Key result: +10% touch rate, 10 sessions: 38-47% vs 13-18% unconditional base rate.
How it was tested: training data 2018-08→2026-08, test data 2018-08→2026-08; out-of-sample design (how it was checked on data it never saw while being built): none.
Horizon: 10 sessions — measured as both the score's own fit horizon and the +10% target window. Coverage: ~1,500 liquid US names (≥$50M/day, >$3) plus the NSE India pool the scanner covers.
Forward record: 37 of 46 published lists have reached their 10-session horizon. Average pick return minus SPY's over the same dates: -4.93 percentage points (standard error 1.53). 35% of the matured picks beat SPY over the horizon.
Stage: Production. Version v1.3; the hub record is as of 6 Nov 2026.
The verdict
FIT FOR LIMITED USE as a reach/volatility selector for the sleeve: the rebuild is a real, evidenced improvement in which names get shown, but it is not a directional edge and its lift is still an in-sample replay.
Development summary: The 2026-08-15 two-track rebuild lifts replayed top-50 hits ~6-8x over the pre-rebuild screen, but no tested design clears 2 standard errors of SPY alpha; the one stable finding is right-tail TOUCH frequency, not direction.
What the review allows
As a daily longlist of names with wider-than-average measured 10-session outcome distributions, for human-reviewed tactical entries inside the sleeve.
The frozen per-pick rationale as a point-in-time record of why a name was picked (RS, ATR, H&L context) — it is not re-read live.
The Track scoreboard (W1-W4, alpha, MFE/MAE) as a forward performance record of picks actually logged.
What the review rules out
Read a pick as a directional forecast — no tested design clears 2 standard errors of SPY alpha in any of 4 regime blocks.
Expect 3-4 of a specific week's top-10 — top-20 (~2.4-2.5/week) is the realistic KPI; top-10 is 1.5-1.8/week even on the best tested design.
Treat the 1.95-vs-0.24 or 38-47%-touch figures as out-of-sample or live-confirmed — both come from a replay over the window used to choose the thresholds.
What it predicts
Item
Detail
Question it answers
Which liquid long names deserve a tracked daily pick, across a mean-reversion track and a momentum track, when an unaided list of ~12 names/week lands 0.29 top-10 hits/week — about 2.4x luck?
Method
Two-track rule screen
How it works
Two-track rule cascade (liquidity/earnings/sector/cooldown gates, per-track score, RS-rank, ATR floor, one H&L veto); frozen thesis logged at pick time, re-scored forward daily off real prices
quality_dip (3 US + 2 India/day) and momentum (2 US + 1 India/day) picks, ranked within each track, with a frozen one-paragraph rationale
Horizon
10 sessions — measured as both the score's own fit horizon and the +10% target window
Coverage
~1,500 liquid US names (≥$50M/day, >$3) plus the NSE India pool the scanner covers
Markets
US, India
Model type
Model · Rule
Stage
Production
Version
v1.3, first recorded 15 Aug 2026
Validation evidence
Training data: 2018-08→2026-08. Test data: 2018-08→2026-08. Out-of-sample design: none.
Headline results from the model's development record, as the hub publishes them.
Measure
Result
Compared with
Note
Top-50 hits/week
1.95
0.24 pre-rebuild screen
2018-08→2026-08 replay of the shipped rules
+10% touch rate, 10 sessions
38-47%
13-18% unconditional liquid-universe base rate
the actionable finding — reach, not direction
Per-trade alpha vs SPY, best of 8 tested designs
≈+0.4%/trade (E)
0 designs clear 2 s.e. in all 4 regime blocks
§5; current screen alone runs −0.22% to +0.02%
Reconstruction fidelity
exact RS/BBW/score/tag match
live swing
earnings-date and H&L gates not reconstructible pre-rebuild; the replay is an upper bound on live performance
Forward record: 37 of 46 published lists have reached their 10-session horizon. Average pick return minus SPY's over the same dates: -4.93 percentage points (standard error 1.53). 35% of the matured picks beat SPY over the horizon.
What the validation review found
Is the design sound? Correctly reframed after the 2026-08-15 audit from an implied directional screen to an explicit reach/touch-frequency selector; the momentum-track rationale text says so. quality_dip's methodology (best score bucket, RS, ATR floor, H&L veto) is coherent with that framing; momentum's is simpler and equally coherent.
Data quality. Frozen-snapshot design is sound (avoids the IBD 'stale state' defect by construction). The panel is ~1,700 names, not the full scanner universe — a real coverage limitation, not a data-quality defect.
Backtest. 's replay is a serious, multi-cut backtest (97 weeks, 4 regime blocks, a real trade-plan simulation) but it is also the SAME study that selected the shipped thresholds — the improvement it reports is not an independent out-of-sample check.
Monitoring. No monitoring layer beyond the tab's own forward scoreboard exists yet; the tier-2 default (monthly PSI + outcome loop) is not evidenced as running for this card specifically.
Limitations
Not a directional edge: the current screen ran −0.22% to +0.02% alpha vs SPY across 4 independent 2-year regimes (2018-2026); no tested variant clears 2 standard errors in either direction.
The headline improvement (1.95 vs 0.24 top-50 hits/week, 38-47% vs 13-18% touch rate) is an in-sample replay over the same window used to pick the thresholds — not yet confirmed out-of-sample or live.
Hard arithmetic ceiling: at ~12-15 distinct names/week out of ~1,500+ liquid names, expecting 3-4 of a specific top-10 is a ~30% precision demand against a 0.67% base rate — top-20 (not top-10) is the realistic weekly KPI.
10.7% of top-20 historical winners were rejected for having no 200-DMA history (recent listings); no substitute trend test has been confirmed shipped.
35% of top-10 historical winners sat below their 200-DMA — the momentum track's 52w-high/50-DMA rule was built to capture this, quality_dip still requires the 200-DMA.
The measured R/R odds (p_target_first, p_stop_first, exp_r) are base rates of the US replay of the STRUCTURAL gates (earnings gate, H&L veto and score gate not rebuilt); India rows get the arithmetic fields only. Expected R is about 0 in every bucket: the fields describe the shape of the bet, not an edge.
Questions
Is the Daily Picks — Two-Track Screen validated?
Its validation review dated 3 Sep 2026 rated it FIT FOR LIMITED USE as a reach/volatility selector for the sleeve: the rebuild is a real, evidenced improvement in which names get shown, but it is not a directional edge and its lift is still an in-sample replay.
What does the Daily Picks — Two-Track Screen predict?
It answers: Which liquid long names deserve a tracked daily pick, across a mean-reversion track and a momentum track, when an unaided list of ~12 names/week lands 0.29 top-10 hits/week — about 2.4x luck? Output: quality_dip (3 US + 2 India/day) and momentum (2 US + 1 India/day) picks, ranked within each track, with a frozen one-paragraph rationale. Horizon: 10 sessions — measured as both the score's own fit horizon and the +10% target window. Coverage: ~1,500 liquid US names (≥$50M/day, >$3) plus the NSE India pool the scanner covers.
How was the Daily Picks — Two-Track Screen tested?
Training data 2018-08→2026-08, test data 2018-08→2026-08, out-of-sample design: none. Top-50 hits/week, 2018-08→2026-08 replay: 1.95 vs 0.24 pre-rebuild screen; +10% touch rate, 10 sessions: 38-47% vs 13-18% unconditional base rate; SPY-relative alpha, any of 8 tested designs: none clears 2 standard errors. Forward record: 37 of 46 published lists have reached their 10-session horizon. Average pick return minus SPY's over the same dates: -4.93 percentage points (standard error 1.53). 35% of the matured picks beat SPY over the horizon.