Alpaca 15m regular session · stat arb comparison · generated 2026-06-05
Google Stat Arb Basket Comparison
你问得对：可以把它写得更像统计套利。所以我对比了单因子 price residual、return residual、等权篮子 spread，以及当前 QQQ tail reversal。结果是：残差策略形式更“stat arb”，但目前没有超过现有 QQQ tail reversal。
Best Overall
Baseline
baseline_lead_QQQ_lb6_q0.05_reversal_none_h1
Baseline Recent 1Y
2.211
Prior 0.926
Best Residual Recent 1Y
1.920
Prior -0.312
Residual OOS Mar+
4.464
Full 0.338
Compared Candidates
Best residual: price_resid_META_bw260_zw520_e2.5_paired_h2. 这是 GOOGL 对 META 的滚动 log-price residual，paired 执行，残差偏高时 short GOOGL / long META beta hedge，残差偏低时反过来。
P&L Curve
Rolling Sharpe
Top Candidates
| Candidate | Family | Full | Recent 1Y | Prior | Mar+ OOS | MaxDD |
|---|---|---|---|---|---|---|
| baseline_lead_QQQ_lb6_q0.05_reversal_none_h1 | lead_lag_tail_reversal | 1.368 | 2.211 | 0.926 | 3.545 | -7.87% |
| price_resid_META_bw260_zw520_e2.5_paired_h2 | price_residual | 0.338 | 1.920 | -0.312 | 4.464 | -9.81% |
| price_resid_META_bw260_zw520_e2.5_google_only_h2 | price_residual | 0.988 | 1.690 | 0.548 | 2.912 | -7.74% |
| return_resid_MSFT_lb6_bw260_zw520_e2.4_google_only_h1 | return_residual | 0.885 | 2.005 | 0.082 | 2.515 | -5.69% |
| return_resid_MSFT_lb6_bw260_zw260_e2.4_google_only_h1 | return_residual | 1.169 | 1.504 | 0.939 | 2.623 | -7.54% |
| price_resid_META_bw260_zw520_e2.5_paired_h1 | price_residual | 0.726 | 1.354 | 0.232 | 3.379 | -5.96% |
| price_resid_MSFT_bw260_zw520_e2.0_paired_h2 | price_residual | 0.973 | 1.251 | 0.788 | 2.911 | -5.03% |
| price_resid_META_bw260_zw260_e1.5_paired_h1 | price_residual | 0.623 | 2.139 | -0.305 | 1.667 | -12.28% |
| price_resid_MSFT_bw260_zw520_e2.0_google_only_h2 | price_residual | 0.771 | 1.864 | 0.206 | 1.830 | -15.10% |
| price_resid_META_bw260_zw260_e1.5_paired_h2 | price_residual | 0.530 | 2.128 | -0.482 | 1.599 | -13.06% |
| price_resid_META_bw260_zw520_e2.5_google_only_h1 | price_residual | 0.752 | 1.399 | 0.321 | 2.644 | -8.47% |
| price_resid_MSFT_bw260_zw260_e2.0_google_only_h2 | price_residual | 0.740 | 2.398 | -0.355 | 1.183 | -21.97% |
| return_resid_META_lb6_bw260_zw260_e1.8_google_only_h1 | return_residual | 0.782 | 1.866 | 0.029 | 1.699 | -10.31% |
| return_resid_META_lb3_bw520_zw260_e2.4_google_only_h1 | return_residual | 0.693 | 1.280 | 0.318 | 2.828 | -6.82% |
| return_resid_META_lb6_bw260_zw260_e2.4_google_only_h1 | return_residual | 1.166 | 1.290 | 1.077 | 2.295 | -5.53% |
| price_resid_META_bw260_zw260_e2.0_paired_h2 | price_residual | 0.978 | 1.656 | 0.452 | 1.803 | -8.64% |
| return_resid_XLK_lb3_bw260_zw520_e2.4_google_only_h1 | return_residual | 0.516 | 1.511 | -0.223 | 2.493 | -9.10% |
| return_resid_META_lb6_bw520_zw520_e1.8_google_only_h1 | return_residual | 1.006 | 1.554 | 0.579 | 1.884 | -6.12% |
| return_resid_META_lb3_bw520_zw520_e2.4_google_only_h1 | return_residual | 0.813 | 1.240 | 0.532 | 2.502 | -7.49% |
| return_resid_MSFT_lb6_bw520_zw260_e2.4_google_only_h1 | return_residual | 1.223 | 1.345 | 1.140 | 1.826 | -7.87% |
Residual Stress
| TC bps | Delay | Full | Recent 1Y | Prior | Mar+ OOS | MaxDD |
|---|---|---|---|---|---|---|
| 1.5 | 1 | 0.338 | 1.920 | -0.312 | 4.464 | -9.81% |
| 1.5 | 2 | 0.329 | 2.153 | -0.340 | 3.929 | -9.65% |
| 1.5 | 3 | 0.194 | 1.440 | -0.218 | 3.544 | -9.52% |
| 3.0 | 1 | 0.215 | 1.760 | -0.423 | 4.251 | -9.85% |
| 3.0 | 2 | 0.203 | 1.975 | -0.451 | 3.692 | -9.70% |
| 3.0 | 3 | 0.067 | 1.241 | -0.329 | 3.226 | -9.53% |
| 5.0 | 1 | 0.052 | 1.545 | -0.570 | 3.961 | -10.53% |
| 5.0 | 2 | 0.035 | 1.734 | -0.599 | 3.367 | -10.52% |
| 5.0 | 3 | -0.103 | 0.973 | -0.477 | 2.790 | -10.19% |
| 10.0 | 1 | -0.356 | 1.001 | -0.936 | 3.206 | -14.77% |
| 10.0 | 2 | -0.383 | 1.124 | -0.967 | 2.524 | -14.77% |
| 10.0 | 3 | -0.525 | 0.303 | -0.843 | 1.663 | -14.55% |
Strategy Eval
PAW daily adapter: Net Sharpe 0.352, CAGR 2.20%, MaxDD -8.91%, Bias CAUSAL — lead-cheat boosts Sharpe more than lag-stale drops it.
Decision
残差方向值得继续研究，但当前不替换 baseline。它的近期 OOS 很强，说明 “GOOGL/META residual” 在 2026 regime 里有信号；但 full Sharpe 只有 0.338，prior context Sharpe 是 -0.312，稳定性不如 baseline。更合理的下一步不是直接替换，而是把 residual 当成 baseline 的 confirm/gate 或小权重 ensemble。