FinRL-DAPO-GRPO-Sentiment-Risk

This repository contains the code and trained model for our FinRL Contest 2025 Task 1 submission:
"A New DAPO Algorithm for Stock Trading".

We propose a novel trading agent that integrates a Group Relative Policy Optimization (GRPO) approach with:

• Insights from the DAPO algorithm (used in LLM preference tuning),
• Sentiment and risk signals derived from LLM-extracted financial news,
• An exponentiated sentiment-risk reward function for more robust decision-making.

Model Highlights

• Framework: Custom RL agent using GRPO with decoupled clipping.
• Reward: Weighted sentiment-risk adjusted return.
• Dataset: FNSPID, based on Nasdaq-100 from 1999–2023.
• Performance:
  • Cumulative return: 230.49% (2020–2023)
  • Info Ratio: 0.37
  • Max Drawdown: -49.11%
  • Outperforms CPPO-DeepSeek 10% baseline

Files

• model_rl.pth: Trained PyTorch model checkpoint.
• README.md: This file.
• (Optionally add: training logs, config files, etc.)

Usage

This model is designed for research use in financial RL.
You can load the model via PyTorch:

import torch

model = torch.load("model_rl.pth")
model.eval()