Source: task5-ashare-leader/output/streaming_hist_2012/v1_4_task37/task37_first_batch_report.md
Last modified: 2026-05-21 22:18:15 · 4,024 bytes

Task #37 - M-filter ON Quality + Weak-Regime Stop Calibration First Batch

Scope: use existing mfilter_sensitivity and stop_ab outputs to answer whether 2023 churn can be mitigated by stronger M-filter ON confirmation or stop-rule changes.

Executive conclusion

First batch result: **M-filter ON confirmation helps 2023, but over-tight confirmation is not a production replacement yet. Stop-rule variants do not solve 2023.**

- More stringent ON confirmation (open10 / open20) can reduce 2023 loss to near 0 by essentially avoiding 2023 exposure, but this also cuts full-window return/Sharpe materially versus v1.3 Y2 baseline.

- Moderate confirmation (open5_close2/3) improves 2023 from the #31 baseline -10.51% to about -5.58%, but full-window Sharpe is still only ~0.72, below v1.3 baseline ~0.86.

- Stop variants (hard8/10/12, hybrid8/10) keep 2023 near -10%, so the main 2023 issue is not solved by simple stop calibration.

Selected evidence

scenario type ret_2023 avg_exp_2023 market_on_pct_2023 full_ann full_sharpe full_mdd
open20_close3 mfilter 0.0000 0.0000 0.0000 0.1123 0.7215 -0.2743
open10_close2 mfilter -0.0333 0.0278 0.0207 0.1048 0.6466 -0.2740
open5_close2 mfilter -0.0558 0.0637 0.0537 0.1248 0.7278 -0.2789
open5_close3 mfilter -0.0558 0.0637 0.0620 0.1238 0.7186 -0.2840
open1_close1 mfilter -0.1141 0.1920 0.0826 0.1234 0.7255 -0.2877
open3_close2 mfilter -0.1169 0.1971 0.0702 0.1230 0.7133 -0.2720
open2_close2 mfilter -0.1169 0.1971 0.0785 0.1088 0.6485 -0.2721
hybrid8_three_axis stop -0.1007 0.1706 0.0826 0.1103 0.6667 -0.2728
hybrid10_three_axis stop -0.1034 0.1864 0.0826 0.1052 0.6524 -0.2741
three_axis_v1_2 stop -0.1141 0.1920 0.0826 0.1234 0.7255 -0.2877
hard12_only stop -0.1226 0.9101 0.0826 0.0397 0.2903 -0.6234
hard8_only stop -0.1306 0.7667 0.0826 0.0577 0.3728 -0.5508
hard10_only stop -0.1387 0.8053 0.0826 0.0454 0.3186 -0.5843

Interpretation

2023 confirms #31: the book lost money during low-quality ON windows. If we require longer ON confirmation, those windows disappear and 2023 improves. But the same filter also misses too much valid bull-regime entry over the full sample.

Stop calibration alone is lower priority: existing stop A/B does not materially improve 2023. The stop system is reacting after low-quality entries; the higher-leverage fix is entry/regime confirmation, not a different stop threshold by itself.

Next experiment

Do not adopt open10/open20 globally. Next #37 batch should test **conditional confirmation only in weak/ambiguous regimes**, e.g.:

1. Keep current v1.3 ON rule in normal regimes.

2. Require extra ON confirmation only when breadth/liquidity/policy sub-factors are mixed or recent M-filter churn is high.

3. Optional: block new buys for N days after an OFF→ON flip unless benchmark breadth confirms; existing holdings still follow normal exit rules.

Decision for v1.4: #37 remains active, but first batch rejects simple global open10/open20 replacement and deprioritizes stop-only fixes.