Window: 2021-01-30 .. 2026-08-16 (ETH/BTC intersect) IS/OOS split: 2025-01-01 | features: 36 (bar-only intersection, 29 base cols) | timestep: 1Day | timesteps trained: 200000 | seed: 0 action: continuous Box[1] in [-1, +1] (target position fraction, scaled by max_position_pct=0.5) PPO: lr=0.0003 n_steps=2048 gamma=0.99 ent_coef=0.01
Scoreboard
| Variant | Days | Ret | Sharpe | PF | MaxDD |
|---|---|---|---|---|---|
| baseline gate60 (OOS, ETH) | 593.0 | +51.38% | 0.98 | 1.41 | -18.16% |
| RL PPO (OOS, ETH+BTC) | 571.0 | -21.13% | 0.04 | 1.01 | -61.83% |
Significance on OOS
- OOS mean daily return: +0.0067% (RL) vs +0.0829% (gate60)
- bootstrap p(RL adds nothing): 0.714
- NULL-R permutation p (2000 circular shifts): 1.000
Verdict
- OOS Sharpe > 0.5: False (0.04)
- OOS Sharpe > baseline: False (0.04 vs 0.98)
- SOL cross-check: SKIPPED (model obs space is 1444 dims, would need 724; honest cross-check requires per-symbol training)
- Verdict: REJECT
Stage 1 (regime switcher) joined the parked list with the same verdict pattern (positive point estimate, no statistical significance). If Stage 2 also REJECTs, the self-learning-bot plan should stop here and only resume with a new mechanism hypothesis. See docs/ROADMAP.md:699-718 for the full parked list.
elapsed: 2151.0s | seed: 0 | timesteps: 200000