Per-signal probability and expected value: premature at about 120 resolved trades a year

[yanivil]

Finding

The review's central architectural proposal, a probability of target-before-stop and an expected R per row with the score demoted, is right as a destination and not supportable now. A logistic model on the tuned rows (risk %, volume ratio, R:R, score, close vs SMA200, SMA50 vs SMA200, breadth, VIX, wait, pattern), fitted on one year and scored on the other:

Train	Score	N train	N scored	AUC	Mean R above / below the median predicted p
Y1	Y2	117	123	0.57	+0.35 / +0.21
Y2	Y1	123	117	0.56	+0.36 / +0.25

The coefficients change sign between years: risk % +0.75 then -0.20, R:R +0.02 then -1.48, volume ratio +0.36 then -0.05. A number with that little information in the morning report would mislead. The quality score is flat across buckets (#82) and gets no weight in the model: it stays a gate, not a ranker, until something validated replaces it.

Revisit when

• a multi-year replay with a point-in-time universe (Point-in-time constituents for multi-year replays #95) gives at least 500 resolved trades including a bear year,
• a model fitted on earlier windows scores AUC >= 0.6 on each later window with stable coefficient signs,
• the top quartile by predicted probability beats the median split in mean R on every window.

Until then the honest per-row information is pattern-level: replay hit rate and mean R per pattern in the wiki (#96).

Status: open. Related: #82.

---

Opened from the external review of 2026-09-07 (mathematical / algorithm recommendations) and the local two-year replay that tested it: 500 sessions to 2026-09-04, horizon 60, each scan seeing 500 bars like the nightly's 2y download, today's constituents. Year 1 = 2024-09-09 to 2025-09-08, never used for any tuning decision; year 2 = 2025-09-09 to 2026-09-04, the year every profile decision was made on. Year 2 reproduces the 2026-09-06 tuned replay (158 signals, 46 %, +0.29 R) within one signal.

[yanivil]

Ten-year point-in-time replay, 2026-09-08: GitHub backtest runs 34246882614 to 34247722648, one calendar year each from 2016 to 2026 (to 2026-09-04), tuned profile, asof=true, bars=500, horizon=60, grid=false; 1690 signals, 1629 traded, +0.23 R pooled, positive in 10 of 11 years. Pooled numbers weight each year's bucket by its traded count.

The first revisit condition is met: 1629 traded signals across eleven years including two bear years. The model test itself (fit on earlier years, score later ones, AUC at or above 0.6 with stable coefficient signs) has not been run. The rows live in the run logs; uploading backtest.json as a workflow artifact would make the fit reproducible without re-running the replays.