A Review of Large Language Models for Stock Price Forecasting from a Hedge-Fund Perspective

In one sentence

Summary

This paper addresses the gap between academic LLM research and practical hedge-fund deployment for stock price forecasting. It positions LLMs as versatile tools for sentiment extraction, financial report analysis, time-series tokenization, and multi-agent trading systems. Unlike prior surveys, it emphasizes the 'hedge-fund lens,' focusing on the robustness, frictions, and regulatory constraints that determine real-world viability rather than just algorithmic novelty. The authors argue that current literature often overstates LLM capabilities by ignoring market microstructure and evaluation biases. The review categorizes LLM usage into five areas: sentiment analysis from news/social media, factual extraction from 10-K filings and earnings calls, relationship mapping via graph construction, tokenization of price series for next-token prediction, and multi-agent decision systems (e.g., FinArena, AlphaAgents). It details methods such as prompt engineering, RAG, and agentic frameworks like LangChain. Experiments cited include StockGPT’s 119% annual return and various sentiment-based baselines. The authors critically assess these results, noting that many studies use short horizons, inadequate baselines, and metrics like MSE that do not reflect trading profitability. Key findings highlight significant pitfalls in current literature. Sentiment signals are fragile due to regime dependence and reflexivity. Dataset designs often suffer from data leakage (e.g., cross-day event drift) and short evaluation windows that miss full market cycles. The paper stresses that ignoring illiquidity premiums and transaction costs can turn profitable backtests into losing strategies, especially for small-cap stocks. It concludes with practical guidelines: use long-horizon datasets, align metrics with risk-adjusted returns (Sharpe, drawdown), control leakage rigorously, and benchmark against non-LLM technical analysis strategies to ensure genuine alpha generation.

Research problem

The paper addresses the disconnect between academic LLM research for stock forecasting and the practical requirements of hedge-fund trading pipelines. It seeks to identify why LLM-based strategies often fail in production despite promising academic results. The core problem is the lack of rigorous evaluation regarding market frictions, data leakage, and regime shifts in existing literature.

Core contributions

• Synthesizes LLM applications in stock forecasting from a hedge-fund perspective, covering sentiment, reports, tokenization, and agents.
• Critiques empirical practices in current literature, highlighting divergences between academic benchmarks and realistic deployment conditions.
• Identifies specific pitfalls: sentiment fragility, data leakage, short-horizon bias, and ignored illiquidity premiums.
• Provides practical guidelines for integrating LLMs into production pipelines, including rigorous leakage control and metric alignment.
• Discusses the limits of stock price predictability and the necessity of regime-switch detection in LLM frameworks.

Method

• Literature review of LLM applications in stock price forecasting and trading systems.
• Critical analysis of empirical practices, focusing on dataset design, evaluation metrics, and leakage controls.
• Assessment of sentiment analysis fragility, including regime dependence and reflexivity effects.
• Evaluation of tokenization methods for time-series data, noting sensitivity to binning and model architecture.
• Review of multi-agent architectures (e.g., FinArena, AlphaAgents) and their aggregation mechanisms.
• Discussion of market frictions, including illiquidity premiums and transaction costs, on strategy performance.

Results & findings

• Sentiment signals are fragile and context-dependent, often failing to generalize across market regimes.
• StockGPT reported 119% annual return, but this ignores illiquidity and transaction costs which can erase profits.
• Data leakage is prevalent, with cross-day event drift and literal label statements contaminating training data.
• Most studies use short evaluation windows, leading to overestimation of strategy robustness.
• Illiquidity premiums of 10-20 bps daily can significantly erode cumulative returns over a year.
• Sentiment from professional outlets may be more predictive than social media, but this requires rigorous validation.
• Multi-agent systems reduce hallucinations but require careful design to avoid overfitting and ensure transparency.

Data & experiments

• Reviews studies using financial news, social media (Reddit, Twitter), 10-K filings, and earnings-call transcripts.
• Analyzes datasets like BigData, CIKM18, and ACL18, noting their short horizons (1-2 years) are insufficient for cycle testing.
• Cites experiments from StockGPT, FinArena, and AlphaAgents, evaluating their reported returns and methodologies.
• Discusses the need for long-horizon, regime-spanning datasets that include both bull and bear phases.
• Highlights the lack of standardized benchmarks that account for transaction costs and slippage.

Limitations & caveats

• As a review, it does not propose new models but synthesizes existing ones.
• Critiques academic literature, which may have selection bias in cited studies.
• Does not provide new empirical results but relies on cited works' findings.
• Illiquidity estimates are based on specific market conditions (Russell 2000) and may vary.
• Sentiment analysis critique assumes regime dependence, which may not hold in all market conditions.
• Recommendations for long-horizon datasets are practical but difficult to implement due to data availability.

Curator notes

Discovery status: round3_promoted. Source links: Enhancing Investment Analysis: Optimizing AI-Agent Collaboration in Financial Research; FinArena: A Human-Agent Collaboration Framework for Financial Market Analysis and Forecasting; GPT-InvestAR: Enhancing Stock Investment Strategies through Annual Report Analysis with Large Language Models; Large Language Models for Financial and Investment Management: Applications and Benchmarks; Large Language Models in equity markets: applications, techniques, and insights; Temporal Data Meets LLM - Explain