Automate Strategy Finding with LLM in Quant investment

In one sentence

Summary

The paper addresses the rigidity and brittleness of traditional deep learning models in quantitative finance by proposing an automated strategy finding framework. It leverages Large Language Models (LLMs) within a risk-aware multi-agent system to generate executable alpha factor candidates from diverse financial data, aiming to overcome the limitations of static factor models in dynamic market environments. The core innovation lies in combining LLM-based exploratory alpha mining with multi-agent evaluation and dynamic weight optimization to create a scalable architecture for financial signal extraction and portfolio construction. The methodology consists of three stages: First, an LLM-based Seed Alpha Factory (SAF) generates and categorizes 100 seed alphas across nine domains (e.g., Momentum, Fundamental) from multimodal financial research documents. Second, a multi-agent system comprising a Confidence Score Agent (CSA) and a Risk Preference Agent (RPA) evaluates these alphas using historical backtesting metrics like Information Coefficient (IC) and Sharpe Ratio, selecting factors based on market status and category balance. Third, a 3-layer Multi-Layer Perceptron (MLP) optimizes the weights of selected alphas to predict future yields. Experiments were conducted on SSE50, CSI300, and SP500 indices using OHLCV data, financial reports, and macroeconomic indicators from 2019 to 2024. The framework significantly outperformed benchmarks, achieving a 53.17% cumulative return on SSE50 (Jan 2023-Jan 2024) compared to the index's -13.22% loss, with superior risk-adjusted metrics (Sharpe: 0.287, Calmar: 1.052). It demonstrated robustness across Chinese and US markets, maintaining positive returns during downturns where benchmarks declined. However, limitations include reliance on input document quality, potential lack of financial intuition in LLM-generated alphas, and the assumption of persistent historical relationships between market conditions and alpha performance, which may fail during regime shifts. The system is primarily equity-focused, with cross-asset applicability requiring further investigation.

Research problem

Traditional quantitative trading methods suffer from rigidity, lacking adaptability to dynamic market conditions, and face challenges in integrating diverse data sources effectively. The paper addresses the brittleness of deep learning models in finance by seeking to automate the discovery and optimization of alpha factors using LLMs and multi-agent systems, ensuring strategies remain robust across varying market regimes.

Core contributions

• A novel LLM-driven framework for identifying formulaic alpha factors, creating a dynamic 'Alpha Factory' from multimodal information.
• Introduction of a multi-agent evaluation system (Confidence Score and Risk Preference Agents) for rigorous, scenario-specific alpha selection.
• Integration of dynamic weight optimization via MLP to adapt alpha factor weights to current market conditions.
• Demonstration of superior performance across Chinese (SSE50, CSI300) and US (SP500) markets compared to traditional ML and LLM baselines.

Method

• LLM-based filtering and categorization of financial research to generate seed alphas.
• Multi-agent system with Confidence Score Agent (CSA) and Risk Preference Agent (RPA).
• Backtesting alphas using historical data to calculate IC and Sharpe Ratio.
• Category-based alpha selection algorithm ensuring diversification across factor types.
• 3-layer MLP for dynamic weight optimization of selected alphas.
• Top-k/drop-n portfolio construction with equal weighting and turnover constraints.

Results & findings

• Achieved 53.17% cumulative return on SSE50 (Jan 2023-Jan 2024) vs. -13.22% for the index.
• Outperformed FinCon (22.47%) and SEP (17.89%) among LLM-based methods.
• Superior risk-adjusted performance: Sharpe 0.287, Sortino 0.208, Calmar 1.052.
• Maintained positive returns in CSI300 (12.78%) and SP500 (2.77%) during H1 2022 downturns.
• Ablation study showed removing the Confidence Score Agent reduced out-of-sample IC by 31.9%.
• LLM-generated alphas showed synergistic effects, with individual low-IC factors contributing to high combined performance.

Data & experiments

• Datasets: SSE50, CSI300, and SP500 constituents with OHLCV data, financial reports, and macroeconomic indicators (2019-2024).
• Benchmarks: XGBoost, LightGBM, MLP, PPO_filter, FinCon, SEP, and the respective index benchmarks.
• Metrics: Information Coefficient (IC), Sharpe Ratio, Sortino Ratio, Calmar Ratio, Volatility, and Cumulative Return.
• Experiments: Backtesting on specific time windows, ablation studies on agent components, and sensitivity analysis on hyperparameters.

Limitations & caveats

• Reliance on input document quality may perpetuate biases in generated alphas.
• LLM-generated alphas may lack financial intuition, leading to theoretically sound but practically infeasible factors.
• Assumes persistent historical relationships between market conditions and alpha performance.
• Validation primarily focused on equity markets; cross-asset applicability is unproven.
• Computational efficiency and scalability of the multi-agent system are not fully addressed.

Curator notes

Discovery status: round2_promoted. Source links: A Multimodal Foundation Agent for Financial Trading: Tool-Augmented, Diversified, and Generalist; InvestLM: A Large Language Model for Investment using Financial Domain Instruction Tuning; LiveTradeBench: Seeking Real-World Alpha with Large Language Models; Revolutionizing Finance with LLMs: An Overview of Applications and Insights