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.
Kelly-sizing simulation across site models — 2026-07-04
Question: would any model be profitable with KellyKelly criterionA formula for how much to stake given your edge. Full Kelly maximizes long-run growth but swings wildly; most people use a fraction of it. sizing? Method: reproduce
each model's 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. probabilities, replay against real closing odds,
bankroll sims (100u start) under flat / quarter / half / full Kelly, two
selection rules (favored-side-always vs Kelly-positive-only). Reproduce with
build_<model>.py (writes picks JSON) then kelly_sim.py.
Answer: only NHL stacking v3 — and only at flat or quarter-Kelly
| model | flat ROIreturn on investmentProfit as a percentage of the money wagered. +2% means $2 profit per $100 bet. | best Kelly cell | full Kelly |
|---|---|---|---|
| NHL stacking v3 (n=63k per-book, 2009-25) | +4.1% @ ev>=3% | +2,059u quarter-Kelly on +EV picks, no bust | ruin |
| NBA Moneyline v1 (n=4,294) | -3.0% | all Kelly variants bust | ruin |
| MLB Moneyline v2 (n=274) | -5.3% | never exceeds 100u | -81% |
| MLB Batter Hits (n=5,408) | -6.6% | >100u cells are path variance (flat ROI of same picks negative) | bust |
The mechanism (disagreement-bucket flat ROI)
Kelly stakes scale with model-market disagreement, so bucket ROI by |p - mkt|: - NHL v3: -3.8% (0-3pt) -> +7.6% (6-10pt) — disagreements carry information; Kelly's overweighting is rewarded. Signature of real edge. - NBA/MLB: flat-to-worse with disagreement — disagreements are model errors; Kelly concentrates the bankroll on the worst bets. - Fake-edge tell: NBA's avg Kelly fraction ~0.15-0.20 (a market-trailing model "seeing" 15-20% edges) busts even quarter-Kelly. NHL's ~0.035-0.04 compounds.
Universal lessons
- Kelly creates edge for nobody — it amplifies the flat-stake sign.
- Full Kelly busts every model, including the profitable one.
- Safe operating point: flat or quarter-Kelly on the +EV subset, only when the underlying flat edge is independently real.
Caveats: single historical path; NHL rows per-book (correlated); same-day slate correlationcorrelationHow closely two things move together, from -1 (opposite) through 0 (unrelated) to +1 (in lockstep). It does not by itself mean one causes the other. not modeled; all backtest — the 2026-07 track-record audit showed backtests flatter reality. If acted on: quarter-Kelly on the NHL v3 alert stream when the season resumes (Oct), judged on live CLVclosing line valueWhether you got a better price than the market settled at. Widely used as a faster signal of skill than profit, which takes ages to measure. first.