Quantum Alpha: OpenReasoning-Nemotron-7B GRPO
Fine-tuned for quantum computing sector signal generation using GRPO (Group Relative Policy Optimization) with actual market returns as the reward signal.
Performance (out-of-sample, 421 evaluation articles)
| Horizon | IC | p-value | Direction Accuracy |
|---|---|---|---|
| 1 day | +0.151 | 0.003 | - |
| 5 days | +0.157 | 0.006 | 58.6% |
| 10 days | +0.160 | 0.008 | - |
| 20 days | +0.159 | 0.024 | - |
Training
- Base: nvidia/OpenReasoning-Nemotron-7B
- Stage 1: SFT on 881 quantum computing articles (V4 training data)
- Stage 2: GRPO with reward function based on actual 5-day cumulative abnormal returns
- 184 training articles with return data, 4 generations per prompt
- Proper train/test split (no data leakage)
- Downloads last month
- 160