New to these reports? Start here
- Dotted-underlined words have a plain-English definition — hover or tap them. Every term is also on the glossary page.
- "Null" means we found nothing, not that something broke. Most reports here are negative results, on purpose — knowing an idea doesn't work is the point.
- Two questions get asked separately. First, is the effect real? Second, is it already priced into the betting odds? An effect can be completely real and still useless to bet on.
- A "calibration" row is a self-check. It runs the same method on something already known to be true. If that fails, the whole report is unreliable — so it's reported alongside the findings.
- If a confidence interval includes zero, the real effect might be nothing at all, so no claim gets made.
NCAAB consensus-stacking evaluation — REJECT (2026-07-04)
The best-prior candidate (soft small-conference markets), tested on our largest capture. Verdict: consensus wins everything.
Data
unified_odds ncaab: 1,637 games / 92,897 snapshots, 2026-02-01→03-20 (one partial season; Nov soft-market games untested — capture began Feb 1). Results via bounded ESPN backfill (espn_results.json 1,943 + espn_history.json 4,304 for features). Matched odds↔results 1,568/1,595 (98.3%). Mean 4.58 books/game at close; NO Pinnacle in this capture.
Experiment (faithful NHL v3 port)
Devigged mean consensus + rest/b2b/form-L10/conference-flag, logistic C=0.5, expanding walk-forwardwalk-forwardEvaluating week by week using only what was knowable before each week, mimicking how the model would actually have been used at the time. (5 folds, 906 OOS predictions).
| OOS n=906 | stacked | raw consensus |
|---|---|---|
| accuracy | 0.6678 | 0.6788 |
| log losslog lossA score for probability forecasts that punishes confident wrong answers harshly. Lower is better. | 0.5854 | 0.5745 |
| brier | 0.2022 | 0.1977 |
Consensus won ALL 5 folds; gap doesn't shrink with more training data. Ablation: full model 0.5854 > consensus-only recalibration 0.5798 > raw consensus 0.5745 — the close is already calibrated AND the features are noise.
Soft-market hypothesis FAILS: low-liquidity (<=3 books, n=129) consensus still wins log loss (0.5480 vs 0.5575). Flat ROIreturn on investmentProfit as a percentage of the money wagered. +2% means $2 profit per $100 bet. at best close monotonically worse with higher claimed edge (-9.1% @1% -> -15.2% @5%) — disagreements are model errors.
Caveats
One partial season, Feb-Mar only (conference play), no sharp book captured. Direction uniform across folds/buckets/ablations though — matches NBA/NFL. Burden of proof is on any future attempt (would need Nov-Dec small-conference capture + a sharp book before re-testing).
Reproduce: run_experiment.py, diagnostics.py; raw numbers in experiment_output.json.