn, W. (2015). Trend momentum. Quantitative Finance, 15, 261–284. dicting stock prices with backpropagation neural networks. Journal of King Saud
Bernard, C., Mallat, S., & Slotine, J. -J. (1998). Wavelet interpolation networks. In Proc. University - Computer and Information Sciences, 26(2), 218–227. http://dx.doi.org/
of ESANN’ 98 (pp. 22–24). Bruges, Belgium. 10.1016/j.jksuci.2013.12.001.
12
Lawrence, S., Giles, C. L., & Tsoi, A. C. (1998). Lessons in neural network training: Sezer, O., Ozbayoglu, M., & Dogdu, E. (2017). A deep neural-network based stock
Overfitting may be harder than expected. AAAI/IAAI. trading system based on evolutionary optimized technical analysis parameters.
Li, S. -T., & Kuo, S. -C. (2008). Knowledge discovery in financial investment for Procedia Computer Science, 114, 473–480.
forecasting and trading strategy through wavelet-based SOM networks. Expert Statista Research Department 2019. Größte Börsen der Welt nach dem Handelsvolumen
Systems with Applications, 34, 935–951. mit Aktien im Jahr 2018. (Accessed 9 March 2020).
Lim, T., Loh, W., & Shih, Y. (2004). A comparison of prediction accuracy, complexity, STOXX 2021. EURO STOXX 50® INDEX 2021. (Accessed 3 March 2021).
and training time of thirty-three old and new classification algorithms. Machine Taspinar, A. (2018). A guide for using the wavelet transform in machine learning.
Learning, 40, 203–228. Accessible via https://ataspinar.com/2018/12/21/a-guide-for-using-the-wavelet-
Liu, H., Tian, H. -q., Pan, D. -f., & Li, Y. -f. (2013). Forecasting models for wind speed transform-in-machine-learning/. (Accessed 9 March 2020).
using wavelet, wavelet packet, time series and artificial neural networks. Applied
Torrence, C., & Compo, G. P. (1998). A practical guide to wavelet analysis. Bulletin of
Energy, 107, 191–208.
the American Meteorological Society, 79, 61–78.
Makridakis, S., Spiliotis, E., & Assimakopoulos, V. (2018). Statistical and machine
Tsaih, R., Hsu, Y., & Lai, C. C. (1998). Forecasting S&P 500 stock index futures with
learning forecasting methods: Concerns and ways forward. PLoS One, 13.
a hybrid AI system. Decision Support Systems, 23, 161–174.
Mallat, S. (1989). A theory for multiresolution signal decomposition: The wavelet
Tsantekidis, A., Passalis, N., Tefas, A., Kanniainen, J., Gabbouj, M., & Iosifidis, A.
representation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 11,
674–693. (2017). Forecasting stock prices from the limit order book using convolutional
Mandelbrot, B. B. (1963). The variation of certain speculative prices. Journal of Business, neural networks. In 2017 IEEE 19th conference on business informatics: Vol. 01, (pp.
36, 394–419. 7–12).
Mandelbrot, B. B., & Taylor, H. (1967). On the distribution of stock price differences. Vellido, A., Lisboa, P. J. G., & Vaughan, J. (1999). Neural networks in business: A
Operations Research, 15, 1057–1062. survey of applications (1992–1998). Expert Systems with Applications, 17, 51–70.
Misiti, M., Misiti, Y., Oppenheim, G., & Poggi, J. (2007). Wavelets and their applications. http://dx.doi.org/10.1016/S0957-4174(99)00016-0.
Wiley-ISTE Ltd. Vogl, M. 2021. Frontiers of Quantitative Financial Modelling: A Literature Review
Mitra, S. (2009). Optimal combination of trading rules using neural networks. on the Evolution in Financial and Risk Modelling after the Financial Crisis
International Business Research, 2(1), 2. (2008-2019). SSRN. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3764570.
Nath, R., & Sahu, V. (2020). The problem of machine ethics in artificial intelligence. Submitted for publication.
AI & Society, 35, 103–111. Vogl, M., & Rötzel, P. G. (2022). Chaoticity versus stochasticity in financial markets:
Nguyen, H. H., & Chan, C. W. (2004). Multiple neural networks for a long term time Are daily S&P 500 return dynamics chaotic? Communications in Nonlinear Science
series forecast. Neural Computing & Applications, 13(1), 90–98. and Numerical Simulation, 108, Article 106218.
Nobre, J., & Neves, R. (2019). Combining principal component analysis, discrete Vui, C. S., Gan, K. S., On, C. K., Alfred, R., & Anthony, P. (2013). A review of stock
wavelet transform and xgboost to trade in the financial markets. Expert Systems market prediction with Artificial neural network, ANN. In 2013 IEEE international
with Applications, 125, 181–194. conference on control system, computing and engineering (pp. 477-482).
Otuteye, E., & Siddiquee, M. (2019). Underperformance of actively managed portfolios:
Walczak, S. (2001). An empirical analysis of data requirements for financial forecasting
Some behavioral insights. Journal of Behavioral Finance, 21, 284–300.
with neural networks. Journal of Management Information Systems, 17, 203–222.
Oussar, Y., & Dreyfus, G. (2000). Initialization by selection for wavelet network
Williams, J. R., & Amaratunga, K. (1997). A discrete wavelet transform without edge
training. Neurocomputing, 34(1–4), 131–143.
effects using wavelet extrapolation. Journal of Fourier Analysis and Applications, 3,
Paliwal, M., & Kumar, U. A. (2009). Neural networks and statistical techniques: A
435–449.
review of applications. Expert Systems with Applications, 36, 2–17.
Yang, Y., & Wang, J. (2021). Forecasting wavelet neural hybrid network with financial
Parasuraman, R., Sheridan, T., & Wickens, C. (2000). A model for types and levels
of human interaction with automation. IEEE Transactions on Systems, Man & ensemble empirical mode decomposition and MCID evaluation. Expert Systems with
Cybernetics, Part A (Systems & Humans), 30(3), 286–297. Applications, 166, Article 114097.
Pati, Y., & Krishnaprasad, P. (1993). Analysis and synthesis of feedforward neural Yu, H., Shen, Z., Miao, C., Leung, C., Lesser, V., & Yang, Q. (2018). Building ethics
networks using discrete affine wavelet transforms. IEEE Transactions on Neural into artificial intelligence. In Proceedings of the 27th International Joint Conference
Networks, 4(1), 73–85. on Artificial Intelligence (pp. 5527–5533).
Peng, Y., Albuquerque, P. H. M., Kimura, H., & Saavedra, C. A. P. B. (2021). Feature Zapranis, A., & Alexandridis, A. (2008). Modelling temperature time dependent speed
selection and deep neural networks for stock price direction forecasting using of mean reversion in context of wheather derivative pricing. Applied Mathematical
technical analysis indicators. Machine Learning with Applications, 5, Article 100060. Finance, 15(4), 355–386.
Petneházi, G. (2021). Quantile convolutional neural networks for value at risk Zapranis, A., & Alexandridis, A. (2009). Weather derivatives pricing: Modelling the
forecasting. Machine Learning with Applications, 6, Article 100096. seasonal residuals variance of an Ornstein–Uhlenbeck temperature process with
López de Prado, M. (2018). Advances in financial machine learning. Hoboken: John Wiley neural networks. Neurocomputing, 73, 37–48.
& Sons Inc. Zhang, Q. (1997). Using wavelet network in nonparametric estimation. IEEE
Puchalsky, W., Ribeiro, G., Veiga, C. da., Freire, R., & dos Santos Coelho, L. (2018). Transactions on Neural Networks, 8(2), 227–236.
Agribusiness time series forecasting using wavelet neural networks and metaheuris- Zhang, Q., & Benveniste, A. (1992). Wavelet networks. IEEE Transactions on Neural
tic optimization: An analysis of the soybean sack price and perishable products Networks, 3(6), 889–898. http://dx.doi.org/10.1109/72.165591.
demand. International Journal of Production Economics, 203, 174–189. Zhang, B., Coggins, R. J., Jabri, M. A., Dersch, D. R., & Flower, B. (2001). Multiresolu-
Rashid, T. (2016). Neuronale netze selbst programmieren. 1. Auflage. Heidelberg: O’Reilly
tion forecasting for futures trading using wavelet decompositions. IEEE Transactions
Media Inc.
on Neural Networks, 12(4), 765–775.
Refinitiv Limited (2019). Thomson reuters equity indices – Corporate action
Zhang, J., Walter, G. G., Miao, Y., & Lee, W. N. W. (1995). Wavelet neural networks
methodology. (pp. 3–6).
for function learning. IEEE Transactions on Signal Processing, 43(6), 1485–1497.
Remus, W., & O’Connor, M. (2001). Neural networks for time-series forecasting. Principles
http://dx.doi.org/10.1109/78.388860.
of forecasting. New York: Springer Science and Business Media.
Sezer, O. B., Gudelek, M. U., & Ozbayoglu, A. M. (2020). Financial time series Zhou, F., Zhou, H. -m., Yang, Z., & Yang, L. (2019). EMD2FNN: A strategy combining
forecasting with deep learning: A systematic literature review: 2005–2019. ArXiv, empirical mode decomposition and factorization machine based neural network for
arxiv:abs/1911.13288. stock market trend prediction. Expert Systems with Applications, 115, 136–151.
13

Machine Learning With Applications: Markus Vogl, Peter Gordon Rötzel (LL.M), Stefan Homes

Uploaded by

Copyright:

Available Formats

Machine Learning With Applications: Markus Vogl, Peter Gordon Rötzel (LL.M), Stefan Homes

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Machine Learning With Applications: Markus Vogl, Peter Gordon Rötzel (LL.M), Stefan Homes

Uploaded by

Copyright:

Available Formats

Machine Learning with Applications 8 (2022) 100302

Contents lists available at ScienceDirect

Machine Learning with Applications

Forecasting performance of wavelet neural networks and other neural

ARTICLE INFO ABSTRACT

relevant alignment focus solely on the intermediate research stream. In Table 1

namely, high compression capabilities in computing the value at a Table 2

the MLP continues to be the best NN for one-day-ahead forecasts Table 5

6. Discussion and implications 6.1. Empirical implications

Table 7 to the data period leads to an increase in performance in general is

You might also like