A Review of Large Language Models for Stock Price Forecasting from a Hedge-Fund Perspective
A comprehensive review examines how large language models are being applied to stock price forecasting in quantitative finance, with particular emphasis on practical challenges often overlooked in academic literature. The analysis, framed from a hedge-fund perspective, addresses critical implementation issues including sentiment analysis fragility, data leakage risks, and market friction constraints that affect real-world trading performance.
Large language models have emerged as powerful tools for financial forecasting, enabling sophisticated extraction of signals from unstructured data sources like earnings calls, financial news, and social media sentiment. This arXiv review synthesizes recent LLM applications across multiple domains—from sentiment classification to multi-agent trading system design—offering rare insight into both opportunities and implementation hazards that practitioners face when deploying these systems at scale.
The critical contribution of this analysis lies in its unflinching examination of practical pitfalls. Academic papers frequently showcase impressive model architectures and backtested returns while downplaying issues like sentiment analysis brittleness, temporal data leakage that artificially inflates performance metrics, and the fundamental limits of stock price predictability. Horizon design—the choice of prediction timeframe—dramatically affects model viability, yet receives minimal attention in most literature. Similarly, illiquidity premia and transaction costs systematically erode theoretical edge in live trading environments.
For hedge funds and institutional traders, this review provides essential stress-testing guidance. LLM-based systems demonstrate vulnerability to market regime changes and novel financial events outside their training distributions. The multi-agent framework discussions highlight coordination challenges and the risk of correlated failures across seemingly independent models. Risk managers must account for LLM hallucinations, citation errors in financial documents, and sentiment misclassifications that cascade through trading pipelines.
Looking forward, the field requires more rigorous validation methodologies that incorporate realistic market microstructure, proper walk-forward testing protocols, and honest performance reporting. Future research should prioritize robustness across market conditions rather than pursuing marginal accuracy improvements on historical datasets. Practitioners implementing LLM-driven strategies must maintain skepticism about backtested results and implement conservative position sizing until live performance validates assumptions.
- →LLMs enable extraction of valuable signals from financial news, transcripts, and social media but suffer from significant fragility in sentiment classification under real-world conditions
- →Data leakage and improper horizon design commonly inflate backtested returns, making academic benchmarks unreliable for evaluating hedge-fund implementation feasibility
- →Illiquidity premia and transaction costs substantially erode LLM-derived trading edges in live markets despite strong historical backtests
- →Multi-agent trading systems face coordination risks and correlated failure modes that traditional risk models fail to capture
- →Practical deployment requires stress-testing across market regimes and conservative position sizing until live performance validates academic assumptions