Alpaca 15m regular session · monthly walk-forward · generated 2026-06-05
Google Intraday Walk-Forward Review
这版用了月度 walk-forward：每个月只看过去 252 天，选择当时训练分数最高且分数大于 0 的一个候选，执行下个月。它比固定参数更防过拟合，但结果也更保守：全样本 Sharpe 0.936，2026 Sharpe 2.514，2026-03 后 Sharpe 4.030。
Full Sharpe
0.936
CAGR 7.17%
2026 Sharpe
2.514
Mar+ 4.030
PAW Net Sharpe
0.909
P0 FAIL
MaxDD
-9.01%
Active 3.99%
P&L Curve
Rolling Sharpe
Strategy Eval
- PAW decision: FAIL / 先不要推进。
- Bias: shift causal, autocorr warning。CAUSAL — lead-cheat boosts Sharpe more than lag-stale drops it
- PAW OOS Sharpe: 4.673; IS Sharpe: 0.550。
- Fill assumptions、portfolio interaction、paper trading stream 仍是 needs_input。
完整 PAW artifacts: /workspace/research/2026-06-05_google_rebound/experiments/EXP-google-rebound-intraday-wf-fast-v1-20260605/best_candidate/paw_eval_report
Recent Monthly Selection
- 2026-01: leadlag_QQQ_lb6_q0.05_reversal_none
- 2026-02: pair_NVDA_bw520_zw260_e2.4_google_only
- 2026-03: leadlag_QQQ_lb6_q0.05_reversal_none
- 2026-04: leadlag_QQQ_lb6_q0.05_reversal_none
- 2026-05: leadlag_QQQ_lb6_q0.05_reversal_none
- 2026-06: leadlag_QQQ_lb6_q0.05_reversal_none
Stress
| TC bps | Delay bars | Full | 2026 | Mar+ OOS | MaxDD |
|---|---|---|---|---|---|
| 1.5 | 1 | 0.936 | 2.514 | 4.030 | -9.01% |
| 1.5 | 2 | 0.266 | 0.620 | -0.089 | -12.99% |
| 1.5 | 3 | -0.641 | -0.003 | -1.582 | -20.34% |
| 3.0 | 1 | 0.490 | 2.057 | 3.540 | -9.89% |
| 3.0 | 2 | -0.191 | 0.013 | -0.820 | -14.80% |
| 3.0 | 3 | -1.118 | -0.661 | -2.316 | -22.17% |
| 5.0 | 1 | -0.110 | 1.437 | 2.872 | -11.81% |
| 5.0 | 2 | -0.801 | -0.801 | -1.795 | -18.31% |
| 5.0 | 3 | -1.749 | -1.539 | -3.287 | -28.77% |
| 10.0 | 1 | -1.612 | -0.141 | 1.145 | -28.83% |
| 10.0 | 2 | -2.307 | -2.819 | -4.169 | -35.45% |
| 10.0 | 3 | -3.294 | -3.686 | -5.620 | -45.64% |
Top Walk-Forward Policies
| Policy | Full | Pre-2025 | 2025 | 2026 | Mar+ OOS | MaxDD |
|---|---|---|---|---|---|---|
| fast_wf_lb252_top1_floor0.0 | 0.936 | -0.099 | 0.963 | 2.514 | 4.030 | -9.01% |
| fast_wf_lb252_top1_floor0.5 | 0.936 | -0.099 | 0.963 | 2.514 | 4.030 | -9.01% |
| fast_wf_lb252_top1_floor-0.5 | 0.936 | -0.099 | 0.963 | 2.514 | 4.030 | -9.01% |
| fast_wf_lb252_top2_floor0.5 | 0.633 | 0.150 | 0.098 | 3.057 | 3.608 | -7.56% |
| fast_wf_lb252_top1_floor1.0 | 0.560 | -0.099 | 0.200 | 2.514 | 4.030 | -10.65% |
| fast_wf_lb252_top2_floor1.0 | 0.522 | 0.150 | -0.037 | 3.057 | 3.608 | -9.70% |
| fast_wf_lb365_top1_floor-0.5 | 0.496 | -0.099 | -0.206 | 3.660 | 3.545 | -9.87% |
| fast_wf_lb365_top1_floor0.0 | 0.496 | -0.099 | -0.206 | 3.660 | 3.545 | -9.87% |
| fast_wf_lb252_top2_floor-0.5 | 0.446 | 0.150 | -0.229 | 3.049 | 3.608 | -8.71% |
| fast_wf_lb252_top2_floor0.0 | 0.446 | 0.150 | -0.229 | 3.049 | 3.608 | -8.71% |
| fast_wf_lb365_top3_floor0.0 | 0.247 | -0.255 | -0.493 | 4.147 | 4.298 | -10.32% |
| fast_wf_lb365_top3_floor-0.5 | 0.247 | -0.255 | -0.493 | 4.147 | 4.298 | -10.32% |
Decision
还不能上 paper。这套结果显示方向越来越接近：最近 regime 的 rolling Sharpe 很高，但全样本和 PAW gate 不够。下一步应该做 regime classifier：只在 mega-cap tech rebound / QQQ mean-reversion regime 打开策略，其他时期空仓；同时要补 factor returns 和 paper fill simulator。