Large Language Models in equity markets: applications, techniques, and insights

In one sentence

Summary

This paper addresses the fragmentation in research regarding Large Language Models (LLMs) in equity investing by providing a comprehensive review of 84 studies published between 2022 and early 2025. The authors aim to map the intersection of LLMs and equity markets, addressing key questions about application trends, technical innovations, and existing limitations. The review positions LLMs as transformative tools that shift investment from manual, structured analysis to automated, real-time insights derived from multimodal data, including news, earnings transcripts, and social media. It emphasizes the potential for LLMs to enhance market responsiveness, risk management, and alpha generation through dynamic, self-learning systems powered by reinforcement learning and multi-agent frameworks. The methodology involves a dual-layered categorization of the selected literature. First, studies are grouped by financial applications, such as stock price forecasting, sentiment analysis, portfolio management, and algorithmic trading. Second, they are classified by technical methodologies, including prompting, fine-tuning, multi-agent frameworks, reinforcement learning, and custom architectures. The authors consolidate findings on datasets used, ranging from financial statements to multimodal data, and systematically compare general-purpose versus finance-specialized LLMs. The review evaluates empirical contributions and methodological innovations, analyzing how LLMs integrate qualitative textual data with structured financial indicators to improve predictive accuracy and generate actionable market signals. Key findings indicate that LLMs significantly improve sentiment extraction and can generate alpha when combined with reinforcement learning to factor market feedback. However, the review identifies critical gaps in scalability, interpretability, and real-world validation. Many studies rely on simulated environments or historical data, lacking robust evaluation frameworks for live trading. The authors highlight the importance of hybrid modeling approaches and architectures that leverage large context windows for complex reasoning. Limitations include data reliability, potential biases, and the challenge of adapting LLMs to diverse investment strategies and asset classes. The paper concludes by proposing future research directions, emphasizing the need for robust evaluation frameworks and transparent, efficient AI-driven financial strategies.

Research problem

The rapid proliferation of research on LLMs in equity investing has created a fragmented landscape with diverse focuses on multi-agent trading, time series forecasting, and domain-specific architectures. There is a lack of comprehensive synthesis that categorizes these efforts by both financial application and technical methodology. Additionally, critical gaps in scalability, interpretability, and real-world validation remain unaddressed, hindering the practical implementation of AI-driven financial strategies.

Core contributions

• Provides a comprehensive review of 84 research studies on LLMs in equity markets from 2022 to 2025.
• Introduces a dual-layered categorization framework: financial applications (e.g., forecasting, sentiment) and technical methodologies (e.g., prompting, multi-agent systems).
• Consolidates findings on datasets used, comparing general-purpose and finance-specialized LLMs.
• Identifies key research trends, commonalities, and divergences across studies.
• Highlights strengths such as improved sentiment extraction and reinforcement learning integration.
• Pinpoints critical gaps in scalability, interpretability, and real-world validation.
• Proposes future research directions including hybrid modeling and robust evaluation frameworks.

Method

• Systematic review of 84 research studies from 2022 to 2025.
• Dual-layered categorization: financial applications and technical methodologies.
• Consolidation of dataset findings and LLM type comparisons.
• Evaluation of empirical contributions and methodological innovations.
• Analysis of trends in sentiment extraction, reinforcement learning, and multi-agent systems.
• Identification of gaps in scalability, interpretability, and real-world validation.
• Proposal of future research directions including hybrid modeling and robust evaluation.

Results & findings

• LLMs enable rapid processing of vast datasets, combining structured metrics with unstructured sources like news and social media.
• Integration of LLMs facilitates evolution from static, rule-based models to dynamic, self-learning systems.
• Improved sentiment extraction and use of reinforcement learning to factor market feedback are key strengths.
• Critical gaps exist in scalability, interpretability, and real-world validation of LLM-driven strategies.
• Hybrid modeling approaches and architectures with large context windows are proposed for future research.
• Robust evaluation frameworks are needed to advance AI-driven financial strategies.
• The review maps the intersection of LLMs and equity markets to provide a foundation for future research.
• Diverse investment strategies and asset classes present challenges for LLM adaptation.

Data & experiments

• Review of 84 research studies sourced from Google Scholar and arXiv between 2022 and early 2025.
• Analysis of datasets ranging from financial statements to multimodal data (news, earnings transcripts, social media).
• Comparison of general-purpose vs. finance-specialized LLMs used in the reviewed studies.
• Evaluation of empirical contributions and methodological innovations across the literature.
• Assessment of studies focusing on stock price forecasting, sentiment analysis, portfolio management, and algorithmic trading.
• Review of technical approaches including prompting, fine-tuning, multi-agent frameworks, and reinforcement learning.

Limitations & caveats

• Relies on a review of existing studies, which may have their own limitations in data and methodology.
• Many reviewed studies lack real-world validation, relying on simulated or historical data.
• Gaps in scalability and interpretability of LLM-driven strategies are identified.
• Challenges in adapting LLMs to diverse investment strategies and asset classes remain.
• Data reliability and potential biases in LLM outputs are noted as critical issues.
• The review focuses on equity markets, potentially limiting generalizability to other financial domains.

Curator notes

Discovery status: round2_promoted. Source links: AlphaFin: Benchmarking Financial Analysis with Retrieval-Augmented Stock-Chain Framework; FinLlama: Financial Sentiment Classification for Algorithmic Trading Applications; Revolutionizing Finance with LLMs: An Overview of Applications and Insights