[Paper Review] Automate Strategy Finding with LLM in Quant Investment

Zhizhuo Kou, Holam Yu+8 more|arXiv (Cornell University)|Sep 10, 2024

Stock Market Forecasting MethodsDecision Sciences3 citations

TL;DR

This paper proposes a novel LLM-powered multi-agent framework for automated alpha factor discovery and dynamic portfolio optimization in quantitative investing. By leveraging LLMs to mine diverse alphas from multimodal financial data and using a dynamic weight-gating mechanism to adaptively combine trading agents based on real-time market conditions, the framework achieves superior performance and stability over state-of-the-art baselines on Chinese stock markets.

ABSTRACT

We present a novel three-stage framework leveraging Large Language Models (LLMs) within a risk-aware multi-agent system for automate strategy finding in quantitative finance. Our approach addresses the brittleness of traditional deep learning models in financial applications by: employing prompt-engineered LLMs to generate executable alpha factor candidates across diverse financial data, implementing multimodal agent-based evaluation that filters factors based on market status, predictive quality while maintaining category balance, and deploying dynamic weight optimization that adapts to market conditions. Experimental results demonstrate the robust performance of the strategy in Chinese & US market regimes compared to established benchmarks. Our work extends LLMs capabilities to quantitative trading, providing a scalable architecture for financial signal extraction and portfolio construction. The overall framework significantly outperforms all benchmarks with 53.17% cumulative return on SSE50 (Jan 2023 to Jan 2024), demonstrating superior risk-adjusted performance and downside protection on the market.

Motivation & Objective

• Address the instability and high uncertainty of existing deep learning models in quantitative trading by introducing a more adaptive and interpretable framework.
• Overcome the rigidity of traditional rule-based alpha mining methods by enabling continuous, data-driven discovery of diverse alpha factors using LLMs.
• Improve strategy robustness across varying market conditions through a multi-agent architecture that evaluates and selects optimal alpha factors in real time.
• Enhance portfolio performance by dynamically weighting trading agents based on current market regimes using a context-aware gating mechanism.
• Demonstrate the framework’s efficacy through extensive backtesting on Chinese stock market data, outperforming SOTA models and benchmarks.

Proposed method

• Utilize Large Language Models (LLMs) to extract and categorize potential alpha factors from multimodal data sources, including numerical market data, financial research papers, and visual charts.
• Construct a diverse pool of trading agents with varying risk preferences through ensemble learning, enabling comprehensive analysis across different market regimes.
• Implement a dynamic weight-gating mechanism that assigns real-time weights to the most relevant agents based on current market conditions, ensuring context-aware strategy composition.
• Integrate LLM-generated alpha signals with historical backtesting across multiple market environments to validate and refine predictive performance.
• Apply a multi-agent architecture to evaluate the relevance and effectiveness of each alpha factor under different market scenarios, enhancing robustness and adaptability.
• Aggregate weighted alpha strategies into a composite investment formula that maximizes risk-adjusted returns through continuous adaptation to market dynamics.

Figure 1. Pipeline of Quant Trading Company

Experimental results

Research questions

• RQ1 Can LLMs effectively mine and generate diverse, interpretable alpha factors from multimodal financial data, including textual, numerical, and visual inputs?
• RQ2 How does a multi-agent system with heterogeneous risk preferences improve the robustness and adaptability of quantitative trading strategies across dynamic market conditions?
• RQ3 To what extent does a dynamic weight-gating mechanism enhance strategy performance by selecting and weighting the most relevant agents based on real-time market regimes?
• RQ4 How does the proposed framework outperform state-of-the-art models and traditional human-driven strategies in terms of risk-adjusted returns and stability on real-world equity data?
• RQ5 Can the integration of LLMs and multi-agent systems lead to a more scalable and maintainable approach to automated alpha discovery in quantitative investing?

Key findings

• The proposed framework significantly outperforms state-of-the-art baselines across multiple financial metrics, including Sharpe ratio, cumulative returns, and maximum drawdown, on Chinese stock market data.
• The integration of LLM-generated alphas with a multi-agent architecture leads to a more stable and adaptive strategy, reducing performance volatility across different market regimes.
• The dynamic weight-gating mechanism effectively identifies and prioritizes the most relevant trading agents based on real-time market conditions, improving overall strategy performance.
• The framework demonstrates superior performance compared to human traders and passive market indices, validating its potential for practical deployment in quantitative investment.
• The LLM-based seed alpha generation process enables incremental updates as new research and data emerge, ensuring long-term relevance and adaptability.
• Extensive backtesting confirms that the multi-agent evaluation system enhances the robustness of alpha factor selection by assessing performance across diverse market scenarios.

Figure 2. (A) An example of the seed alpha. (B) Its equivalent expression tree. (C) Step-by-step computation of this seed alpha on an example time series.