Bridging Language Models and Financial Analysis: A Survey of Datasets, Models, and Applications

In one sentence

Summary

The paper addresses the interdisciplinary gap between computer science’s focus on predictive scalability and finance’s emphasis on causal inference and validation. It aims to synthesize recent LLM developments to guide their practical integration into the financial sector, where cautious adoption and long-term validation are prioritized. The survey highlights that while LLMs offer transformative potential for processing multifaceted financial data, their implementation remains slower than technological innovation due to these domain-specific requirements. By mapping the current landscape, the work seeks to provide a roadmap for researchers and practitioners to leverage emerging techniques effectively in high-stakes financial environments. The methodology involves a comprehensive review of textual data tasks, including classification, information extraction, summarization, and question answering, alongside an analysis of model architectures from BERT to decoder-based LLMs. The authors categorize applications into trading (sentiment analysis, time-series forecasting, fundamental analysis), risk modeling, ESG scoring, M&A forecasting, and financial agents. They examine specific systems like LLMFactor, StockGPT, and FinRobot, evaluating their use of Chain-of-Thought, Retrieval-Augmented Generation, and multi-agent simulations. The survey also covers experimental economics using LLM agents, noting their ability to replicate human subject experiments with comparable results in behavioral finance contexts. Key findings indicate that LLMs significantly enhance fundamental analysis by handling large contexts and enabling deductive reasoning, surpassing traditional models in interpreting complex narratives. The survey identifies promising future directions, including advanced prompting techniques, RAG for knowledge-based inference, tool-augmented models, and multimodal capabilities. However, it notes that research on LLMs in ESG and M&A remains in early stages, with challenges in scoring consistency and high-stakes validation. The paper concludes that while LLMs show immense potential, their full impact requires systematic application of these innovations to achieve greater accuracy and adaptability in complex financial modeling and prediction tasks.

Research problem

The paper addresses the significant gap between the rapid advancement of Large Language Models (LLMs) and their slower, cautious adoption in the financial industry. It highlights the interdisciplinary disconnect where computer science prioritizes predictive accuracy and scalability, while finance emphasizes statistical inference, causal relationships, and long-term validation. The core problem is how to effectively bridge this divide to unlock the transformative potential of LLMs for processing complex, multifaceted financial data, including text, tables, and charts, which traditional methods struggle to analyze efficiently.

Core contributions

• Provides a comprehensive survey of LLM applications in finance, covering datasets, models, and specific use cases like trading, risk, and ESG.
• Identifies and categorizes key financial tasks such as text classification, information extraction, summarization, and question answering.
• Analyzes the evolution from traditional statistical models to modern LLMs, highlighting their superior performance in handling unstructured financial text.
• Explores advanced LLM techniques including Chain-of-Thought, Retrieval-Augmented Generation, and multi-agent systems in financial contexts.
• Outlines future research directions for integrating cutting-edge LLM innovations into financial reasoning, decision-making, and analysis.
• Discusses the use of LLM agents in experimental economics, demonstrating their viability as alternatives to human subject studies.

Method

• Conducts a systematic literature review of LLM research in finance, focusing on recent advancements.
• Categorizes financial tasks into text classification, information extraction, summarization, and question answering.
• Analyzes the application of various LLM architectures, including encoder-based and decoder-based models.
• Evaluates the use of advanced prompting techniques like Chain-of-Thought and Tree-of-Thoughts in financial reasoning.
• Examines the integration of Retrieval-Augmented Generation (RAG) to improve accuracy and reduce hallucinations.
• Reviews multi-agent systems and simulations for trading and experimental economics.
• Assesses the potential of multimodal LLMs for integrating textual, numerical, and visual financial data.
• Identifies gaps in current research and proposes future directions for LLM applications in finance.

Results & findings

• LLMs demonstrate superior performance in processing complex, unstructured financial text compared to traditional machine learning models.
• Chain-of-Thought and Retrieval-Augmented Generation significantly improve LLM reasoning and accuracy in financial tasks.
• LLM-based agents can replicate human subject experiments in behavioral economics with comparable results.
• Multimodal LLMs effectively integrate textual and numerical data for enhanced stock forecasting and portfolio management.
• LLMs enable sophisticated fundamental analysis by handling large contexts and performing deductive reasoning on financial narratives.
• Current LLM applications in ESG and M&A forecasting are promising but remain in early stages with consistency challenges.
• Multi-agent systems show potential for enhancing trading accuracy through debate and collaborative prediction mechanisms.

Data & experiments

• Reviews datasets for financial text classification, information extraction, summarization, and question answering.
• Analyzes models trained on financial corpora, including 10-K filings, earnings call transcripts, news articles, and social media.
• Examines experimental setups for sentiment analysis, stock price prediction, and risk modeling using LLMs.
• Discusses benchmarks for financial question answering and the evaluation of LLM reasoning capabilities.
• Covers simulations of financial markets and trading behaviors using multi-agent LLM systems.
• Notes the lack of standardized benchmarks for emerging areas like ESG scoring and M&A forecasting with LLMs.

Limitations & caveats

• The survey is descriptive and does not present new empirical results or benchmarks.
• Research on LLMs in ESG and M&A forecasting is noted as being in early stages with limited validation.
• Challenges in scoring consistency and high-stakes validation for LLMs in finance remain unresolved.
• The rapid pace of LLM development means some cited works may become outdated quickly.
• The gap between predictive accuracy and causal inference in finance is not fully bridged by current LLM methods.
• Practical adoption in finance is hindered by cautious integration requirements and regulatory constraints.

Curator notes

Useful for synthesis across data, models, and applications.