# Alpha-R1: Alpha Screening with LLM Reasoning via Reinforcement Learning

<p align="center">
  <img src="assets/Alpha-R1.png" alt="Alpha-R1" style="width: 100%; height: auto;">
</p>

<p align="center">
  <a href="https://huggingface.co/FinStep/Alpha-R1/blob/main/README_en.md">English</a> | <a href="https://huggingface.co/FinStep/Alpha-R1/blob/main/README.md">中文</a>
</p>

**Alpha-R1** is a reasoning-enhanced LLM for quantitative alpha selection: built on Qwen3-8B and trained with GRPO reinforcement learning ([verl](https://github.com/volcengine/verl)) using a market-feedback reward. It reasons over **semantic factor descriptions** — how each factor works, when it works, and when it fails — and selects the Alpha101 factors that best fit current market conditions.

- 📄 Paper: [arXiv:2512.23515](https://arxiv.org/abs/2512.23515)
- 💻 Code: [FinStep-AI/Alpha-R1](https://github.com/FinStep-AI/Alpha-R1) (inference pipeline / qlib backtesting / training config)
- 📜 License: MIT

## Model Overview

<p align="center">
  <img src="assets/framework.png" alt="Alpha-R1 framework overview" style="width: 100%;">
</p>

| Item | Content |
|---|---|
| Base model | [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) |
| Training | GRPO (verl) with a market-feedback reward |
| Input | Decision-context prompt: concatenated semantic factor descriptions `α_des` |
| Output | Selected factors listed in `<alpha_list>` |
| Candidate pool | 82 Alpha101 factors (as screened in the paper) |
| Recommended decoding | temperature=0 (greedy), top_p=0.7, max_new_tokens=4096 |

## Quick Start

### transformers

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "FinStep/Alpha-R1"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype="bfloat16", device_map="auto")

prompt = "<decision context: concatenated factor descriptions>"  # see the GitHub repo for the prompt builder
inputs = tokenizer.apply_chat_template(
    [{"role": "user", "content": prompt}],
    add_generation_prompt=True, return_tensors="pt",
).to(model.device)

# paper setting: temperature=0 (greedy), top_p=0.7
out = model.generate(inputs, max_new_tokens=4096, do_sample=False)
print(tokenizer.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))
```

### vLLM

```python
from vllm import LLM, SamplingParams

llm = LLM(model="FinStep/Alpha-R1")
params = SamplingParams(temperature=0.0, top_p=0.7, max_tokens=4096)
outputs = llm.chat([[{"role": "user", "content": prompt}]], params)
```

For the full end-to-end pipeline (factor description generation → Alpha-R1 inference → output parsing → qlib strategy backtest), see the [GitHub repository](https://github.com/FinStep-AI/Alpha-R1).

## Output Contract

The model lists the selected factor ids inside `<alpha_list>...</alpha_list>`, e.g.:

```
<alpha_list>alpha001, alpha021, alpha053</alpha_list>
```

Validation and parsing scripts are provided under `src/alpha_r1/parsing/` in the GitHub repository.

## Performance

12-month out-of-sample testing (2025-01-01 to 2025-12-31, paper Table 1):

<p align="center">
  <img src="assets/main_results.png" alt="Backtest NAV comparison on S&P 500 (left) and CSI 300 (right)" style="width: 100%;">
</p>

<table>
  <thead>
    <tr>
      <th rowspan="2">Type</th>
      <th rowspan="2" width="160">Method</th>
      <th colspan="3">S&amp;P 500</th>
      <th colspan="3">CSI 300</th>
    </tr>
    <tr>
      <th>AR (%)</th>
      <th>SR</th>
      <th>MDD (%)</th>
      <th>AR (%)</th>
      <th>SR</th>
      <th>MDD (%)</th>
    </tr>
  </thead>
  <tbody>
    <tr><td rowspan="9">Non-LLM</td><td>Buy&nbsp;&amp;&nbsp;Hold</td><td>19.34</td><td>0.80</td><td>18.75</td><td>22.16</td><td>1.31</td><td>10.49</td></tr>
    <tr><td>PCA</td><td>7.98</td><td>0.27</td><td>17.30</td><td>2.93</td><td>0.17</td><td>14.46</td></tr>
    <tr><td>XGBoost</td><td>3.49</td><td>0.03</td><td>18.45</td><td>8.99</td><td>0.50</td><td>16.26</td></tr>
    <tr><td>LightGBM</td><td>-5.42</td><td>-0.43</td><td>20.93</td><td>18.44</td><td>1.05</td><td>14.92</td></tr>
    <tr><td>A2C</td><td>10.82</td><td>0.40</td><td>17.70</td><td>22.96</td><td>1.20</td><td>14.86</td></tr>
    <tr><td>PPO</td><td>7.68</td><td>0.25</td><td>14.97</td><td>14.96</td><td>0.81</td><td>12.95</td></tr>
    <tr><td>DDPG</td><td>2.53</td><td>-0.02</td><td>15.04</td><td>1.97</td><td>0.12</td><td>16.54</td></tr>
    <tr><td>TD3</td><td>5.54</td><td>0.14</td><td>16.58</td><td>8.66</td><td>0.52</td><td>10.26</td></tr>
    <tr><td>SAC</td><td>37.60</td><td>1.44</td><td>15.18</td><td>9.77</td><td>0.56</td><td>11.68</td></tr>
    <tr><td rowspan="5">LLM</td><td>Gemini&nbsp;2.5&nbsp;Pro</td><td>14.23</td><td>0.55</td><td>17.01</td><td>16.29</td><td>0.90</td><td>14.01</td></tr>
    <tr><td>Claude&nbsp;3.7&nbsp;Sonnet</td><td>10.92</td><td>0.40</td><td>18.88</td><td>10.13</td><td>0.57</td><td>14.49</td></tr>
    <tr><td>DeepSeek&#8209;R1</td><td>21.94</td><td>0.93</td><td><b>14.36</b></td><td>14.66</td><td>0.81</td><td>14.60</td></tr>
    <tr><td>Qwen3&#8209;8B</td><td>12.85</td><td>0.47</td><td>19.52</td><td>15.44</td><td>0.79</td><td>14.38</td></tr>
    <tr><td><b>Alpha&#8209;R1&nbsp;(Ours)</b></td><td><b>47.87</b></td><td><b>1.62</b></td><td>16.91</td><td><b>40.57</b></td><td><b>2.23</b></td><td><b>6.58</b></td></tr>
  </tbody>
</table>

Out-of-domain generalization without retraining (paper Table 2): 80.54% AR (SR 2.46) on Russell 2000 and 73.52% AR (SR 2.80) on CSI 1000. AR = annualized return, SR = excess Sharpe ratio, MDD = max drawdown.

## Training

Alpha-R1 is trained on Qwen3-8B with GRPO using verl and a market-feedback reward (`R_final = R_adjusted - P_structural`, paper Section 3.4). The training configuration and a reference reward implementation live in the `training/` directory of the GitHub repository.

## Limitations

- This model is intended for academic research; its outputs do not constitute investment advice.
- Factor selection depends on the upstream description-generation and backtesting pipeline (see the GitHub repository); the model alone does not produce tradable signals.

## Citation

```bibtex
@article{jiang2025alphar1,
  title={Alpha-R1: Alpha Screening with LLM Reasoning via Reinforcement Learning},
  author={Jiang, Zuoyou and Zhao, Li and Sun, Rui and Sun, Ruohan and Li, Zhongjian and Li, Jing and Jiang, Daxin and Bai, Zuo and Hua, Cheng},
  journal={arXiv preprint arXiv:2512.23515},
  year={2025}
}
```

## License

This project is released under the [MIT License](https://opensource.org/licenses/MIT).