# A New DAPO Algorithm for Stock Trading

**Authors:** Ruijian Zha, Bojun Liu  
**arXiv:** [2505.06408](https://arxiv.org/abs/2505.06408)  
**Date:** 2025-05-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2505.06408v1/

## Key points

- Ruijian Zha and Bojun Liu from Columbia University explore reinforcement learning techniques in financial trading
- Focus on Dynamic Sampling Policy Optimization (DAPO) algorithm combined with Large Language Models (LLMs)
- Introduce a novel trading agent integrating Group Relative Policy Optimization (GRPO) with insights from DAPO
- Agent incorporates LLM-based signals related to risk and sentiment from financial news sources
- Performance evaluated on NASDAQ-100 index using FNSPID dataset
- Improved DAPO algorithm achieves cumulative return of 230.49% and Information Ratio of 0.37, outperforming CPPO-DeepSeek baseline
- Training time reduced from 8 hours to 2.5 hours over 100 epochs, along with decreased RAM usage
- Research demonstrates potential for data-efficient trading agents by combining reinforcement learning algorithms like DAPO with LLM-based signals from financial news sources

## Summary

In their paper titled "A New DAPO Algorithm for Stock Trading," Ruijian Zha and Bojun Liu from the Department of Computer Science at Columbia University explore the application of reinforcement learning techniques in financial trading. They focus on the Dynamic Sampling Policy Optimization (DAPO) algorithm, which has shown promising results when combined with Large Language Models (LLMs). The study introduces a novel trading agent that integrates an enhanced Group Relative Policy Optimization (GRPO) algorithm with insights from DAPO. Additionally, the agent incorporates LLM-based signals related to risk and sentiment extracted from financial news sources. The performance of this approach is evaluated on the NASDAQ-100 index using the FNSPID dataset. Results indicate that the improved DAPO algorithm achieves a cumulative return of 230.49% and an Information Ratio of 0.37, surpassing the CPPO-DeepSeek baseline. Furthermore, the proposed method significantly reduces training time from approximately 8 hours to 2.5 hours over 100 epochs while also decreasing RAM usage. Overall, this research showcases a scalable path towards developing data-efficient trading agents by leveraging reinforcement learning algorithms like DAPO in conjunction with LLM-based signals derived from financial news sources. The findings highlight the potential for advancements in automated trading strategies through innovative combinations of machine learning techniques and real-world financial data analysis.