A Survey of Large Language Models for Financial Applications: Progress,
Prospects and Challenges

Many earlier models, such as those based on Recurrent Neural Networks (RNNs), specifically Long Short-Term Memory (LSTM), have demonstrated a capacity for achieving a degree of language understanding over text sequences and performing textual work [63]. However, due to these models’ architectural constraints, they struggled with long-term dependencies. Specifically, they encountered challenges in maintaining context over long text sequences, understanding complex expressions, dealing with large datasets and handling unstructured data efficiently [63], [64]. This limitation is particularly evident when applying in financial sector, where the volume of documentation is vast, and the need for accurate and concise summaries is critical [65].
 
LLMs, which leverage the transformer model architecture, on another hand, have significantly advanced the field’s capabilities. The transformer architecture, characterized by its innovative self-attention mechanism, allows LLMs to process, understand and generate text based on massive datasets they have been trained on [66], [67]. This breakthrough is instrumental in overcoming the challenges faced by earlier models. By efficiently managing long-term dependencies and contextual information over large volumes of text, LLMs can streamline complex financial narratives into concise summaries and extract relevant information [66], [67]. This process retains essential insights and enables more efficient information processing.
 
Summarization and Extraction: Recent research has effectively utilized LLMs to summarize and extract financial document information [68], [69], [70]. Given that these financial documents are often lengthy, which can exceed the token limits of many LLMs, various studies have introduced frameworks by dividing long documents into shorter segments or utilized specific models to address the challenges of processing extensive financial texts [71], [72]. Recently, Yepes et al. [73] propose an expanded approach to chunking documents for Retrieval Augmented Generation (RAG) by using structural elements rather than just paragraphs, which improves chunk size determination without tuning. Furthermore, some papers propose segmenting long reports into ten distinct sections, such as management’s discussion and analysis, financial highlights, and business overview, to streamline the summarization process [74], [75]. Similarly, Khanna et al. [76] utilize the Longformer-Encoder-Decoder (LED) model, a transformer architecture first introduced by Beltagy et al. [77], which employs a self-attention mechanism scalable with sequence length, making it suitable for analyzing long financial reports.

Beyond handling the long size of the document, research has expanded into multilingual and domain-specific challenges. This includes summarizing financial documents across multiple languages [78]; customizing language models to tackle the adaptation challenges to Japanese financial terminology [79]; automating the finetuning process for text summarization models in the cryptocurrency domain without requiring human annotation [80]; adopting multitask learning strategies to classify, detect and summarize financial event [81]; addressing the challenge of ensuring accuracy and reducing errors in financial information extraction [82]; extracting information from annual report to enhance stock investment strategies [83].
 
Managing Diverse Document Structures: Despite the effectiveness of LLMs in handling textual financial data, they often face challenges with PDF document formats that incorporate images, charts, and tables. This challenge may arise from their primarily text-based nature, which finds it difficult to interpret the complex spatial layouts crucial for understanding such multimodal documents [84]. One simple approach to this issue involves converting PDF files into machine-readable plain text. For instance, in the Automated Financial Information Extraction (AFIE) framework proposed by Yue et al. [85], tables are transformed into text using PLAIN serialization. This method uses spaces and newline characters to separate cells and rows, respectively. This effectively integrated table data with regular paragraphs for LLMs to process uniformly.

However, this conversion process may change the document’s spatial layout and potentially lead to the loss of crucial information embedded in charts or tables. To address this, the team at JP Morgan developed DocLLM [86], a layout-aware generative language model specifically designed for multimodal document understanding. DocLMM utilizes bounding box information to understand the spatial arrangement of elements within the documents. It enhances document understanding by modifying the attention mechanism in transformers to concentrate on the cross-alignment between textual and spatial modalities.
 
Name-Entity Recognition: Name-entity recognition (NER) is a subtask of information extraction and plays a crucial role in extracting meaningful information from various financial sources [87], [88]. In the financial domain, it is used to extract specific entities such as company names, financial terminology, stock symbols, financial indicators, monetary values from news articles, financial reports and market summaries [89]. This information is crucial for financial downstream tasks, such as industry classification, sentiment analysis, credit scoring, fraud detection and regulatory compliance reporting [90].

Traditionally, NER is approached through Rule-Based Methods, Machine Learning Techniques, or Deep Learning Techniques [91]. Rule-based methods depend on handcrafted linguistic and grammatical rules. They offer high precision for well-defined patterns but suffer from limited scalability [87]. Machine Learning Techniques include both supervised and unsupervised approaches. Supervised approaches utilize a comprehensive set of engineered features, such as word-level characteristics and list lookups, alongside machine learning algorithms, such as Hidden Markov Models [92], Decision Trees [93] and Support Vector Machine [94], to identify and classify entities in text. Unsupervised learning approaches extract and classify named entities by employing clustering, leveraging lexical resources and patterns, and analyzing corpus statistics [91]. While machine learning offers flexibility and can handle diverse data types, it relies heavily on the availability of labeled data for supervised learning and can lack interpretability in unsupervised learning [87]. Deep learning methods utilize advanced architectures such as Bidirectional Long Short-Term Memory (BiLSTM) networks, self-attention-based transformers, and Conditional Random Fields (CRF) for tag decoding to effectively learn and represent both word and character-level features from large datasets. These approaches significantly enhance model performance by enabling the capture of complex patterns and long-range dependencies in text [87].

With the emergence of deep learning methods, LLMs are now increasingly used in NER within the financial domain [95], [96]. The ability of LLMs to leverage vast pre-trained knowledge and sophisticated language understanding can significantly enhance the accuracy and efficiency of entity recognition in complex financial texts [95]. Recently, Hillebrand et al. [97] propose KPI-BERT, a new system that utilizes advanced techniques NER and Relation Extraction (RE) to identify and connect key performance indicators (KPIs) such as ”revenue” or ”interest expenses” within German financial documents. This system relies on an end-to-end trainable architecture based on BERT. It combines an RNN with conditional label masking for sequential entity tagging, followed by relation classification. Further research has utilized LLMs for NER to improve the efficiency and accuracy of XBRL (eXtansive Business Reporting Language) tagging [98]; identify similar peer companies [99]; detect key entities of negative news information [100]; extract relevant phrases for entities [101].

Despite LLMs have demonstrated exceptional generalization capabilities, they sometimes come with high training and inference costs, especially when processing long financial documents. To address these issues, Zhou et al. [102] propose UniversalNER, a model that employs targeted distillation with mission-focused instruction tuning to train cost-efficient student models for open NER. This approach not only reduces the computational burden but also achieves remarkable NER accuracy without direct supervision.

In financial text analysis, summarizing and extracting key information from documents is crucial for quickly understanding and processing important data within lengthy and complex texts [103]. Following the extraction of pertinent information, the next step involves utilizing this information for solving downstream financial tasks. This section will introduce two main activities central to this application: constructing financial relationships and textual classification. These efforts are vital for leveraging the extracted information to enhance decision-making and analytical processes in the finance sector.
 
Financial Relation Construction: Constructing financial relationships, particularly through the use of knowledge graphs, represents a powerful methodology for organizing and making sense of the extracted entities and their interrelations from extensive and complex financial datasets [104]. Knowledge graphs consist of interconnected descriptive structures about entities (objects, events, people, etc.), the attributes of those entities, and the relationships that link them together. This framework offers a structured way of representing relationships within data and enables sophisticated analyses to be derived from them [105], [106].

Upon the identification and extraction of entities (such as companies, individuals, financial instruments, events, etc.), along with the relationships between these entities (such as ownership, transactions, legal disputes, etc.), this information can be systematically organized into a graph format for further construction. Within a knowledge graph, entities are represented as nodes, and relationships are denoted as edges that connect these nodes. This structure provides a visual and programmable method to explore and comprehend the connections among different entities within the financial ecosystem. With the construction of the knowledge graph, financial analysts and systems can employ graph analytics and machine learning algorithms to discover insights, recognize patterns, and predict future events [107].

Recent advancements in LLMs have led researchers to explore the potential of using information extracted by LLMs to construct and analyze knowledge graphs in the financial sector [108], [109], [110]. Notably, Trajanoska et al. [108] generate a knowledge graph by leveraging LLMs to extract structured Environmental, Social, and Governance (ESG) information from sustainability reports, using a format of triples consisting of node-edge-node, to enable deeper analysis and understanding of corporate sustainability practices. Similarly, Cheng et al. [111] develop a Semantic-Entity Interaction Module. This module combines a language model with a Conditional Random Field (CRF) layer to comprehend the interaction between entities and their semantic contexts in texts. It automatically constructs financial knowledge graphs from brokerage research reports without the need for explicit financial knowledge or extensive manual rules.

Moreover, financial research analysts often face challenges in identifying critical documents, key entities, and important events during their research on complex financial subjects. Mackie and Dalton [112] tackle these issues by developing automated methods to create detailed, query-specific knowledge graphs from documents and entities.

As illustrated above, knowledge graphs have demonstrated their utility in information retrieval. A special case within this domain is the translation of Natural Language (NL) into Graph Query Language (GQL). This process enhanced querying experiences by leveraging the relational data within knowledge graphs, offering advantages over traditional text-to-SQL methods. However, this approach is challenged by the complexity of accurately mapping NL to GQL syntax and the lack of domain-specific examples, making it difficult to fine-tune LLMs for precise alignment with graph databases in specialized fields [106]. To address this, Liang et al. [113] develop a pipeline that employs LLMs to generate NL-GQL pairs from financial graph databases without labeled data. This process involved creating template pairs with ChatGPT and refining them through a self-instruction method. Subsequently, LLMs were fine-tuned with these pairs using the LoRA technique to align the models with the specific knowledge contained in graph databases.

Knowledge graphs can also be used to significantly enhance question-answering systems. Wang et al. [114] introduce an innovative Knowledge Graph Prompting (KGP) for multi-document question answering (MD-QA). Their approach constructs a knowledge graph from multiple documents, highlighting semantic or lexical relationships between passages or document structures. An LLM-based graph traversal agent then uses this knowledge graph to gather contextually relevant information, thereby enhancing the LLM’s accuracy in answering questions

Another beneficial aspect of knowledge graphs is their ability to be enriched over time through the use of LLMs. Li [115] presents FinDKG, a dynamic knowledge graph with LLMs used in financial domain. FinDKG incorporates a temporal layer in its structure, which allows it to reflect and adapt to changes in financial markets, economic indicators, and thematic trends. This dynamic approach provides valuable insights for thematic investing, making it possible to identify and leverage long-term industry shifts and economic trends for strategic investment decision-making.

There exist other financial relation extraction studies using LLMs, though not necessarily for knowledge graph construction [116], [117], [118], [119]. Ghosh et al. [120] propose the Mask One At a Time (MOAT) framework, which masks one entity at a time, extracts contextual embeddings using a domain-specific language model (SEC-BERT), and combines these embeddings with additional features to train a neural network for accurately classifying relationships between financial entities. Similarly, Rajpoot and Parikh [121] employ in-context learning with GPT models, utilizing both a learning-free dense retriever (KNN with OpenAI embeddings) that relies on the similarity of embeddings to find the most relevant examples, and a learning-based retriever trained to select the most similar example in the training set for each test example by estimating the probability of the output given the input and a candidate training example as the prompt. Focusing on multi-type Chinese financial event relation extraction, Wan et al. [122] propose the CFERE framework, which employs a Core Verb Chain for event identification, constructs a Syntactic Semantic Dependency Parsing graph to combine events into pairs, and enhances BERT with an Event Core Embeddings layer to capture semantic meanings. These studies demonstrate the potential of LLMs and innovative approaches in advancing financial relation extraction, ultimately contributing to the research value of making use of financial information and helping investors make better investment decisions.
 
Textual Classification: Textual classification plays a crucial role in organizing and understanding large volumes of unstructured data within the financial domain. This classification task can be further categorized into several sub-tasks, such as industry/company classification and document/topic classification. By effectively classifying and organizing this information, businesses and researchers can extract valuable insights and make informed decisions. The utilization of these classification techniques, in conjunction with the establishment of financial relationships, is essential for leveraging the extracted information to enhance decision-making and analytical processes within the finance sector.

Company or industry classification involves grouping companies into distinct categories based on shared characteristics such as business activities and market performance, with the aim of creating coherent and differentiated groups. Identifying similar company profiles is a fundamental task in finance, with applications spanning investment portfolio construction, securities pricing, and financial risk attribution. Traditionally, financial analysts have relied on industry classification systems, such as the Global Industry Classification System (GICS), the Standard Industrial Classification (SIC), the North American Industry Classification System (NAICS), and the Fama French (FF) model, to identify companies with similar profiles [123]. However, these systems do not provide a means to rank companies based on their degree of similarity and require time-consuming, effort-intensive manual analysis and data processing by domain experts [123].

Recently, a team at BlackRock [124] explores a novel approach to company classification using LLMs. They investigated the use of pre-trained and fine-tuned LLMs to generate company embeddings based on business descriptions from SEC filings. Their study aimed to assess the embeddings’ ability to reproduce GICS classifications, benchmark LLM performance on various downstream financial tasks, and examine the impact of factors such as pre-training objective, fine-tuning, and model size on embedding quality. The results showed that LLM-generated embeddings, particularly those from fine-tuned Sentence-BERT models, could accurately reproduce GICS sector and industry classifications and outperform them on tasks like identifying similar companies based on return correlations and explaining cross-sectional equity returns.

Interestingly, knowledge graphs can also be used to enrich industry classification and improve the performance of domain-specific text classification tasks. Wang et al. [125] propose a novel Knowledge Graph Enriched BERT (KGEB) model that integrates external knowledge from a local knowledge graph with word representations. They demonstrated the effectiveness of their approach by constructing a large dataset based on companies listed on the Chinese National Equities Exchange and Quotations (NEEQ) and showing that the KGEB model outperforms competitive baselines, including graph convolutional network, Logistic Regression, TextCNN, BERT, and K-BERT, achieving an accuracy of 91.98% and an F1 score of 90.89%.

Document or topic classification is another crucial sub-task within the broader scope of textual classification in the financial domain. This task involves categorizing financial documents or texts, such as news articles [126], [127] or company filings [128], [129], into predefined topics or themes. Alias et al. [130] propose a novel approach that utilizes the FinBERT model to extract and categorize relevant topics of Key Audit Matters (KAM) from the annual reports of publicly listed companies in Bursa Malaysia. Similarly, Burke et al. [131] fine-tune the FinBERT model to classify accounting topics within three unlabelled financial disclosures, including custom notes to the financial statements, the Management’s Discussion and Analysis section, and the risk factor section.

Another important classification task in the financial domain involves categorizing Environmental, Social, and Governance (ESG) information. This task requires identifying and classifying ESG-related data, such as carbon emissions, diversity and inclusion, and corporate governance practices, from multiple sources including corporate sustainability reports, news articles, and social media posts. In a recent study, Lee and Kim [132] propose an ESG classifier that can discriminate ESG information by fine-tuning a pre-trained language model. The classifier was trained on a manually labeled dataset constructed from sustainability reports of Korean companies across five sectors and achieved a classification accuracy of 86.66% for the four-class classification problem (Environment, Social, Governance, and Neutral). Similarly, Mehra et al. [133] develop a domain-specific language model called ESGBERT to enhance the classification of ESG-related text by fine-tuning BERT’s pre-trained weights using ESG-specific text and further fine-tuning the model for classification tasks.

Textual classification techniques, including industry/company classification and document/topic classification, play a vital role in organizing and understanding large volumes of unstructured data within the financial domain. Recent advancements in LLMs and knowledge graph integration have significantly improved the accuracy and efficiency of these classification tasks. The successful application of these techniques can further provide valuable insights and support informed decision-making in various financial contexts, such as investment portfolio construction, risk assessment, and ESG analysis.

First, we outline the significant milestones in sentiment analysis in this section, leading up to the era before LLMs like ChatGPT and BERT revolutionized the field. Additionally, it highlights key applications within the financial domain, demonstrating the impact of sentiment analysis on various applications.
 
Lexicon-Based Methods: Early sentiment analysis relied on lexicon-based approaches, where the sentiment of a text was inferred based on the presence of predefined words associated with positive or negative sentiments. These methods, simple yet effective for certain applications, include the General inquirer [137], the Linguistic Inquiry and Word Count (LIWC) lexicons [138], SO-CAL [139], and Loughran and McDonald’s (LM) word lists [140].

One of the strengths of lexicon-based methods is their simplicity and interpretability. However, their performance can be limited by the context-dependency of sentiment expressions and the inability to capture the sentiment expressed by complex linguistic constructs such as sarcasm or irony. Despite these limitations, lexicon-based methods have been effectively applied in finance, particularly in analyzing investor sentiment from financial news or social media content [141], [142], [143].
 
Machine Learning Methods: With the advent of machine learning, financial sentiment analysis (FSA) experienced significant advancements. ML-based methods can be broadly categorized into supervised and unsupervised learning. When doing FSA, supervised learning approaches require labeled data and include techniques such as Support Vector Machines (SVM) [144], Naive Bayes [145], KNN (K-Nearest Neighbor) [146], Random Forests [147] and Multi-layer perceptrons (MLPs) [148]. Unsupervised learning, in contrast, does not require labeled data and typically involves clustering techniques to discern sentiment [149].

In finance, ML has been used to predict market movements based on sentiment in financial news and social media, showcasing its ability to capture financial sentiment nuances [166]. Machine learning methods offer the advantage of being able to capture complex patterns in data that are not apparent to lexicon-based methods. However, they require large datasets for training, and the versatility is limited on a specific domain.
 
Embedding-Based Methods: The introduction of word embeddings marked a significant milestone in general sentiment analysis. Embedding-based methods represent textual information in a high-dimensional space where semantically similar words are closer together. This representation captures not only the sentiment but also the context of words, leading to improved performance in sentiment analysis tasks. Mikolov et al.’s introduction of Word2Vec [167] in 2013 was a pioneering development in this domain. Word2Vec employs neural networks to learn word associations from large datasets, generating embeddings that capture a wide array of linguistic relationships and nuances. The innovative aspect of Word2Vec lies in its ability to learn high-quality word vectors from vast datasets efficiently. It offers two architectures for this purpose: Continuous Bag of Words (CBOW) and Skip-gram. CBOW predicts target words from context words, while Skip-gram does the opposite, predicting context words from a target word, making it particularly effective for capturing semantic and syntactic word relationships.

Subsequent to Word2Vec, several other embedding models have emerged, further advancing the field. Notable among these are Global Vectors for Word Representation (GloVe) [168], which introduces an unsupervised learning algorithm for obtaining vector representations of words through aggregating global word-word co-occurrence statistics from a corpus; FastText [169], which extends Word2Vec to consider subword information, thereby enhancing the representation of rare words; and Embeddings from Language Models (ELMo) [170], which leverages bidirectional language models to generate contextually enriched word embeddings.

Beyond word-level embeddings, there has been a push towards capturing longer contextual dependencies. An exemplar in this area is Doc2Vec, also known as Paragraph Vector, introduced by Le and Mikolov [171]. Doc2Vec extends the Word2Vec paradigm to support document-level embeddings, enabling the capture of document-wide contextual information which is crucial for tasks requiring comprehension of extended textual content. By learning fixed-length feature representations from variable-length pieces of texts, Doc2Vec facilitates a deeper understanding of document semantics, thereby broadening the applicability of embedding techniques in sentiment analysis and beyond.

Embedding-based methods have the advantage of capturing contextual complexity and semantic relationships between words, significantly improving the accuracy of sentiment analysis. This has made them popular in FSA as well. Sohangir et al. [172] highlight the effectiveness of these methods in financial domains, demonstrating their ability to extract sentiment from large volumes of unstructured financial data with high accuracy.

However, they are not without drawbacks. A notable limitation is their dependency on large datasets for training, which might not always be feasible in specialized domains. Additionally, while adept at semantic understanding, they may overlook slight differences in syntax and require retraining to adapt to new language uses or vocabularies. Pre-trained embeddings can also perpetuate biases present in their training data, leading to potential issues in fairness and representation. Despite these challenges, embedding-based methods are crucial in advancing natural language understanding and have paved the way for large language models like BERT and GPT-3, which build on these embeddings to achieve state-of-the-art NLP performance.

The advent of ChatGPT and other LLMs represents a pivotal milestone in the domain of FSA. Nowadays, these models have demonstrated their effectiveness in numerous tasks and offer several unique advantages for FSA applications.

Firstly, LLMs excel in deciphering the complexities of financial language, adeptly navigating informal expressions, emojis, memes, and specialized terminology across social media and financial blogs [58], [60], [173], [174], [175], [151]. Their proficiency in identifying subtleties like irony, sarcasm, and sector-specific jargon is vital for accurately analyzing sentiments across various formats, from tweets to comprehensive financial reports [176], [6].

Secondly, LLMs’ ability and great potential to process multimodal data, including images, audio, and video, are essential for comprehensive sentiment analysis in financial contexts like earnings calls [155] and FOMC Meetings [177]. This capability allows for the integration of non-verbal cues and visual data into the sentiment analysis process [37].

Thirdly, LLMs’ ability to process extensive documents enables thorough analysis of detailed financial reports and lengthy articles, ensuring no sentiment-bearing information is overlooked. This feature is particularly beneficial for evaluating the sentiments expressed in annual reports, earnings transcripts, and extensive financial narratives [157].

Moreover, LLMs exhibit enhanced resilience to adversarial attacks or deceptive information tactics that could be encountered in FSA tasks. Their advanced algorithms and broader contextual understanding help in identifying and mitigating misleading or manipulative sentiment indicators, enhancing the reliability of sentiment analysis outcomes. Leippold [156] highlights the contrast between traditional keyword-based sentiment analysis methods and LLMs in the face of adversarial attacks. The research involved using GPT-3 to substitute negative words with synonyms to assess model robustness, showcasing FinBERT’s enhanced resilience against adversarial attacks over traditional keyword-based methods.

We further delve into the recent advancements in the integration of LLMs within FSA, categorically analyzing the impact and contributions according to diverse data sources. We embark on this exploration by categorizing the data into four key segments: Social Media and News, Corporate Disclosures, Market Research Reports, and Policy and Economic Indicators. This structured approach enables a comprehensive understanding of how LLMs have revolutionized the domain of FSA, offering unprecedented insights and analytics capabilities.
 
Social Media and News: Social media platforms like Twitter, general online forums like Reddit, and finance-specific forums such as StockTwits, along with financial blogs and microblogs, have become rich sources of data for FSA. These platforms are crucial due to their rich repositories of real-time, unstructured textual content that mirrors public sentiment regarding financial markets, specific stocks, and the overall economic environment. The immediacy and public nature of the discussions on these platforms make them an invaluable resource for capturing the mood of the market, which can be predictive of future market movements. Su et al. [150] leverage BERT for extracting sentiment and semantic insights from Twitter, facilitating improved covariance estimation and enhancing portfolio optimization. The integration of text-derived covariance data into mean-variance optimization resulted in superior performance in this work, especially during COVID-19 crash period. Furthermore, Steinert and Altmann [60] employ GPT-4 for sentiment analysis of microblogging messages on the Stocktwits platform, outperforming the naive buy-and-hold strategy for Apple and Tesla stocks by a significant margin, which underscores the potential of LLMs in predicting stock price movements through sentiment analysis. Despite the efficacy of LLMs in sentiment analysis, social media sources present unique challenges, including the vast volume of information, the colloquial language often used, possible selective bias, and the presence of misinformation or inaccuracies in the messages shared, which complicate the task of accurately capturing and interpreting market sentiments [178].

News represents another crucial data source, which shares many similarities with social media in terms of rapid dissemination and broad reach, but it generally focuses more on objective events. Contrary to the often subjective and personal nature of social media, news content typically originates from more prestigious and established media outlets, including the renowned newspapers like The New York Times, television broadcasters like CNN and BBC, as well as finance-specific publications such as The Economist. The credibility and professionalism of journalists and writers in these outlets lend a higher trustworthiness to the content, despite sometimes at the cost of timeliness. Evidence increasingly supports the advantages of post-ChatGPT LLMs over earlier approaches, particularly in analyzing the sentiment of news headlines. Lopez-Lira and Tang [152] investigate ChatGPT’s effectiveness in predicting stock market returns, illustrating its capability to accurately assign sentiment scores to headlines and outperform earlier models such as GPT-2 and BERT. Besides, Fatouros et al. [153] reveal that GPT-3.5 offers considerable improvements over FinBERT in analyzing forex-related news headlines. Similarly, Luo and Gong [154] report noteworthy success with the open-source Llama2-7B model [26], achieving performance that exceeded previous BERT-based approaches and conventional methods like LSTM with ELMo. These researches underscore the significance of advanced LLMs in decision-making and quantitative trading.

In this digital age, the phenomenon of real-time news is becoming increasingly prevalent. Distributed via live broadcasts or online platforms, these news sources manage to strike a balance between accuracy and immediacy, offering timely insights into market conditions and public events that could influence financial sentiments [179]. Chen et al. [180] investigate using advanced LLMs like BERT, RoBERTa, and OPT for sentiment analysis and stock prediction. These models significantly outperform traditional methods such as Word2vec by capturing complex text information and providing a more accurate contextual understanding. It also demonstrates that LLM-based models achieve higher Sharpe ratios and better performance. Crucially, the research reveals that news information is incorporated into stock prices with a delay due to limits-to-arbitrage, creating opportunities for real-time trading strategies to exploit these inefficiencies. This underscores the potential of LLMs in real-time financial text mining.
 
Corporate Disclosures: Corporate disclosures is increasingly recognized for its significance in FSA. This section delves into three primary categories of corporate disclosures: Earnings Calls, Corporate Communications, and Regulatory Filings and Legal Documents (e.g., SEC filings), each highlighted for its importance and accompanied by pertinent studies.

Earnings calls are crucial for providing insights into a company’s financial health, strategic direction, and management’s perspective on performance and future prospects. The sentiment analysis of earnings calls transcripts can reveal underlying tones and sentiments that may influence investor decisions and market perceptions. Cook et al. [155] evaluate the performance of local LLMs in interpreting financial texts, particularly focusing on analyzing the tone and content of bank earnings calls during the post-pandemic era and the early 2023 banking stress. They show that local LLMs are effective for analyzing financial communications, demonstrating a shift in bank earnings call content towards more homogeneity and less positive sentiment during periods of increased banking stress. Leippold [156] demonstrates the susceptibility of financial sentiment analysis to adversarial attacks using GPT-3, highlighting the need for LLMs to ensure the reliability of AI in financial text processing.

Corporate communications encompass a wide range of official statements, press releases, and announcements made by a company to its stakeholders. The sentiment embedded within these communications can significantly affect how stakeholders perceive the company’s current state and future outlook. LLMs can process these communications to assess the sentiment and identify potential market-moving information. For instance, Kim et al. [157] illustrate that ChatGPT can significantly streamline and clarify corporate disclosures for investors by reducing the length and amplifying the sentiment of the content, while also revealing the prevalent issue of ’bloat’—excessive, redundant, or irrelevant information in financial reports—that can obscure the true insights needed for informed investment decisions.

Regulatory Filings and Legal Documents are essential for compliance, governance, and transparency, providing a wealth of information on a company’s operations, risks, and financial condition. LLMs can process these complex documents and identify sentiment-related information, such as litigation risks, accounting irregularities, and management changes. Aparicio et al. [158] introduce BioFinBERT, a fine-tuned language model that utilizes sentiment analysis of regulatory filings and legal documents, such as 10-Q, 10-K, 6-K, and 20-F reports, along with biotech company press releases, to execute market orders and predict stock price movements in the biotechnology sector. Another paper [159] investigates how corporations adjust their regulatory disclosures to be more machine-readable in the age of AI, influencing both the sentiment expressed and the speed of information dissemination in financial markets.
 
Market Research Reports: Market research reports, which encompass a wide range of data including economic indicators, industry analysis, and consumer behavior, are crucial for informed decision-making in finance. The significance of analyst reports and investment research lies in their detailed analysis and recommendations on securities, offering a profound understanding of market trends and potential investment opportunities. Analyst ratings, such as ”buy,” ”hold,” or ”sell” recommendations, provide another concise evaluation of a security’s future performance, serving as a valuable guide for investors. These ratings are based on rigorous financial analysis and are closely monitored by investors to assess market sentiment and make strategic investment choices [160].
 
Policy and Economic Indicators: In the field of financial sentiment analysis, particularly with respect to policy and economic indicators, a significant focus has been placed on the analysis of Federal Open Market Committee (FOMC) meeting minutes, European Central Bank (ECB) policy decisions, and other key indicators like non-farm payroll data, unemployment rates, inflation rates, and GDP growth. These sources are critically important for understanding the market dynamics and guiding investment decisions based on the sentiment derived from policy decisions and economic reports.

The FOMC meeting minutes are a important source of information for understanding the U.S. Federal Reserve’s monetary policy stance [181], [182]. These minutes provide a detailed account of the discussions and deliberations that take place during FOMC meetings, shedding light on the economic outlook, inflation expectations, and potential interest rate changes [161]. Researchers have employed LLMs to analyze the sentiment and tone of FOMC meeting minutes. Kim et al. [162] investigate that while FinBERT outperforms traditional techniques in predicting negative sentiment within FOMC statements, there is a need for further enhancements and exploration of alternative approaches to optimize the analysis of FOMC texts and gain more comprehensive economic insights. Gössi et al. [163] present a fine-tuned FinBERT model with a Sentiment Focus method, which significantly improves the sentiment analysis accuracy of complex financial sentences in FOMC Minutes, particularly those containing conjunctions with contradicting sentiments.

The ECB is responsible for setting monetary policy for the eurozone, and its policy decisions have a significant impact on financial markets [183]. ECB policy decisions, including interest rate changes and asset purchase programs, are closely monitored by investors and analysts [184], [185]. Recent research has utilized LLMs to analyze the sentiment and impact of ECB policy decisions on financial markets [164]. Utilizing the FinBERT model, Kanelis and Siklos [165] reveal that sentiments from monetary policy speeches explain the tone of press conference statements, while financial stability speeches offer little explanatory power, highlighting the LLM’s ability to provide detailed sentiment analysis in economic communication.

In addition to FOMC meeting minutes and ECB policy decisions, several other economic indicators and research papers are relevant to FSA. Non-farm payroll data and unemployment rates provide insights into the labor market and can have a significant impact on market sentiment [186]. Inflation rates and GDP growth are also closely watched indicators, as they reflect the overall health of the economy [187], [188]. Applying LLMs to analyze the sentiment and impact of these economic indicators on financial markets deserves further exploration for future research.

Deep learning has revolutionized time series analysis, offering powerful tools for modeling and forecasting sequential data [189], [190], [191]. Prominent deep learning models like LSTM networks and CNNs have demonstrated significant effectiveness in capturing temporal dependencies and anomalies in time series data [192], [193], [194].

With the recent surge in popularity of LLMs, these tools are increasingly being utilized to assist in time series tasks [195], [196]. They offer a multitude of auxiliary functions such as generating additional features from textual data and producing descriptive statistics, as we have discussed in Section 3.1 and 3.2, which can enhance the accuracy of time series models by tapping into a broader spectrum of information beyond the original data.

Beyond these supportive roles, LLMs are also being employed to directly analyze time series data [197], [198], a development supported by several factors. This is primarily attributed to LLMs’ ability to understand and process sequential data, which is a common trait between text and time series. Also, the prevalent Transformer architecture underlying most LLMs has proven effective in various time series tasks [199], [200], [201]. Furthermore, LLMs exhibit remarkable multimodal capabilities, suggesting that their pre-training on vast datasets, even if solely text-based, imparts general inference and reasoning abilities beyond the specific data modality [202]. This characteristic not only serves as supportive evidence to the direct application of LLMs in time series analysis but also paves the way for future multimodal foundation models [203].

Several notable works have demonstrated the efficacy of LLMs in time series analysis. Pioneering efforts by Zhou et al. [204] demonstrate the versatility of LLMs across tasks such as forecasting, anomaly detection, classification, and imputation. Using a GPT-2 backbone, they establish the potential for LLMs to process and model time series data effectively. Gruver et al. [205] further explore the zero-shot capabilities of pretrained LLMs for time series forecasting. Through appropriate tokenization of time series data, they found that LLMs can implicitly understand temporal patterns and generate forecasts without explicit training. Jin et al. [206] apply the concept of reprogramming to enhance LLM performance in time series analysis. This technique translates time series data into representations more readily understood by LLMs, leading to state-of-the-art forecasting results. Beyond direct LLM applications, researchers are focusing on developing foundation models specifically for time series analysis [207], [208]. These efforts aim to establish a new paradigm for time series modeling, leveraging the power of techniques in LLMs to capture complex temporal dependencies.

Recent research has explored the utility of LLMs in the domain of financial time series forecasting, demonstrating both the potential and the limitations of these advanced computational tools. This section reviews key studies that have contributed to our understanding of how LLMs can be applied to predict stock market movements and other financial indicators.

LLMs can be directly used for stock forecasting, as stated in [209]. Their research explores NASDAQ-100 stock prediction with LLMs, and demonstrates that by integrating diverse data sources, LLMs not only provide robust predictions but also enhance the explainability. The study emphasizes the importance of instruction-based fine-tuning and chain of thought reasoning, which have been shown to significantly improve the performance of LLMs over traditional statistical models in this field. Another way is to integrate LLMs to enhance the other neural networks. Chen et al. [210] introduced a framework that leverages ChatGPT to enhance graph neural networks (GNN) for stock movement prediction. Their approach adeptly extracts evolving network structures from textual data and incorporates these networks into GNNs for predictive tasks. The experimental results indicate that this model consistently outperforms state-of-the-art deep learning-based benchmarks with higher annualized cumulative returns and reduced volatility.

Besides, LLMs are notable for their capability to be integrated in multimodal data analysis as discussed in the previous section, which can be crucial when analyzing alternative data. For instance, Wimmer and Rekabsaz [211] introduce innovative models that leverage both textual and visual data to forecast market movements. Utilizing CLIP-based models, their research shows significant outperformance against established benchmarks in predicting the trends of the German share index. Metrics such as Precision, F1 Score, Balanced Accuracy, and others show the effectiveness of these multimodal approaches. Another noteworthy study is the RiskLabs framework, which combines various types of financial data, including textual and vocal information from earnings conference calls, market-related time series data, and contextual news data [212]. The framework’s multi-stage process starts with extracting and analyzing these data using LLMs, followed by processing time-series data to model risk over different timeframes. RiskLabs employs multimodal fusion techniques to combine these varied data features for comprehensive multi-task financial risk prediction. The empirical results demonstrate the framework’s effectiveness in forecasting both volatility and variance in financial markets, indicating the potential of LLMs in financial risk assessment.

However, the application of LLMs in financial forecasting is not without challenges. Xie et al. [213] specifically assess ChatGPT’s performance in zero-shot multimodal stock movement prediction tasks and find that it underperforms when compared to both traditional machine learning models and other state-of-the-art techniques. Their findings highlight the necessity for ongoing research to enhance the predictive capabilities of LLMs in complex financial environments. On the other hand, Lopez-Lira and Tang [152] examine how well these models, particularly GPT-4, can predict stock market returns using news headlines as input. Their results indicate that advanced LLMs significantly outperform both traditional models and earlier versions of LLMs. Notably, the models show higher efficacy especially after negative news and for smaller stocks, a phenomenon explained through theories of information diffusion, arbitrage limitations, and investor sophistication. The debate on the effectiveness of LLMs in financial forecasting remains open, with evidence supporting both their limitations and potential.

Though early challenges exist, research reveals considerable promise for LLMs in financial time series forecasting. Explainability, comprehensive understanding of news, and multimodal integration stand out as compelling areas for future investigation and refinement. However, they also mark the challenges and the necessity for further research to fully realize the potential of LLMs in this domain.

Anomaly detection is a fundamental task in various domains, particularly in finance where identifying unusual patterns or outliers is crucial [214]. For instance, identifying fraudulent transactions or unusual account activity is a top priority for financial institutions. Anomaly detection algorithms can flag potentially fraudulent behavior, preventing financial losses [215]. Besides, market manipulation schemes, such as pump-and-dump tactics, can be detected through anomaly detection in trading volumes and price patterns [216]. Anomaly detection is also valuable in risk assessment and mitigation strategies, since anomalies in market trends or macroeconomic indicators can signal potential risks.

Financial time series data, like stock prices, can be highly complex, characterized by volatility, seasonality, and non-linear relationships. Traditional statistical approaches, though robust, often struggle to encapsulate the full spectrum of these complexities, thereby constraining their anomaly detection capabilities. The development of deep learning has catalyzed a fundamental transformation, offering novel methodologies that hold great promise for this domain [217], [218]. Particularly, LLMs have emerged as a pivotal method, demonstrating remarkable efficacy in anomaly detection across a myriad of tasks, as evidenced by recent scholarly works [217], [219]. For instance, Park [220] introduces an LLM-based multi-agent framework that synergizes traditional statistical methods with AI-driven analytics. This innovative fusion is exemplified through an application to the S&P 500 index, showcasing a marked enhancement in the efficiency, precision, and automation of anomaly detection in financial markets, thereby diminishing the dependency on manual interventions. The integration of LLMs into financial time series anomaly detection will likely become increasingly valuable, which has the potential to not only address the limitations of conventional techniques, but also reduce manual processes and enhance algorithmic trading systems that capitalize on market anomalies, paving the way for more sophisticated and automated trading systems.

Besides forecasting and anomaly detection, the capabilities of LLMs offer promising potential within several other domains of financial time series analysis.
 
Classification: Financial time series can be classified into various categories based on trends, volatility, or other characteristics. LLMs can learn these complex patterns and assign labels accordingly. For instance, they could classify stocks as ”growth” or ”value,” or identify different market regimes (bullish, bearish, etc.) [221]. LLMs can efficiently classify financial time series data by understanding and predicting patterns that are indicative of specific financial behaviors. This includes the applications of sentiment analysis (section 3.2) and anomaly detection (section 3.3.3) that we have already discussed.
 
Data Augmentation: The limited size and variability of financial datasets can sometimes hinder machine learning models. Generative AI offers a path toward data augmentation, which involves generating synthetic data that can be used for training machine learning models, ensuring robustness despite the original limitations of the dataset. A recent paper by Nagy et al. [222] introduces a generative AI model for end-to-end limit order book modeling, demonstrating the use of a token-level autoregressive generative model to produce realistic order flow in financial markets. This model utilizes structured state-space layers to efficiently handle long sequences of order book states and tokenized messages. The model shows promising performance in approximating data distribution and forecasting mid-price returns, suggesting potential applications in high-frequency financial reinforcement learning. While this work focuses on generative AI rather than directly employing LLMs, its approach and insights are relevant for augmenting financial time series data, highlighting the versatility of generative models in this domain. By simulating various market scenarios, LLMs can help in creating a richer, more diverse dataset that aids in building more accurate predictive models [223].
 
Imputation: Financial time series data often suffers from missing values due to errors or unavailability. Imputation refers to the method of filling in missing or incomplete data points in financial time series. LLMs have a good potential to fill these missing values based on their superior generative capability [224]. This is particularly useful in maintaining the quality and continuity of financial data analysis. Accurate imputation helps in avoiding biases or inaccuracies that might occur due to gaps in the data, thus ensuring more reliable financial assessments and forecasts.
 
In summary, LLMs demonstrate significant potential in financial time series analysis, offering capabilities in forecasting, anomaly detection, pattern classification, data augmentation, imputation, and more. Their ability to process and understand complex financial data opens avenues for novel approaches to market analysis. As LLM research progresses, we can anticipate continued advancements in the application of these models within the financial time series domain.

Financial planning involves setting financial goals, assessing current financial situations, and devising strategies to achieve those goals. This process includes analyzing income, expenses, investments, and risk management to create a comprehensive plan for long-term financial stability and growth.

In a corporate context, LLMs can be utilized to support various aspects of financial planning. For instance, LLMs can analyze market trends and competitor data to help organizations develop business strategies. Nguyen and Tulabandhula [225] examine the use of generative AI models, such as GPT-4 and other transformer-based models, for business strategy development. By employing Named Entity Recognition (NER) and Zero-shot Classifiers (ZSC) to automatically extract and classify relationships between companies, they created dynamic signed business networks that reflect the competitive and collaborative market landscape. This method provides business stakeholders with insights into market conditions and supports strategic decision-making.

Moreover, LLMs can streamline financial planning processes, as demonstrated by Ludwig and Bennetts [226]. By integrating ChatGPT into financial planning practices, they illustrate how financial planners could leverage this AI model to enhance client communication and provide immediate, semi-personalized responses to common financial concerns, such as preparing for economic recessions. They also highlight ChatGPT’s role in client education and its ability to simplify complex financial concepts for better understanding. Despite these benefits, the authors emphasize the need for human oversight to ensure the accuracy and quality of the advice provided, addressing potential limitations of the models.

In personal financial planning, LLMs can help individuals create customized strategies for long-term financial well-being. A recent study by Lakkaraju et al. [227] evaluates the performance of LLM-based chatbots, ChatGPT and Bard, in providing personal financial advice. The study covers various aspects of personal finance, including decisions related to bank accounts, credit cards, and certificates of deposits (CDs). It assesses how these models handle complex financial interactions and make recommendations across different languages and dialects, such as English, African American Vernacular English, and Telugu. Their findings reveal that while ChatGPT often provides more personalized and accurate responses, both models face challenges, including mathematical errors, lack of visual aids to support explanations, and difficulties in processing non-English queries effectively. The paper emphasizes the need for improvements in these LLMs to enhance their reliability and inclusivity when applied to financial planning, a topic that will be further discussed in section 5.

Additionally, LLMs can optimize budgeting strategies by incorporating AI-driven recommendations into individual and household financial models. de Zarzà et al. [228] present an optimization framework for individual budget allocation to maximize savings and extend this approach to household finances, addressing the complexities of multiple incomes and shared expenses. In high-net-worth contexts, LLMs can also be used to simulate various tax scenarios, identify optimal tax strategies, and provide proactive advice based on changing tax law to minimize tax liabilities and maximize financial growth [229].

The integration of LLMs in financial planning has the potential to transform how individuals and businesses approach their financial objectives. By leveraging the data processing and analysis capabilities of LLMs, financial planning can become more efficient, accurate, and personalized. As research and development in this field continue to progress, LLMs are poised to become vital tools in the financial planning environment, allowing users to make educated and strategic decisions. The examples discussed in this section highlight the diverse range of applications and the potential for LLMs to revolutionize financial planning practices for both corporate and personal contexts.

LLMs are revolutionizing investment recommendations and wealth management by analyzing financial data, forecasting market trends, and optimizing portfolios. They provide personalized advice based on individual risk profiles and preferences, which improves robo-advisors and investment strategies. However, the integration of LLMs in wealth management needs regulatory frameworks to assure fairness, effectiveness and informed decision-making in conjunction with human expertise.
 
LLM in Investment Advisory: LLMs play a crucial role in enhancing the capabilities of robo-advisors by providing personalized and automated investment recommendations. For example, Huang et al. [230] highlight the effectiveness of platforms like Wealthfront and Betterment, which employ AI algorithms to deliver customized asset management plans aimed at optimizing investment returns based on individual user profiles. The study emphasizes the importance of consistent use, transparency, and user-centric design in maximizing the benefits of intelligent advisors. To build user trust and improve the overall effectiveness of robo-advisors, the authors recommend focusing on key areas such as enhancing transparency, designing intuitive user interfaces, and offering tailored financial services for individual needs.

Similarly, Lu et al. [231] explore the potential of ChatGPT in generating investment portfolio recommendations. Using textual data from the Wall Street Journal and Chinese policy announcements, the researchers evaluate ChatGPT’s ability to generate portfolios that outperform the market. Through fine-tuning and performance measurements, the study demonstrates that ChatGPT can achieve a monthly three-factor alpha of up to 3%, particularly when analyzing policy-related news. They highlight the importance of model parameters, such as the “temperature” setting, in influencing the recommendations’ creativity and accuracy, indicating that generative AI, with appropriate tuning, can be a valuable tool for financial advisors.

Another development in the field is the Cogniwealth system, introduced by Ramyadevi and Sasidharan [232]. This platform utilizes the Llama 2 model as a financial advisor. The system leverages NLP and machine learning techniques to assist both professional fund researchers and laymen investors by providing personalized investment recommendations and financial insights. Cogniwealth’s ability to handle user-provided data and deliver human-like responses through an intuitive interface ensures high levels of adaptability, user-friendliness, and engagement.
 
Impact on Investment Strategies: LLMs are transforming the landscape of investment strategies, offering the potential to deliver more accurate, diverse, and accessible investment advice. A prime example of this is demonstrated in Ko and Lee’s study [233], where they showcase ChatGPT’s ability to construct portfolios with superior diversity and performance compared to randomly selected ones. This finding highlights the potential of LLMs to serve as valuable advisory tools for both professional portfolio managers and individual investors, democratizing access to advanced investment strategies.

LLMs can also impact the development of algorithmic trading strategies by automating the creation of accurate and executable code for technical indicators. The study conducted by Noguer i Alonso and Dupouy [234] compares the capabilities of various LLMs, such as GPT-4-Turbo, Gemini-Pro, Mistral, Llama 2, and Codellama, in generating code that runs correctly and matches baseline implementations. The study emphasizes the importance of well-designed prompts and the models’ ability to handle complex financial calculations for successful code generation.

Recently, Kim et al. [235] investigates the capability of an LLM, specifically GPT-4 Turbo, to perform financial statement analysis comparable to that of professional human analysts. By providing standardized and anonymous financial statements, the study examines the model’s ability to predict future earnings without any narrative or industry-specific context. The findings reveal that the LLM not only outperforms human analysts in predicting earnings changes, particularly in challenging scenarios, but also matches the performance of specialized state-of-the-art machine learning models. The authors claim that the model’s predictions derive not from its training memory but from generating useful narrative insights about a company’s future performance, thus eliminating look-ahead bias. To address this bias, the research design uses a consistent anonymized format for financial statements across firms, making it virtually impossible for the model to infer a firm’s identity. Additionally, the statements do not contain any dates and use relative years, mitigating concerns about the model leveraging macroeconomic trends from specific years. Furthermore, trading strategies based on the LLM’s predictions demonstrate higher Sharpe ratios and alphas compared to those based on other models.

Another promising application of LLMs in investment strategies is the analysis of annual reports to extract valuable insights, thereby enhancing stock investment strategies. Gupta [83] introduces a framework that utilizes GPT-3.5 to streamline the process of analyzing comprehensive 10-K filings of a company. By combining the generated insights with historical stock data, the study demonstrates that machine learning models trained on these LLM-generated features can outperform traditional market benchmarks, such as the S&P 500 index. This approach highlights the potential of integrating LLMs with historical data to improve the accuracy of stock predictions and enhance investment strategies.

Moreover, Zhang et al. [236] introduce BreakGPT for detecting financial breakouts. BreakGPT’s multi-stage structure improves the accuracy and stability of detecting true and false breakouts in financial markets by systematically analyzing price movements and order flows. The model’s superior performance compared to ChatGPT-3.5 and ChatGPT-4 makes it a valuable tool for traders and investors in detecting financial breakouts.

However, despite these promising developments, Chuang and Yang [237] raise an important concern regarding the implicit biases present in pre-trained language models, such as BERT and FinBERT. The study reveals that these models exhibit significant biases towards certain stocks and industry sectors, which can influence the quality and fairness of investment recommendations. They emphasize the need for awareness and mitigation of such biases in financial decision-making systems to ensure more reliable and fair investment advice. This research highlights the importance of careful model training and evaluation in financial contexts to develop robust and accountable financial advisory systems.
 
Regulatory and Ethical Considerations: The application of LLMs in financial advisory services has raised significant regulatory and ethical concerns. Caspi et al. [238] examine the regulatory landscape, highlighting key concerns such as maintaining fiduciary duties, ensuring transparency, and preventing conflicts of interest. They discuss potential regulatory strategies to address the challenges posed by generative AI, emphasizing the need for effective regulation that balances innovation with consumer protection. Moreover, Niszczota and Abbas [239] investigate the financial literacy of GPT models, revealing GPT-4’s near-perfect score on financial literacy tests. However, they also find that individuals with lower financial knowledge tend to rely more heavily on GPT’s advice.

Lakkaraju et al. [240] also compare the effectiveness and fairness of LLM-based chatbots (ChatGPT and Bard) to a rule-based chatbot (SafeFinance) in providing personal financial advice. They find that while ChatGPT and Bard generate fluent responses, they exhibit inconsistencies and biases across different user groups and languages. In contrast, SafeFinance provides reliable answers, albeit with limited generalization. The study demonstrates the need for improvements in LLM-based systems to ensure fairness and accuracy in financial advisement.

While LLMs have demonstrated potential in transforming financial advisory services, their application raises important regulatory and ethical considerations. Effective regulation should balance innovation with consumer protection, while educating users about the limitations and potential biases of AI-driven financial recommendations is essential to promote informed decision-making.

Operational risk management and compliance are critical components in the financial sector, as they help safeguard the integrity of financial institutions, protect consumers, and maintain the stability of the entire financial system. However, the increasing complexity of financial products, ever-changing regulations, and the constant threat of fraudulent activities pose significant challenges for financial institutions. LLMs are emerging as powerful tools that enhance these processes by providing sophisticated analytical capabilities. By leveraging LLMs, financial institutions can improve the accuracy of audits, streamline compliance verification, and detect inconsistencies more efficiently. This enables financial institutions to make informed decisions in critical areas such as Financial Auditing and Regulatory Compliance, and Fraud Detection and Risk Management, ultimately enhancing their operational resilience and ensuring compliance with regulatory requirements.
 
Financial Auditing and Regulatory Compliance: Financial auditing involves the systematic examination of financial records and statements to ensure accuracy and compliance with regulations. LLMs are increasingly being used to enhance these processes by improving the accuracy and efficiency of text matching and regulatory interpretation [241]. A study conducted by Hillebrand et al. [242] introduces ZeroShotALI, which stands for Zero-Shot Automated List Inspection. It combines GPT-4 and a domain-specific SentenceBERT model to enhance the matching of text segments from financial reports with specific legal requirements. This system significantly improves the efficiency and accuracy of financial audits compared to traditional methods.

Moreover, another study conducted by Cao and Feinstein [243] examines the use of LLMs (such as GPT-4, GPT-3.5, Claude-3-Opus, Gemini-1.5-Pro) for interpreting complex financial regulations, specifically focusing on Basel III capital requirements. Effective prompt design and document loading methods guide LLMs in translating regulatory texts into a concise mathematical framework, aiming to significantly enhance regulatory interpretation accuracy.

In addition, by analyzing firms’ public narrative disclosures with GPT-4, Choi and Kim [244] develop a novel measure of tax audit periods at the firm level. Their measure shows high conformity with data released by the IRS (the Internal Revenue Service) and reveals that tax audits lead to reduced tax avoidance, decreased capital investments, and increased stock volatility.

LLMs have shown potential in uncovering inconsistencies and contradictions in financial reports. A study conducted by Deußer et al. [245] develops an innovative approach to identifying discrepancies in financial reports by leveraging the power of LLMs such as GPT-4 and Llama. The study employs embedding-based paragraph clustering to efficiently detect contradictions across various datasets, including both annotated and unannotated financial reports. By utilizing sentence-pair data, document-level data, and intelligent bucketing systems, the researchers optimize the query process for the LLMs, enabling them to effectively pinpoint inconsistencies and contradictions. The results of this study demonstrate significant enhancements in the accuracy and efficiency of financial audits, ultimately reducing the time and effort required to conduct thorough and reliable financial report audits.
 
Fraud Detection and Risk Management: Fraud detection and risk management are critical components of maintaining financial integrity and stability. LLMs offer advanced capabilities to detect fraudulent activities and manage risks through sophisticated data analysis and pattern recognition. A study conducted by Feng et al. [246] highlights the potential of LLMs to revolutionize credit scoring and risk assessment. By instruction tuning, LLMs can match or surpass traditional credit scoring models, leading to more inclusive and comprehensive evaluations. However, the study also emphasizes the need to address biases within LLMs to ensure fair financial decision-making.

Furthermore, Cao et al. [212] present a novel framework named RiskLabs that leverages LLMs to predict financial risk by integrating data from various sources. By processing and fusing features from diverse data types, including textual and vocal information from Earnings Conference Calls (ECCs), market-related time series data, and contextual news data surrounding ECC release dates, RiskLabs outperforms traditional methods and existing models in forecasting financial risks, providing a more comprehensive understanding of market dynamics.

Several papers explore the application of LLMs in fraud detection. Zhao et al. [247] introduce an innovative GPT-based model for identifying fraudulent activities in payment systems, which excels in capturing detailed behavioral sequences through temporal and contextual analysis. Yang et al. [248] introduce the FinChain-BERT model, which enhances fraud detection accuracy by focusing on key financial terms and optimizing model performance. Similarly, Bhattacharya and Mickovic [249] demonstrate the effectiveness of the BERT model in detecting accounting fraud in financial reports by fine-tuning the model on Management Discussion and Analysis sections of annual 10-K reports from the Securities and Exchange Commission (SEC) database, outperforming existing benchmark models.

While LLMs have shown great potential in fraud detection and risk management, it is crucial to acknowledge and address the inherent biases that may exist within these models. Biases in LLMs can lead to unfair and discriminatory practices in financial decision-making. Ongoing research and development efforts are necessary to mitigate these biases and ensure the responsible and ethical deployment of LLMs in the financial sector.

Real-time reasoning enables instant and dynamic interactions between users and AI-powered systems. By leveraging the vast knowledge and understanding of LLMs, financial institutions can deploy chatbots, virtual assistants, and question-answering systems that provide accurate, relevant, and timely information to customers and stakeholders. These real-time applications streamline customer support, simplify complex financial transactions, and offer immediate access to financial insights and advice.
 
Chatbots and Virtual Assistants: Chatbots and virtual assistants are changing the way financial institutions interact with customers and streamline internal processes. By leveraging the capabilities of LLMs, these AI-driven tools can further provide more personalized and effective assistant, thereby enhancing customer satisfaction and boosting organizational efficiency. For instance, Aggarwal et al. [250] present a multi-purpose NLP chatbot that incorporates LLM models, including ChatGPT, BERT, and DistilBERT. The proposed system incorporates emotion recognition, multilingual support, and voice conversion. The chatbot demonstrates exceptional performance in providing personalized financial advice, understanding and responding to human emotions, and maintaining functionality in offline modes.

In another study, Yue and Au [251] introduce GPTQuant, a conversational AI chatbot designed to facilitate investment research. GPTQuant leverages few-shot learning and LangChain’s integration to generate Python code for backtesting and strategy analysis. The chatbot uses prompt templates to activate GPT-3’s capabilities, demonstrating efficacy in portfolio construction, rebalancing, and factor score querying.

Lastly, Yadav et al. [252] introduce a virtual assistant that utilizes LLMs to enhance the financial reconciliation process. The assistant automates the generation of SQL queries from natural language inputs, streamlining and expediting reconciliation, research, and validation processes for accountants. Utilizing a retrieve-and-refine strategy with retrieval-augmented generation (RAG) and few-shot prompting, the virtual assistant achieved 95% accuracy in generating correct SQL queries for real-world questions related to account reconciliation. This integration of LLMs significantly improves the accuracy and efficiency of generating SQL queries, demonstrating the potential of LLMs to automate repetitive and time-consuming tasks in financial reconciliation.
 
Question-Answering: Question-answering systems powered by LLMs have shown remarkable progress in understanding and responding to complex queries related to financial documents. Recent studies have focused on enhancing the numerical reasoning capabilities of these systems, enabling them to handle multi-step calculations and extract relevant information from various data sources. For example, Arun et al. [253] develop a pipeline utilizing fine-tuned LLMs, such as Llama-2-7B and T5, to analyze financial reports and answer numerical reasoning questions. By extracting and serializing tables from PDFs, generating embeddings, and training on the FinQA dataset, the authors demonstrated the potential for real-time analysis of financial reports. The study concludes that with appropriate fine-tuning and methodologies, LLMs could significantly enhance the efficiency and accuracy of financial data analysis, enabling swift and informed decision-making in dynamic market environments through rapid extraction and interpretation of crucial data points.

Furthermore, Phogat et al. [254] introduce zero-shot prompting techniques (ZS-FinPYT and ZS-FinDSL) for LLMs including GPT-3, GPT-3.5-turbo, and GPT-4 to perform complex numerical reasoning over financial documents. By encoding reasoning into Python/DSL(domain-specific languages) programs, these techniques mitigate arithmetic limitations. Evaluations on datasets such as FinQA, ConvFinQA, and TATQA demonstrate superior performance compared to baselines, particularly in table/text data, multi-step reasoning, and numerical questions.

In a related study, Srivastava et al. [255] investigate the mathematical reasoning capabilities of LLMs on financial documents. They introduce a novel prompting strategy, EEDP (Elicit-Extract-Decompose-Predict), designed to enhance LLM performance in scenarios requiring multi-step numerical reasoning. Extensive experimentation with multiple LLMs across financial datasets reveals that EEDP outperforms baseline strategies like Direct Prompting, Chain of Thought (CoT), and Program of Thoughts (PoT). The study highlights the potential of structured prompting strategies in improving LLM performance for complex reasoning tasks and identified common error types, emphasizing the need for precise information extraction.

Moreover, Xue et al. [103] propose a cutting-edge dialogue system designed specifically for the finance sector, named WeaverBird. It leverages a LLM with GPT architecture fine-tuned on extensive financial corpora. This enables WeaverBird to understand and provide informed responses to complex financial queries, such as investment strategies during inflation. The system’s performance is further enhanced by integrating a local knowledge base and search engine, allowing it to retrieve relevant information and generate responses conditioned on web search results, complete with proper source references for enhanced credibility. Comparative evaluations across a broad spectrum of financial question-answering tasks demonstrate WeaverBird’s superior performance compared to other models, positioning it as a powerful tool for financial dialogue and decision support.

The financial markets are dynamic and complex, requiring advanced tools to navigate effectively. LLMs have proven to be powerful allies in this domain by enabling the creation of intelligent trading agents that can process vast amounts of data and execute trades with high precision. These agents leverage LLMs’ NLP capabilities to interpret and synthesize financial news, market reports, and historical data, significantly improving market predictions and trading strategies. StockAgent [261], for instance, explores the potential of AI-driven trading systems to simulate and analyze stock market behaviors under various external influences. It is a multi-agent system powered by LLMs designed to mimic real investor behaviors and assess the impact of external factors like macroeconomic events, policy changes, and financial reports on trading activities. The study finds that different LLMs, such as GPT-3.5 Turbo and Gemini, exhibit distinct trading behaviors and preferences, with GPT agents showing more diverse and independent trading styles compared to the more homogeneous and trend-following behavior of Gemini agents. This variation suggests that LLM-based systems can offer personalized investment strategies and insights. The research also highlights that removing financial information or communication channels like BBS (Bulletin Board System) can significantly alter trading behaviors and market dynamics, underscoring the complexity and interdependence of factors influencing stock trading.

A notable advancement in LLM applications is integrating multimodal data—textual, numerical, and visual—into trading agents. FinAgent [262] exemplifies this by combining these data types to support quantitative and high-frequency trading including stocks and Crypto. Its diversified memory retrieval system and tool augmentation features enable FinAgent to interact with various data sources and tools, enhancing adaptability and performance in dynamic trading environments.

LLM-based trading agents excel in continuous learning and adaptation as well. FINMEM [263] introduces a layered memory and character design, enhancing the agent’s ability to process hierarchical financial data and convert insights into trading decisions. The memory module of FINMEM, inspired by human cognitive processes, includes working memory and layered long-term memory components. This design allows FINMEM to categorize and prioritize information based on its relevance and timeliness, retaining critical insights longer and enabling agile responses to new investment cues. Through real-world testing and continuous learning, FINMEM evolves its trading strategies, demonstrating improved decision-making and adaptability in volatile financial environments. Similarly, QuantAgent [264] focuses on self-improvement through a two-layer loop system. The inner loop refines responses using a knowledge base, while the outer loop involves real-world testing and knowledge enhancement. This iterative approach enables QuantAgent to autonomously extract financial signals and uncover viable trading opportunities, showcasing LLMs’ dynamic potential.

Integrating human expertise with AI capabilities is another significant advancement. The Alpha-GPT series, including Alpha-GPT [265] and Alpha-GPT 2.0 [266], emphasizes human-AI interaction in the alpha mining process. Alpha-GPT 2.0 further introduces a human-in-the-loop framework for iterative refinement of investment strategies. These agents interpret trading ideas and translate them into effective strategies, providing insightful and actionable alphas. By leveraging both human expertise and AI capabilities, this approach enhances the efficiency and creativity of the alpha mining process, leading to more effective investment decisions.

Simulating markets and economic activities has long been a critical aspect of financial research and policy analysis. Traditional simulators, typically grounded in econometric models and system dynamics, have been the cornerstone of this effort. These simulators rely on historical data and established economic theories to predict future market behaviors. For instance, models like the Vector Autoregression (VAR) model and the Dynamic Stochastic General Equilibrium (DSGE) model are widely used for economic forecasting and policy analysis [267], [268]. While they offer a structured and mathematically rigorous approach, traditional simulators often struggle with the complexity and dynamism inherent in real-world economic systems. They are generally static, assuming rational behavior and equilibrium, which can limit their accuracy and adaptability to unforeseen economic shocks or behavioral intricacies.

In contrast, agent-based simulators represent a significant advancement in the simulation of economic activities. These models consist of autonomous agents, each with distinct behaviors and decision-making processes. These agents interact within a defined environment, allowing for the emergence of complex macroeconomic phenomena from the bottom up. The primary advantage of ABM lies in their flexibility and ability to model heterogeneous agents with varying strategies and interactions. This approach can capture the non-linear dynamics of markets, such as feedback loops, market sentiments, and adaptive behaviors [269].

However, agent-based simulators are not without their challenges. One significant drawback is the computational complexity, as simulating numerous agents with intricate interactions demands substantial processing power. Additionally, the development of realistic agent behaviors and interaction rules requires deep domain expertise and can be time-consuming. Moreover, while agent-based simulators can model emergent phenomena, the validation of these models against real-world data remains a challenging task, often requiring extensive calibration and sensitivity analysis [270].

The integration of LLMs with agent-based simulators represents a cutting-edge development in economic simulations. With their advanced NLP capabilities, LLMs can enhance the perception, reflection, and decision-making processes of agents within simulators. This hybrid approach leverages the strengths of both technologies: the detailed and adaptive behaviors modeled by agent-based simulators and the comprehensive data processing and learning capabilities of LLMs.

Research by Li et al. [271] exemplifies the potential of this integration by demonstrating the ability to simulate complex macroeconomic activities. Their study, EconAgent, shows how LLM-empowered agents can realistically model economic activities by processing economic data through advanced mechanisms. These agents can simulate human-like decision-making processes, providing a comprehensive understanding of how different economic factors interact. This enables more accurate predictions of economic trends and the effects of policy changes. Equipped with layered memory systems, these agents can adapt their strategies based on real-time data inputs and historical analysis, making them highly effective for forecasting and policy simulation.

Similarly, Horton [272] explores the use of LLMs as computational models for economic simulations. By endowing LLMs with preferences and decision-making frameworks, their approach allows the simulation of human-like economic behavior. These simulations are particularly valuable for social science experiments and exploring economic scenarios, providing insights that can inform policy and strategy. The study introduces Homo Silicus agents, designed to emulate human economic agents by incorporating principles of behavioral economics. This enables the agents to make decisions based on a mix of rational analysis and emotional factors, providing a more realistic simulation of economic activities and market behaviors.

Furthermore, Zhao et al. [273] investigate the competitive behaviors of LLM-based agents in a simulated environment, demonstrating how competition among agents can lead to the emergence of innovative strategies and enhanced performance. They propose CompeteAI, a framework that simulates a virtual town where restaurant agents compete for customers, revealing how competition drives agents to continually adapt and improve their strategies, aligning with established sociological and economic theories.

The evolution from traditional simulators to agent-based models and now to LLM-empowered agents marks a significant stride in the field of economic simulation. The integration of LLMs with ABM offers a promising avenue for more realistic and adaptive modeling of economic activities, capturing the complex interplay of factors that drive markets and economies. This hybrid approach not only enhances our understanding of economic dynamics but also provides a powerful tool for forecasting and policy analysis.

The integration of LLMs into financial processes has reformed the way financial tasks are automated, offering enhanced capabilities for workflow generation and strategic planning. These applications streamline operations and provide robust solutions for complex financial tasks.

One notable application is FlowMind [274], which presents an innovative approach to automating financial workflows using LLMs. FlowMind leverages the capabilities of models like GPT to generate workflows on-the-fly, addressing the limitations of traditional robotic process automation that relies on predefined tasks. The system uses a structured lecture recipe to ground LLM reasoning with reliable APIs, mitigating issues such as hallucinations and ensuring data privacy by avoiding direct interaction with proprietary code. FlowMind includes a feedback loop that allows users to inspect high-level descriptions of the generated workflows and provide adjustments, enhancing the system’s adaptability. This approach is demonstrated using the NCEN-QA dataset, a benchmark for evaluating workflow generation in financial question-answering tasks, where FlowMind significantly outperforms traditional methods. This framework showcases the potential of LLMs to automate complex, spontaneous tasks in financial services while maintaining data integrity and security.

Another application is AUCARENA [275], which evaluates strategic planning and execution in auction environments to assess the strategic reasoning capabilities of LLM agents. In the ascending-price auctions, LLM agents like GPT-4 compete, managing budgets and adapting strategies in real-time. Utilizing a Belief-Desire-Intention model, agents update beliefs, adjust desires, and replan based on auction developments. This setup allows for a detailed analysis of how LLM agents manage resources, adhere to goals, and adapt to new information in competitive contexts. The study shows that LLM agents, especially GPT-4, are effective in strategic planning and resource management, though sometimes outperformed by simpler methods, highlighting areas for further improvement in LLM design. AUCARENA demonstrates the potential of LLMs to enhance decision-making processes in complex, competitive scenarios.

The use of multi-agent systems in financial analysis leverages the strengths of LLMs to enhance the robustness and accuracy of financial strategies. Multi-agent systems improve trading performance by simulating various agent interactions and providing a more comprehensive analysis of the tasks. TradingGPT [276] exemplifies this approach with its innovative multi-agent framework designed for financial trading. It organizes memory into three distinct layers: short-term, medium-term, and long-term, each governed by a custom decay mechanism that matches human cognitive processes. In TradingGPT, agents can engage in inter-agent communication and debate, enhancing their decision-making capabilities. Each agent is equipped with individualized trading characters, such as risk-seeking, risk-neutral, and risk-averse profiles, which enrich the diversity of perspectives and improve decision-making robustness. By leveraging layered memory processing and consistent information exchange, this framework demonstrates augmented adaptability to historical trades and real-time market cues, significantly enhancing automated trading outcomes. Aside from trading tasks, SocraPlan [277] leverages multi-agent reasoning with LLMs for effective corporate planning. This framework conducts comprehensive market research, customer profiling, product usage analysis, and sales strategy formulation. By combining human insights with AI capabilities, SocraPlan enhances corporate planning, enabling businesses to devise strategies that are both innovative and grounded in detailed market analysis. SocraPlan employs a multi-agent architecture where each agent specializes in a different aspect of corporate planning, such as competitive analysis, customer segmentation, or trend forecasting. These specialized agents collaborate to provide a holistic view of the market, which helps businesses make informed strategic decisions.

Multi-agent systems also benefit the in analyzing financial sentiments or textual information, which is a critical component of market analysis and strategy formulation as we have discussed in Section 3.1 and 3.2. An example is HAD [278], which indicates heterogeneous agent discussion, employing specialized agents focused on different types of errors common in FSA. This framework ensures that each of the agents focuses on particular errors, such as sarcasm, aspect mismatches, and temporal expressions, making the system robust against common pitfalls in sentiment analysis. The HAD framework has shown significant improvements in accuracy and F-1 scores across multiple datasets, proving its efficacy in refining sentiment analysis for financial texts. Another example is [279], which introduces a multi-agent framework that automates the verification of information between loan applications and bank statements, powered by both open-sourced models like Llama 3 and close-sourced models such as GPT-4. Despite higher operational costs, this approach is more economical and faster than manual reviews, offering a reliable solution for structured finance auditing and compliance.

Moreover, multi-agent systems can be used for monitoring and anomaly detection in financial markets. Park [220] introduces a sophisticated multi-agent framework designed to improve the validation and interpretation of financial data anomalies. The framework employs a network of specialized LLM agents, each focusing on distinct tasks such as data conversion, web-based expert analysis, utilization of institutional knowledge, cross-checking, and report consolidation. This collaborative approach enhances the efficiency and accuracy of anomaly detection, reducing the need for manual verification. By applying this framework to the S&P 500 index, the study demonstrates significant improvements in anomaly detection, showing that LLM-based agents can autonomously and accurately identify and interpret anomalies in financial market data, thereby supporting more effective financial market monitoring and decision-making

Besides the multi-agent systems, agents can interact with itself in an autonomous way as well [280]. The Self-Reflective LLM framework SEP [281], which means Summarize-Explain-Predict, addresses this need by enabling the generation of explainable stock predictions. SEP combines verbal self-reflective agents with Proximal Policy Optimization (PPO) to provide autonomous and explainable predictions. This framework allows agents to self-reflect on their decision-making processes, ensuring that the predictions are not only accurate but also interpretable. By enhancing the explainability of stock predictions, SEP improves accuracy, transparency, and trustworthiness among investors and analysts.
 
In summary, the integration of LLMs into agent-based modeling in finance offers significant advancements in trading, investment, financial analysis, and economic simulation. These applications demonstrate the versatility and effectiveness of LLMs in enhancing decision-making, strategy formulation, and market analysis. Future research in this area promises to further refine these systems, improving their accuracy, efficiency, trustworthiness, and adaptability in the ever-evolving financial landscape [282], [283], [284].