Alpaca 15m regular session · long sample · generated 2026-06-05
Google Intraday 长样本复核
结论：60 天 GOOGL/MSFT residual 的高 Sharpe 不能跨样本成立。把样本拉到 2024-01-02 到 2026-06-04 后，原候选全样本 Sharpe 变成 -0.354；长样本最佳候选也只有全样本 Sharpe 1.368。2026 片段仍然很强，说明这是近期 Google/mega-cap regime 的线索，不是已经足够扎实的 paper 策略。
Long Best
1.368
leadlag_QQQ_lb6_q0.05_reversal_none
Best 2026
3.660
OOS Mar+ 3.545
Old 60d Candidate
-0.354
2026 only 2.112
PAW Net Sharpe
1.386
P0 FAIL / daily adapter
最佳长样本策略在干什么
leadlag_QQQ_lb6_q0.05_reversal_none：看 QQQ 过去 6 根 15m bar 的极端波动。如果 QQQ 短线大涨/大跌进入最极端 5% 区间，就反向交易 GOOGL，赌 GOOGL 对大盘科技冲击出现短线回归。它只交易 GOOGL，不需要配对 MSFT，执行比原 residual pair 简单。
PAW / Bias
- PAW daily adapter: Net Sharpe 1.386, CAGR 11.64%, MaxDD -7.73%。
- Bias shift test: CAUSAL — lead-cheat boosts Sharpe more than lag-stale drops it。
- Autocorr warning: WARN — mean AC 0.000 < 0.5; positions jumpy, shift test less informative。这是因为日内策略的日末仓位很跳，PAW daily adapter 对它不是完美工具。
- 实质防未来函数逻辑仍在 intraday engine：信号用 completed bar，收益用下一根 bar 的 shifted position。
Top Candidates
| Candidate | Family | Full | Pre-2025 | 2025 | 2026 | Mar+ OOS | MaxDD | Active |
|---|---|---|---|---|---|---|---|---|
| leadlag_QQQ_lb6_q0.05_reversal_none | lead_lag | 1.368 | 0.328 | 1.440 | 3.660 | 3.545 | -7.87% | 5.26% |
| leadlag_CRM_lb1_q0.05_momentum_googl_up | lead_lag | 0.954 | 0.520 | -0.034 | 4.128 | 3.449 | -6.47% | 2.93% |
| leadlag_CRM_lb3_q0.1_momentum_googl_up | lead_lag | 0.745 | 0.262 | 0.254 | 2.546 | 2.862 | -5.16% | 5.74% |
| pair_META_bw520_zw520_e2.4_google_only | residual_pair | 0.909 | 0.313 | 0.787 | 2.347 | 2.321 | -6.20% | 3.25% |
| pair_AMZN_bw520_zw520_e1.8_google_only | residual_pair | 0.722 | 0.730 | -0.119 | 2.756 | 3.302 | -9.41% | 5.68% |
| pair_META_bw520_zw260_e2.4_google_only | residual_pair | 0.618 | -0.095 | 0.499 | 2.232 | 3.128 | -7.06% | 3.71% |
| pair_AMZN_bw260_zw520_e1.8_google_only | residual_pair | 0.772 | 0.468 | 0.469 | 2.094 | 2.454 | -8.79% | 5.75% |
| leadlag_META_lb6_q0.05_reversal_none | lead_lag | 0.793 | -0.013 | 0.486 | 3.033 | 2.507 | -4.86% | 5.32% |
| pair_NVDA_bw520_zw520_e1.8_google_only | residual_pair | 0.669 | 0.670 | 0.444 | 1.323 | 2.565 | -7.55% | 5.91% |
| pair_AMZN_bw260_zw260_e1.8_google_only | residual_pair | 0.696 | 0.233 | 0.320 | 2.521 | 2.497 | -8.34% | 6.31% |
| leadlag_CRM_lb1_q0.05_momentum_none | lead_lag | 0.678 | 0.505 | -0.242 | 3.768 | 3.050 | -9.60% | 5.02% |
| leadlag_XLK_lb6_q0.05_reversal_none | lead_lag | 0.378 | -0.186 | -0.441 | 4.283 | 3.480 | -12.27% | 5.51% |
| leadlag_CRM_lb1_q0.05_momentum_qqq_up | lead_lag | 0.542 | 0.334 | -1.026 | 4.182 | 3.947 | -9.80% | 2.78% |
| pair_NVDA_bw260_zw520_e2.4_google_only | residual_pair | 0.529 | 0.089 | 0.172 | 2.127 | 2.706 | -8.48% | 3.47% |
| pair_AMZN_bw1040_zw260_e1.8_google_only | residual_pair | 0.552 | 0.406 | -0.318 | 3.143 | 3.081 | -10.85% | 6.32% |
Stress
| Candidate | TC bps | Delay | Full | 2026 | Mar+ OOS | MaxDD |
|---|---|---|---|---|---|---|
| leadlag_QQQ_lb6_q0.05_reversal_none | 1.5 | 1 | 1.368 | 3.660 | 3.545 | -7.87% |
| leadlag_QQQ_lb6_q0.05_reversal_none | 1.5 | 2 | 0.672 | 2.844 | -0.195 | -10.71% |
| leadlag_QQQ_lb6_q0.05_reversal_none | 1.5 | 3 | 0.379 | 1.586 | -1.427 | -6.93% |
| leadlag_QQQ_lb6_q0.05_reversal_none | 3.0 | 1 | 0.928 | 3.156 | 2.960 | -9.03% |
| leadlag_QQQ_lb6_q0.05_reversal_none | 3.0 | 2 | 0.207 | 2.267 | -0.917 | -14.39% |
| leadlag_QQQ_lb6_q0.05_reversal_none | 3.0 | 3 | -0.147 | 0.961 | -2.152 | -9.42% |
| leadlag_QQQ_lb6_q0.05_reversal_none | 5.0 | 1 | 0.336 | 2.473 | 2.161 | -12.72% |
| leadlag_QQQ_lb6_q0.05_reversal_none | 5.0 | 2 | -0.414 | 1.487 | -1.879 | -19.38% |
| leadlag_QQQ_lb6_q0.05_reversal_none | 5.0 | 3 | -0.847 | 0.120 | -3.112 | -16.50% |
| leadlag_QQQ_lb6_q0.05_reversal_none | 10.0 | 1 | -1.159 | 0.722 | 0.112 | -23.76% |
| leadlag_QQQ_lb6_q0.05_reversal_none | 10.0 | 2 | -1.954 | -0.479 | -4.222 | -32.48% |
| leadlag_QQQ_lb6_q0.05_reversal_none | 10.0 | 3 | -2.566 | -1.977 | -5.422 | -36.64% |
| pair_MSFT_bw520_zw260_e1.8_paired | 1.5 | 1 | -0.354 | 2.112 | 3.506 | -14.10% |
| pair_MSFT_bw520_zw260_e1.8_paired | 1.5 | 2 | -1.816 | -1.667 | 0.914 | -20.96% |
| pair_MSFT_bw520_zw260_e1.8_paired | 1.5 | 3 | -1.523 | -0.491 | -0.159 | -17.55% |
| pair_MSFT_bw520_zw260_e1.8_paired | 3.0 | 1 | -1.465 | 1.172 | 2.598 | -24.26% |
| pair_MSFT_bw520_zw260_e1.8_paired | 3.0 | 2 | -3.178 | -2.817 | -0.310 | -31.84% |
| pair_MSFT_bw520_zw260_e1.8_paired | 3.0 | 3 | -2.961 | -1.709 | -1.550 | -28.92% |
| pair_MSFT_bw520_zw260_e1.8_paired | 5.0 | 1 | -2.953 | -0.104 | 1.356 | -36.65% |
| pair_MSFT_bw520_zw260_e1.8_paired | 5.0 | 2 | -4.949 | -4.329 | -1.928 | -44.22% |
| pair_MSFT_bw520_zw260_e1.8_paired | 5.0 | 3 | -4.832 | -3.330 | -3.398 | -41.79% |
| pair_MSFT_bw520_zw260_e1.8_paired | 10.0 | 1 | -6.578 | -3.343 | -1.844 | -60.56% |
| pair_MSFT_bw520_zw260_e1.8_paired | 10.0 | 2 | -8.950 | -7.890 | -5.753 | -66.21% |
| pair_MSFT_bw520_zw260_e1.8_paired | 10.0 | 3 | -9.047 | -7.217 | -7.756 | -64.74% |
Decision
不能上 paper。更准确的下一步是继续 research：保留 QQQ→GOOGL reversal 作为 regime-aware 线索，加入 walk-forward 参数选择、成交价模型、factor returns、以及更长历史/更多市场状态；原 GOOGL/MSFT 60 天 residual 候选应降级为近期 regime hit，不作为 paper 候选。