Extended-hours
mover audit.
Frozen A3 walk-forward rankings were re-judged against the next U.S. session from 04:00–16:00 ET using Alpaca SIP historical bars. This page is diagnostic research evidence, not a trading-return claim. Split/reverse-split exclusion and targeted NO_BAR provenance checks remain pending.
← Back to product shellWhat the old target is missing
As the move threshold rises, premarket peaks and extended-hours-only movers become more prominent. That is the main reason this audit is not reducible to a Top20 reranking problem.
Recall / precision by move threshold
Recall asks how much of the true mover set was captured by each frozen rank cut. Precision asks how many names inside that cut actually crossed the move threshold.
| Move | Movers | R@10 | P@10 | R@20 | P@20 | R@50 | R@100 | R@300 | P@300 | Median rank |
|---|---|---|---|---|---|---|---|---|---|---|
| +10% | 44,745 | 1.54% | 57.02% | 2.96% | 54.67% | 6.60% | 11.81% | 27.72% | 34.17% | 654 |
| +20% | 12,703 | 3.76% | 39.50% | 6.77% | 35.54% | 13.44% | 21.70% | 40.74% | 14.26% | 376 |
| +30% | 6,252 | 6.03% | 31.16% | 9.85% | 25.45% | 18.27% | 27.80% | 46.35% | 7.98% | 268 |
| +50% | 2,918 | 8.50% | 20.50% | 12.89% | 15.54% | 22.21% | 31.70% | 48.94% | 3.93% | 198 |
| +100% | 1,171 | 8.45% | 8.18% | 12.21% | 5.91% | 21.35% | 30.06% | 43.64% | 1.41% | 158 |
Miss taxonomy
+30% movers beyond rank 3002,649 / 6,252 · 42.4%
+30% model-ineligible or no OOF score705 / 6,252 · 11.3%
+50% movers beyond rank 3001,003 / 2,918 · 34.4%
+50% model-ineligible or no OOF score487 / 2,918 · 16.7%
NO_BAR across all signal rows3,844 / 660,944 · 0.58%
Current research decision
Direction: extended-hours target redesign is the leading candidate, not yet the final verdict.The frozen decision protocol requires the corporate-action / NO_BAR integrity pass before formally promoting a next-stage design. Raw evidence already shows that true 04:00–16:00 movers are too often outside Top300 and that a material fraction of movers exist only under the extended-hours target. If clean metrics preserve this pattern, Close V3 should target the actual D-day 04:00–16:00 opportunity directly while preserving expanding / walk-forward OOF and benchmark blindness.