Block uncertainty
Resample 20-session blocks 2,000 times so overlapping outcomes are not treated as independent rows.
MODEL 1 · STEP 5Stress
Stress
the result.
Run 5 is the post-hoc least-bad development candidate. Nine predeclared gates decide whether that ranking is strong enough to consider deployment.
MARKET CONTEXT · OUTSIDE MODEL FREEZEASTS
These readings update independently. They are not model inputs, deployment evidence, or a trading recommendation.
- Latest price
- $62.31 · fallback
- 20D momentum
- -13.4% Close-to-close · 20 completed sessions
- RSI 14
- 47.2 neutral · Wilder method
- Relative volume
- 1.0× Versus the prior 20 completed sessions
TECHNICALS THROUGH · TIMESTAMPED FALLBACK
Does the post-hoc least-bad candidate earn consideration?
The answer depends on evidence, not rank. Run 5 can be least bad among the five inspected candidates and still fail against the prior, uncertainty, calibration, stability, and null benchmarks.
REFERENCEFrozen Run 5 OOS path
CHALLENGEFour stress families
DECISION1 of 9 gates pass
FOUR STRESS FAMILIES
Try to disprove the post-hoc ranking.
The family-wise null circularly shifts the frozen OOS label path against frozen forecasts. It excludes shifts within 63 sessions of zero and protects against selecting the strongest of five runs.
Family-wise null
Circularly shift the frozen OOS label path and compare the strongest of five development runs.
Calibration
Audit slope, intercept, five-bin calibration error, and confidence coverage.
Sensitivity
Change block length, inspect regimes, remove extreme-event windows, and measure feature drift.
STRESS RESULTS
Run 5 remains weaker than the prior.
The primary 20-session bootstrap keeps zero inside the Brier improvement interval and 0.50 inside the AUC interval.
Run 5 · shallow boosting
Worse than the fold-local prior
Lower bound remains below 0.50
Frozen-forecast circular-shift null
Above the 5% evidence gate
20-SESSION BLOCK BOOTSTRAP · 95% INTERVALS2,000 repetitions
ROC AUC0.3800–0.5379
0.50
Brier improvement−0.00498 to +0.00131
0
Run 5 remains the post-hoc least-bad candidate, but that label only ranks the five inspected runs. Eight of nine deployment evidence gates fail, so no model should move forward.
GATES PASSED1 / 9
STEP 5 EVIDENCE GATES
One pass cannot rescue eight failures.
These are conservative research guardrails, not nine independent hypothesis tests. The observed and required values stay visible.
| Gate | Rule | Observed | Required | Result |
|---|---|---|---|---|
| G1 | Positive Brier improvement with 20-session lower bound above zero | −0.00170; lower −0.00498 | Point > 0; lower > 0 | FAIL |
| G2 | ROC AUC and its 20-session lower bound clear no-skill | 0.4577; lower 0.3800 | AUC ≥ 0.52; lower > 0.50 | FAIL |
| G3 | Family-wise frozen-forecast circular-shift tail rate | 96.59% | < 5% | FAIL |
| G4 | Run 5 improves Brier score on Run 1 with paired lower bound | +0.00197; lower −0.00222 | Point > 0; lower > 0 | FAIL |
| G5 | Calibration slope | −1.319 | 0.70 to 1.30 | FAIL |
| G6 | Absolute calibration intercept | 0.098 | ≤ 0.10 | PASS |
| G7 | Five-bin expected calibration error | 6.44% | < 5% | FAIL |
| G8 | Outer folds with positive Brier improvement | 6 / 12 · 50% | ≥ 60% | FAIL |
| G9 | Block-length, event-window, and supported-regime sensitivity | Lower −0.00503; event −0.00242 | All sensitivity conditions pass | FAIL |
PHASE B / SEPARATE REPAIR AUDITS
No lawful candidate fixes all three weaknesses.
The repair program tests flatness, equal-frequency qcut calibration error, and fold consistency separately. Each row keeps its own nine-gate denominator.
| Candidate | Probability spread | qcut ECE | Positive folds | Individual result |
|---|---|---|---|---|
| Prior-rank Ridge family output | 0.4642 to 0.4976 Std. dev. 0.0108 Median within-fold std. dev. 0.000569 | 0.060283 | 9 / 12 | 2 / 9 / G4 / G8 |
| Empirical-bin family member | 0.4257 to 0.5433 Std. dev. 0.030792 Median within-fold std. dev. 0.027692 / maximum 0.039467 | 0.055738 | 8 / 12 | 2 / 9 / G6 / G8 |
| Non-flat state model | 0.3225 to 0.5410 Std. dev. 0.0370 Median within-fold std. dev. 0.021151 | 0.043606 | 6 / 12 | 1 / 9 / G7 |
| Neutral-anchor Core Ridge | Prior fallback in 9 / 12 folds Std. dev. 0.0120 Median within-fold std. dev. 0.000000 across all folds / 0.006596 active-only | 0.060468 | 2 / 12 | 1 / 9 / G4 |
| Legacy direction-aware Core Ridge | Compressed legacy output | 0.066126 | 4 / 12 | 1 / 9 / G4 |
The selected calibration family passes G4 and G8. The non-flat state model passes G7. The unselected empirical-bin member passes G6 and G8. Neutral-anchor and the corrected legacy direction-aware audit pass G4 only. These passes belong to different probability paths and cannot be combined.
The selected prior-rank Ridge family reaches 9 of 12 positive folds, yet its median within-fold forecast standard deviation is only 0.000569 and qcut ECE is 6.028%. The non-flat model spans 32.25% to 54.10%, has a 0.021151 median within-fold standard deviation, and passes qcut ECE at 4.361%, but wins only 6 of 12 folds and loses to the prior overall.
The empirical-bin family member was not selected. It passes G6 and G8, posts +0.000871 Brier improvement and 0.502218 AUC, spans 42.57% to 54.33%, and reaches 8 of 12 positive folds. Its median within-fold standard deviation is 0.027692, but qcut ECE is 5.5738%, missing G7 by 0.5738 percentage points. It remains development-only and does not raise the best gate count above 2 of 9.
This corrected audit uses the original frozen reference, qcut ECE, seed 20260730, and 5/20/63-session blocks. The superseded legacy gate file must not be used. Its only pass is G4: +0.003139 / lower +0.000281. qcut ECE is 0.066126, and the family-wise tail rate is 0.970779.
Calibration family gates / Non-flat state gates / Neutral-core gates / Corrected legacy direction-aware gates. The Phase A table above remains unchanged.
This is a direction-probability exercise, not a trading backtest. It does not model positions, overlapping trades, execution, transaction costs, or portfolio risk.
STEP 5 OUTPUTmodel1_evidence_gate_v2
One frozen audit of uncertainty, family-wise null evidence, calibration, folds, regimes, drift, and event sensitivity.
NEXTStep 6 · Summarize the model
Turn the evidence into a durable recommendation.