What actually survives
Nothing in the signal. The apparatus, though, is worth keeping, and it is the answer to how any of this gets pointed at something real. Five rules, and the backtest earns each one:
- Score against the price, never against accuracy. The NCAAF stats model picks 77.1% of winners and still loses 1.6% a bet. NCAAB picks 72.7% and loses 3.6%. A model that agrees with the favourite is a mirror, not an edge. The desk already has this bruise: call C-001's favourite rule wins 74% of the time and loses 5.92c a contract.
- Walk forward or don't bother. Every model here trains only on seasons before the one it predicts. Train on the whole history and the gematria model's 63.7% NCAAB accuracy would have looked like a discovery instead of a team-name lookup table.
- Shuffle the calendar. If dates carry meaning, real dates must beat shuffled dates. Across six sports they never did at the 5% level. That single test is what separates a date-numerology claim from a date-numerology belief, and it costs an afternoon.
- Count your tries. 113 rules were tested. One cleared p<0.05; chance alone expects 5.7. Any leaderboard of rules, wallets or strategies needs this correction stapled to it, or the top of the table is just the widest sample error.
- Preregister the kill. The criteria were written before the run: a confidence interval above zero on moneyline ROI, or above 52.4% on the spread. Nothing reached it, so nothing ships. Deciding the bar afterwards is how a null becomes a strategy.
The honest use of a gematria engine is as a control. It is a signal generator with no possible mechanism, so whatever it scores is the noise floor for the harness that scored it. If a real signal cannot beat this, the real signal is noise too.
Method and limits
Six sports, 22,990 scored games: NFL 2004–26 (5,984), NCAAB 2025–26 (5,691), MLB 2025–26 (4,927), NHL 2024–26 (2,809), NBA 2024–26 (2,642), NCAAF FBS 2025 (937). Every model is fit on prior seasons only. Moneyline returns use the real posted price; spreads are scored at −110, so 52.4% is break-even. MLB and NHL have no spread column because their ±1.5 run and puck lines are not −110 markets. Confidence intervals are 95%. The permutation test reshuffles game dates and re-runs the oracle; perm p is the share of shuffles that matched or beat the real calendar.
What this does not show. Absence of an effect in 22,990 games is not proof of absence in every possible symbol scheme — it is a bound. The bound is tight: the widest confidence interval here would have caught a genuine three-point edge. The NBA, NHL, MLB and NCAAB windows are two seasons or fewer, so their bounds are looser than the NFL's twenty-two. Sample sizes on individual rules run from 89 to 1,211, and the small ones are where the flattering numbers live, which is the point of showing n first.
Provenance. The figures on this page are transcribed from the backtest's own result files into symbols.js; the strategy rows are its cross-sport summary and the rule rows are the top 25 of 113. The 113-rule aggregates are computed over the full set. No order was submitted at any point and there is no order-submission code, which is the correct amount for a signal that measures as noise.