Verdict: Next-Day Upside Ranker (M-023) is rated “NOT FIT (Use A: closing-auction buy / next-open sell; Use B: '+3–10% tomorrow' picker)” in its validation review dated 6 Oct 2026.
Key result: Rank IC vs overnight, WF 2017–25: +0.0364 [+0.0293, +0.0422].
How it was tested: training data 2016-01→2025-12, test data 2017-01→2025-12; out-of-sample design (how it was checked on data it never saw while being built): walk-forward yearly.
Horizon: 1 night (close t -> open t+1). Coverage: US listings ≥$10M/day and ≥$5 (t10, ~2,270 names in 2026); trained on t2 (≥$2M/day, ≥$1) 2016→ incl. recovered delistings.
Forward record: 23 of 24 published lists have reached their 1-session horizon. Average pick return minus SPY's over the same dates: -0.20 percentage points (standard error 0.83). 50% of the matured picks beat SPY over the horizon.
Stage: Validated — NOT FIT (shadow only). Version v1; the hub record is as of 6 Nov 2026.
The verdict
NOT FIT for both uses; shadow (paper) scoring only.
Development summary: A small, persistent overnight ranking signal (IC +0.036) that does not survive as a book: the 2026 acceptance window fails every gate, the decision-time gap removes up to 43% of the top-10 excess, and fills are unmeasured. NOT FIT;…
What the review allows
Shadow scoring at 15:35 ET with paper MOC/MOO settlement, to measure the decision-time gap and fills
What the review rules out
Choose trades from the list (Use A)
Present it as finding stocks that rise 3–10% tomorrow (Use B)
Quote WF_REPORT book figures as expected returns
What it predicts
Item
Detail
Question it answers
Which liquid US stocks gain most overnight, close to next open? A pre-close daily list, shadow only.
Method
XGBoost regressor
How it works
XGBoost regressor on 53 within-date ranks, 3 seeds
Inputs
Close location in day's range, ATR% (volatility), Volume surprise today, 21-day overnight return drift, Today's open-to-close move
Output
Daily top-20 paper list at 15:35 ET, settled at official close/open prints
Horizon
1 night (close t -> open t+1)
Coverage
US listings ≥$10M/day and ≥$5 (t10, ~2,270 names in 2026); trained on t2 (≥$2M/day, ≥$1) 2016→ incl. recovered delistings
Markets
US
Model type
Model · ML/Stat
Stage
Validated — NOT FIT (shadow only)
Version
v1, first recorded 6 Oct 2026
Validation evidence
Training data: 2016-01→2025-12. Test data: 2017-01→2025-12. Out-of-sample design: walk-forward yearly.
Headline results from the model's development record, as the hub publishes them.
Measure
Result
Compared with
Note
Rank IC vs overnight, WF 2017–25
+0.0364
B4 12-1 +0.0362 · B6 +0.0179 · random +0.0003
Top-10 net excess (20 bp), WF
+27.4 bp/day
B6 −6.9 · B3 −8.3
Top-10 net excess (20 bp), 2026 acceptance
−7.8 bp/day
B6 +2.7 · B3 −0.6
15:40 decision-time proxy, top-10 net 20 bp
+6.9 [−2.4, +16.9]
16:00 bar +27.4
The hub's standard metrics for this model class (only those the record measured).
Metric
Sample / window
Compared with
Value
Range (CI)
Rank IC (headline)
WF 2017–25 · close→open
12-1 momentum +0.0362
0.0364
[0.0293, 0.0422]
Excess return at horizon
WF 2017–25 · top-10 · gross
12-1 momentum +14.2
+47.4 bp
[36.8, 59.2] bp
Precision at k
WF 2017–25 · top-10 · +3% next day
ATR-matched base
1.31×
n/a
Forward record: 23 of 24 published lists have reached their 1-session horizon. Average pick return minus SPY's over the same dates: -0.20 percentage points (standard error 0.83). 50% of the matured picks beat SPY over the horizon.
What the validation review found
Is the design sound? Daily cross-sectional ranking with gradient-boosted trees on within-date ranks and a demeaned, clipped overnight label is a sound design; all value sits in close→open, so any use depends on deciding before the close and filling at both auctions — none evidenced. The picks are volatile, low-priced day-t winners (overnight continuation), not the pre-registered panic-drop reversal.
Data quality. Feature timing audit 160/160; shuffled-label IC −0.0033; shifted-label IC +0.032 (persistent, no leak); source panel split- but not dividend-adjusted (negligible).
Backtest. WF top-10 net 20 bp +27.4 [+17.6, +39.3]; SR 2.09 [1.31, 2.95]; max drawdown −46.0%; worst night −29.3%.
Against simple benchmarks. Beats B6 by +34.4 bp/day in 2017–2025 (paired); loses to B6 by −10.4 (p 0.41) in 2026.
Independent replication. Exact: every WF_REPORT §2–§3 figure for X6/X1/L1/B6 replicates within tolerance; B3 differs only through clv = 0 ties.
Sensitivity. 15:40 proxy −36% IC / −43% top-10 excess; liquid universes ≈ 0 net; 40 bp costs leave top-10 +7.4 [−2.4, +19.3]; regime-dependent (VIX top tercile).
Decision-time gap: training saw the 16:00 close, production sees a ~15:20 ET delayed bar; the 15:40 proxy cut IC 36% and the top-10 excess 43% (net 20 bp +27.4 -> +6.9 [−2.4, +16.9]). The most important inputs (clv_1d, vol_z_1d, oc_1d, r_1d) are the ones the last 20 minutes move.
Auction fills unmeasured: TradingView close/open are assumed fillable MOC/MOO prints at 20 bp; picks are low-priced (median $17.6, 28% < $10) and the 40 bp top-10 net was +7.4 [−2.4, +19.3]; liquid universes (dv20 ≥ $200M) earn ≈ 0 net. No imbalance data is recorded.
Regime concentration: 2020–21 carry the walk-forward (top-10 net +82.4 / +59.0); 2017–19 and 2022 ≈ 0; VIX top tercile +49.6 vs +16 bp; max drawdown −46%; worst night −29.3%; 2026 acceptance net −7.8 [−27.5, +9.4].
The '+3–10% next day' claim is unsupported: top-10 P(y_cc ≥ 3%) is 1.31× the ATR-matched rate (1.16× in 2026), the +3–10% band 1.12×, while P(y_cc ≤ −3%) is 1.37× — a volatility selection, not an upside picker (VAL F-02).
Survivorship bounded only for recovered delistings; the production source (ohlcv) differs from the training panel (2026-09-03 parity: X6 rank rho 0.991, top-10 overlap 80%).
300-round cap binds (best iterations 298/294/296 in the frozen refit).
Questions
Is the Next-Day Upside Ranker validated?
Its validation review dated 6 Oct 2026 rated it NOT FIT for both uses; shadow (paper) scoring only.
What does the Next-Day Upside Ranker predict?
It answers: Which liquid US stocks gain most overnight, close to next open? A pre-close daily list, shadow only. Output: Daily top-20 paper list at 15:35 ET, settled at official close/open prints. Horizon: 1 night (close t -> open t+1). Coverage: US listings ≥$10M/day and ≥$5 (t10, ~2,270 names in 2026); trained on t2 (≥$2M/day, ≥$1) 2016→ incl. recovered delistings.
How was the Next-Day Upside Ranker tested?
Training data 2016-01→2025-12, test data 2017-01→2025-12, out-of-sample design: walk-forward yearly. Rank IC vs overnight, WF 2017–25: +0.0364 [+0.0293, +0.0422]; Top-10 net excess (20 bp), WF 2017–25: +27.4 bp/day [+16.8, +39.2]; Same, 2026 acceptance (169 sessions): −7.8 bp/day [−27.5, +9.4] — gates S1–S4 fail. Forward record: 23 of 24 published lists have reached their 1-session horizon. Average pick return minus SPY's over the same dates: -0.20 percentage points (standard error 0.83). 50% of the matured picks beat SPY over the horizon.
Out-of-sample: results measured on data the model never saw while it was built or tuned.
Walk-forward: the model is refit on data up to a date and tested on the period after it, then the window rolls forward, so every test is out of sample.
Rank IC: information coefficient: the rank correlation between a model's scores and the returns that followed; 0 means no skill.
Basis points (bp): hundredths of a percentage point: 25 bp is 0.25%.
Standard error (s.e.): the typical size of the error in an estimate; a result more than 2 standard errors from zero is unlikely to be luck.
Confidence interval: the range that likely contains the true value (95% unless stated); [a, b] on this page is that range.
SPY: the SPDR S&P 500 ETF, the US market benchmark used here.
ATR: average true range: a stock's typical daily price range, often expressed as a percentage of its price.
Shadow: scored and tracked on paper only; no list is published and nothing is acted on.
12-1 momentum: the return over the past 12 months, leaving out the most recent month.