Task #37 Conditional ON Confirmation Proxy Report
Date: 2026-05-21
Scope
This batch tests conditional M-filter ON confirmation using available proxy signals rather than true M-filter factor-strength components.
Available output did not expose the full 4-factor M-filter strength decomposition as a clean production-ready feature. The proxy grid therefore used recent market_on churn / run-length style conditions to decide when to replace the baseline ON signal with stricter open5 or open10 confirmation.
This is a research proxy only. It should not be treated as a production spec candidate unless later reproduced with explicit factor-strength inputs.
Decision
The conditional churn proxy batch is rejected for v1.4 production.
It does not satisfy the inclusion requirement of improving 2023 while preserving full-window Sharpe/MDD. The best full-window Sharpe in the proxy grid is only about 0.63, far below the v1.3 Y2 full-window anchor near 0.86 and below the first-batch mfilter sensitivity baseline near 0.72 in this diagnostic framework.
Evidence
Selected rows from conditional_confirmation_proxy_grid.csv:
| scenario | 2023 return | 2023 avg exposure | 2023 trades | full ann. | full Sharpe | full MDD |
|---|---|---|---|---|---|---|
| baseline_open1_close1 | -11.41% | 19.20% | 36 | 9.07% | 0.618 | -26.40% |
| cond_basechurn_w20_gt3_use_open10 | -11.41% | 19.20% | 36 | 9.34% | 0.632 | -27.08% |
| cond_basechurn_w20_gt3_use_open5 | -11.41% | 19.20% | 36 | 9.25% | 0.626 | -25.34% |
| global_open5_close2 | -5.58% | 6.37% | 16 | 8.68% | 0.594 | -27.20% |
| cond_basechurn_w60_gt2_use_open10 | -9.43% | 16.53% | 28 | 8.27% | 0.586 | -27.56% |
| global_open10_close2 | -3.33% | 2.78% | 10 | 6.07% | 0.462 | -29.22% |
Two patterns are clear:
1. Proxy variants that preserve full-window Sharpe mostly leave 2023 near baseline loss.
2. Variants that improve 2023 reduce exposure/trades and damage full-window Sharpe materially.
Interpretation
The churn proxy is too blunt. It behaves like a weak version of global ON confirmation: it removes some false ON exposure, but also blocks too much valid entry exposure across the full sample.
This supports the first-batch conclusion rather than overturning it:
- Global stricter ON confirmation is not viable.
- Churn-proxy conditional confirmation is also not viable.
- Stop-only changes remain lower priority because the main issue is low-quality entry windows, not post-entry stop threshold alone.
Next Research Direction
Do not continue expanding churn/run-length proxy grids unless a specific new hypothesis is added.
If #37 continues, the next useful version should use actual regime-strength inputs, preferably:
1. 4-factor M-filter vote count, e.g. require extra confirmation only when 2/4 or 3/4 factors pass.
2. Breadth-strength input, e.g. benchmark/component breadth above trend.
3. Liquidity/policy sub-factor strength if already available in the M-filter implementation.
Decision gate remains strict: a candidate must improve the 2023 weak-regime result while keeping full-window Sharpe and MDD close to or better than v1.3 Y2. Otherwise #37 should stay out of v1.4 production and remain only an operational governance insight.