From Feedback Loops to Policy Updates: Reinforcement Fine-Tuning for LLM-Based Alpha Factor Discovery
The paper introduces extsc{QuantEvolver}, a framework designed for alpha factor discovery in quantitative trading. It addresses limitations of existing LLM-based methods by converting feedback into policy updates, thus improving the efficiency of factor generation. Extensive experiments show that extsc{QuantEvolver} outperforms traditional methods in producing high-quality and diverse factor pools.
- ▪Modern quantitative trading relies on systematic models for alpha factor discovery.
- ▪Existing LLM-based methods often suffer from context explosion and feedback drift.
- ▪extsc{QuantEvolver} uses reinforcement fine-tuning to optimize factor generation.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Computational Engineering, Finance, and Science arXiv:2605.15412 (cs) [Submitted on 14 May 2026] Title:From Feedback Loops to Policy Updates: Reinforcement Fine-Tuning for LLM-Based Alpha Factor Discovery Authors:Lingzhe Zhang, Tong Jia, Yunpeng Zhai, Zixuan Xie, Chiming Duan, Minghua He, Philip S. Yu, Ying Li View a PDF of the paper titled From Feedback Loops to Policy Updates: Reinforcement Fine-Tuning for LLM-Based Alpha Factor Discovery, by Lingzhe Zhang and 7 other authors View PDF HTML (experimental) Abstract:Modern quantitative trading increasingly relies on systematic models to extract predictive signals from large-scale financial data, where alpha factor discovery plays a central role in transforming market observations into tradable signals.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
