Integrating Large Language Models in Financial Investments and Market Analysis: A Survey

In one sentence

Summary

This survey addresses the integration of Large Language Models into financial decision-making, categorizing existing research into four primary frameworks: LLM-based pipelines, hybrid integration methods, fine-tuning approaches, and agent-based architectures. It positions LLMs as tools that enhance traditional quantitative models by processing unstructured data like news and filings, thereby improving real-time analytical capabilities and interpretability in complex financial environments. The paper systematically reviews techniques such as Retrieval-Augmented Generation (RAG), Chain-of-Thought reasoning, and Parameter-Efficient Fine-Tuning (PEFT) to contextualize how these methods mitigate knowledge cutoffs and improve domain-specific reasoning. The authors analyze numerous empirical studies, including MarketSenseAI, Ploutos, and LLMoE, which utilize datasets ranging from S&P 500 indices to proprietary financial news and SEC filings. Experimental designs often involve comparing LLM-generated signals against traditional baselines using metrics like cumulative returns, Sharpe ratios, and prediction accuracy. For instance, MarketSenseAI achieved up to 72% cumulative returns on S&P 100 stocks, while LLMoE demonstrated over 25% improvement in return metrics for Microsoft and Apple. The survey also examines fine-tuning efforts using LoRA on models like Llama-2 and GPT-3.5, and agent systems like FinCon that employ multi-agent collaboration for stock analysis and trading simulations. Key findings indicate that hybrid models and multi-agent systems generally yield superior performance in stock selection, risk assessment, and portfolio optimization compared to single-model approaches. However, the survey highlights significant limitations, including the susceptibility of LLMs to hallucinations, the high computational cost of fine-tuning, and the lack of standardized benchmarks for financial LLMs. It concludes that future research must focus on adaptive fine-tuning, specialized financial architectures, and robust human-AI collaboration frameworks to ensure reliability and regulatory compliance in live trading environments.

Research problem

The paper addresses the challenge of integrating Large Language Models into financial investment strategies to enhance decision-making beyond traditional quantitative methods. It seeks to categorize and evaluate how LLMs process unstructured data (news, filings) alongside numerical data to improve stock prediction, risk assessment, and portfolio management. The survey aims to identify effective frameworks, techniques, and limitations in current LLM applications within the financial domain.

Core contributions

• Categorizes LLM integration in finance into four distinct frameworks: pipelines, hybrid methods, fine-tuning, and agent-based architectures.
• Reviews key technical enablers such as RAG, Chain-of-Thought, and PEFT/LoRA for financial applications.
• Synthesizes empirical results from diverse studies, highlighting performance gains in stock selection and portfolio optimization.
• Identifies critical limitations including hallucination risks, data scarcity, and the need for standardized benchmarks.
• Proposes future research directions focusing on adaptive learning, specialized architectures, and human-AI collaboration.

Method

• Systematic literature review of LLM applications in financial decision-making.
• Categorization of research into four main frameworks: pipelines, hybrid, fine-tuning, and agents.
• Analysis of technical components: RAG for data retrieval, CoT for reasoning, PEFT for efficient adaptation.
• Evaluation of empirical results from cited studies, focusing on returns, accuracy, and risk metrics.
• Comparison of model performance across different asset classes and market conditions.
• Assessment of limitations and challenges such as hallucination, latency, and regulatory compliance.

Results & findings

• MarketSenseAI achieved up to 72% cumulative returns and 10-30% excess alpha on S&P 100 stocks.
• LLMoE showed over 25% improvement in return metrics for MSFT and AAPL using an Mixture-of-Experts approach.
• GPT-InvestAR outperformed the S&P 500 index by using LLM-generated features from 10-K filings.
• MuSA framework achieved higher Sharpe ratios and annualized returns on DJIA and S&P 500 using DRL and sentiment analysis.
• Fine-tuned GPT-3.5 provided clearer, more logically consistent answers for investment queries than baseline models.
• Hybrid models combining LLMs with traditional indicators (MAs, VIX) improved stock trend prediction accuracy.
• Self-reflective agents in SEP framework outperformed baselines in both predictive accuracy and explanation quality.

Data & experiments

• Reviews studies using S&P 500, S&P 100, KOSPI200, and proprietary datasets like Benzinga news and SEC filings.
• Evaluates models on tasks including stock movement prediction, sentiment analysis, and portfolio weight optimization.
• Compares LLM-based strategies against traditional quantitative baselines and other AI models using return and accuracy metrics.
• Analyzes the impact of different prompting techniques (few-shot, CoT) and fine-tuning methods (LoRA) on model performance.
• Examines multi-agent systems in simulated trading environments like FinArena and TwinMarket.

Limitations & caveats

• LLMs are prone to hallucinations, which can lead to inaccurate financial insights.
• High computational costs associated with fine-tuning and running large models.
• Lack of standardized benchmarks for evaluating financial LLMs.
• Difficulty in adapting to rapidly changing market conditions with static models.
• Regulatory and compliance challenges in deploying LLMs for financial decision-making.
• Dependence on the quality and availability of unstructured financial data.

Curator notes

Discovery status: round3_promoted. Source links: Automate Strategy Finding with LLM in Quant investment; Can Large Language Models beat wall street? Evaluating GPT-4’s impact on financial decision-making with MarketSenseAI; Enhancing Investment Analysis: Optimizing AI-Agent Collaboration in Financial Research; FinArena: A Human-Agent Collaboration Framework for Financial Market Analysis and Forecasting; FinCon: A Synthesized LLM Multi-Agent System with Conceptual Verbal Reinforcement for Enhanced Financial Decision Making; GPT-InvestAR: Enhanc