The projection

For each team we take last season's points (regressed toward the league mean, because small samples lie), combine it with the opponent's points allowed, then adjust for:

• Pace — combined tempo of both offenses → more or fewer possessions.
• Scheme — the coordinator's family (Reid, Shanahan, McVay, Air Raid…) carries a scoring bias.
• 2026 changes — a manual nudge for a new QB, coordinator, or major roster move.
• Home field — worth about 1.3 points.

That yields projected points for each side → a total, a spread, and a win probability.

The edge

We never copy Vegas — we build our number first, then compare it to every book you enter:

• Total edge flagged when our total differs from the book by 2.5+ points.
• Spread edge flagged at 2.0+ points of disagreement.
• Moneyline value flagged when our win probability beats the book's implied odds by 6%+.

With 10 books offering different numbers, the edge is finding the one that's softest on a given side.

Honest expectations (what we actually measured)

We backtested this on 6,700 real games (2000–2025) with closing lines. The verdict: our model picks straight-up winners at ~60% (better than a coin flip, below the ~66% you'd get just betting the favorite), and hits ~50% against the closing spread — i.e., no edge vs the sharpest number. Swapping in EPA didn't change it: the closing line already prices team efficiency. So we don't pretend to beat it. The real edges live in player props (softer markets, projectable volume) and line-shopping soft/early numbers across books. Anyone promising 70% ATS is selling you something.

What makes it independent

Most public models are just re-derived from last year's box scores — which means they echo the same info the books already priced in. Our coaching-tree/scheme layer gives us a view the market underweights early in the season, when personnel changed but the scheme lineage predicts how a new staff will actually play.