ALERA: An Entropy-Based LLM Multi-Agent Framework for Dynamic Risk Quantification in Quantitative Investing

Abstract

The financial services industry increasingly relies on data-driven quantitative investing, with large language models (LLMs) advancing decision-making through improved financial semantic understanding, news-driven forecasting, and trading signal generation. However, traditional risk metrics like volatility, Value-at-Risk, and Maximum Drawdown often fail to capture the dynamic complexity of markets, leading to undetected risks or excessive trading. We propose ALERA, an LLM-based multi-agent framework that uses entropy to quantify decision-making uncertainty and assess risk across four dimensions: expert disagreements, historical experience, market volatility, and stock correlations. ALERA mitigates long-term information forgetting with a memory pool and incorporates stock interactions for robust risk assessment. Our contributions include pioneering entropy-based risk quantification, a unified framework for dynamic trading strategies, and extensive empirical validation showing superior performance in return and Sharpe ratio compared to baseline methods. Specifically, ALERA achieves an annualized return of 26.32% and a Sharpe Ratio of 1.71 for AAPL, with a cumulative return of 6.31% and a maximum drawdown of 3.63% over a three-month period from August to November 2024, outperforming baselines like MACD, KDJ+RSI, ZMR, and SMA.