We test the strategies this industry sells.When they fail, we publish the failure.
These are not opinions. They are pre-registered experiments with fixed decision rules, observed numbers, and a verdict rendered against those rules — not against whether we wanted the result to be favorable.
Every entry below is a strategy commonly sold in this industry. None of them survived our tests.
These figures are transcribed from an internal study whose raw computation is not published. They are not an independent audit of the named sportsbooks.
L-15
Market-level close-prediction
Mechanism
A Ridge model on 5-fold grouped cross-validation by game, predicting the change in median Shin no-vig probability from entry snapshot to closing line. The pre-registered model used the full pre-entry feature set (alpha = 1.0); a post-hoc entry-probability-only Ridge was run as the diagnostic that exposed the artifact. 776 entry-window labels across 216 games, drawn from the 241-game clean-close corpus.
Pre-registered rule
Pre-registered decision threshold: grouped-CV Spearman r ≥ 0.15 continue; r < 0.10 stop, no appeal.
Observed
Totals grouped-CV Spearman r = 0.490. That number looks promising, but it is a textbook corr(X, Y−X) artifact: corr(p_entry, p_close) = 0.40, so corr(p_entry, Δ) = −0.51 by identity. A Ridge on p_entry alone matches the full model (r = 0.503, R² = 0.284). Moneyline grouped-CV r = 0.091 with negative R². Totals no-vig P(over) itself sits in a narrow band around 50% (standard deviation 1.35 percentage points), which is what makes the artifact dominate.
Verdict
Kill. The apparent signal is measurement error dressed up as forecast skill. There is no tradable close-prediction edge at public-data cadence on this corpus.
L-16A
Per-book shading
Mechanism
For each book, measured the mean deviation of its entry price from the consensus close (e = p_book_entry − p_median_close). Liang-Zeger standard errors clustered by game. Anti-shade bet: when |p_book − p_median_entry| ≥ 0.005, take the opposite side; score CLV against the consensus close.
Pre-registered rule
Pre-registered kill clauses: no book has |t| > 2 with n ≥ 50, or no such book shows positive mean CLV on ≥150 bets.
Observed
11 executable books, 863–947 labels each. The largest absolute t-statistic is betus at +1.48 (mean e = +0.12pp, n = 901). No book clears |t| > 2. On the |deviation| ≥ 0.5pp subset, anti-shade CLV is +1.2 to +1.8pp, but that subset is small and the pattern is outlier noise, not persistent shade.
Verdict
Dead. No book persistently shades the totals close in a way that survives the pre-registered gate. Do not build a copier or an outlier screen on this corpus.
L-16B
Cross-book lead-lag
Mechanism
At ~19-minute cadence, computed the first-order autocorrelation of each book's price changes (r_b,t = p_b,t − p_b,t−1). A leads B if ρ₁(A,B) > 0.1 and ρ₁(B,A) ≤ 0 with Benjamini-Hochberg q < 0.05. 110 ordered pairs tested; first 120 games for selection, last 121 for holdout CLV.
Pre-registered rule
Pre-registered kill clause: no qualifying pair has positive mean CLV on ≥150 holdout trades.
Observed
Raw-lead count = 0. BH-lead count = 0. The largest raw ρ₁ is bovada → draftkings at 0.075; the reverse is +0.054, so it is not one-way. No pair is simulated because none clears the selection gate.
Verdict
Dead. No book leads another at this cadence in a way that survives the pre-registered gate. Do not build a copier or a lag-screen on this corpus.
L-17
Price-path geometry
Mechanism
Six path-geometry features computed strictly before the pre-entry snapshot: realized variation, increment autocorrelation, sign-change rate, dispersion half-life, cross-sectional skew, and staleness fraction. All six passed decimation. Ridge with alpha = 1.0, grouped 5-fold CV by game, predicting realized CLV on 210 MLB totals games with ≥10 pre-entry snapshots.
Pre-registered rule
Pre-registered decision threshold: grouped-CV Spearman r ≥ 0.15 continue; r < 0.10 stop, no appeal.
Observed
Totals grouped-CV Spearman r = 0.091 (n = 203 games, p = 0.19, R² = −0.049). Moneyline r = −0.047, R² = −0.072. Spreads r = 0.209 with negative R². Mean CLV is −0.00045 (sd 0.010). Univariate realized variation is −0.192, but it is one of six looks and was not used to change the decision.
Verdict
Stop. The pre-registered rule fires at r < 0.10. The research program ends on this corpus. No booster, no second experiment, no appeal.
Why this exists
The sports prediction industry is built on strategies that sound plausible and are never tested in the open. We run those tests, pre-register the rules before we look at the data, and publish the result whether it helps us or hurts us.
These are things we tested and found not to work. We publish them so you do not pay for the same mistakes.
Sports wagering is real risk. Only stake what you can afford to lose without changing your week. If you or someone you know has a gambling problem, call 1-800-GAMBLER.