Multi-Turn Tool-Using LLM Agents Enable Adaptive Alpha Mining
ORIGINAL / AlphaPilot: Multi-Turn Tool-Using LLM Agents for Adaptive Alpha Mining with Reinforcement Learning
Existing formulaic alpha mining methods typically exploit only a subset of signals among market text, historical data, and quantitative evaluation. AlphaPilot integrates market descriptions, LLM financial knowledge, real-data evaluator feedback, and a retained factor pool within a multi-turn reinforcement learning framework, and performs day-by-day search at inference to adapt to shifting market regimes, thereby jointly leveraging multiple signal types for alpha mining.
01 ABSTRACT
The paper introduces AlphaPilot, a reinforcement learning framework that trains multi-turn tool-using LLM agents for adaptive formulaic alpha mining. The authors report that across three independent out-of-sample rolling backtests on CSI300 and CSI500, mean total returns were 42.93% and 54.19%, with mean Sharpe ratios of 2.46 and 2.86, and that it attained the highest mean total return, Sharpe, and RankIC among the evaluated methods on both universes. These figures are results reported by the authors under specific backtest settings.
02 KEY FINDINGS
- Across multiple turns, the agent proposes factor expressions conditioned on market descriptions, queries real-data evaluator feedback, and updates a retained factor pool.
- At inference time, it performs day-by-day search and pool updates to adapt to shifting market regimes.
- Across three independent out-of-sample rolling backtests on CSI300 and CSI500, reported mean total returns were 42.93% and 54.19%, with mean Sharpe ratios of 2.46 and 2.86.
AI GENERATED SUMMARY / DISCOVERED BY ARXIV CS.CL · ARXIV CS.LG