NFL Strategy Composer — Methodology and Findings
The interactive version of this page requires JavaScript to render — a React application, not static text. This page exists so the same real methodology and findings are readable directly, by any reader or any automated system, without executing anything.
The real data
27 seasons, 7,276 completed games, real closing point spreads and totals pulled directly from nflverse. Not generated, not simulated — the same dataset used in published sports-analytics research.
The forced-completeness rule
A strategy must specify a side (home/away/favorite/underdog/over/under) and at least one additional real filtering condition — divisional game, roof type, surface, rest-day advantage, week range, spread range, or total-line range. A bare "always bet home" with zero conditions is rejected by the tool itself; that's a single-variable rule, not a strategy.
The validation split — real forward-walking, not a re-test on the same data
| Window | Seasons | Games | Role |
|---|---|---|---|
| In-sample | 1999–2018 | 5,300 | Where a strategy is discovered and its conditions are tuned |
| Forward-walk | 2019–2025 | 1,976 | Automatically re-tested here, unchanged, on data the strategy never touched |
Confidence weighting — the actual mechanism
Every result is reported as a real Wilson score 95% confidence interval, not a bare win percentage. A
strategy matching 8 games and going 6-2 produces a very different, much wider confidence band than one
matching 400 games at the same win rate — the interval makes that difference visible instead of hiding it
behind one number. Any result with fewer than 42 matched games (the same minimum-sample threshold used
elsewhere on this site, in exploratory_pass_42) is flagged directly as too small to trust, with
the wide interval shown rather than suppressed.
The real breakeven line
52.38% — the actual win rate required to profit against standard -110 odds. Every strategy result is plotted against this line, not a naive 50%, so "beating chance" and "beating the actual vig" aren't conflated.
What this tool does not claim
It does not claim any specific strategy built with it will be profitable going forward. It reports what a given rule set actually did, in-sample and out-of-sample, with honest confidence bounds attached. The forward-walk split exists specifically so a strategy that only worked by coincidence on the data it was tuned against gets caught, visibly, rather than presented as validated.