Does the engine produce edge that simple logistic models don’t?
Each cell reports the metric ± a BCa bootstrap 95% confidence interval (Efron 1987). Cross-cell claims are Benjamini–Yekutieli corrected at q = 0.10 across the day’s full set of trials (López de Prado 2018, Ch. 6). The “Null” column is the no-skill baseline: predicting the in-sample base rate every time. A cell is colored teal (↑) only when its CI is fully above the comparator’s CI after FDR correction; rose (↓) when fully below; ink when CIs overlap.
| Metric | LLM | Logistic-24 | Logistic-canonical | Null |
|---|---|---|---|---|
Brier | 0.030 [—, —] | 0.250 [0.249, 0.250] | 0.250 [0.250, 0.251] | 0.250 — |
IC | +0.253 [—, —] | — [—, —] | -0.137 [—, —] | +0.000 — |
Log-loss | — [—, —] | 0.693 [—, —] | 0.694 [—, —] | 0.693 — |
q-values use Benjamini–Yekutieli over n_trials_attempted = 5000 cells.