What the model uses

Every prediction starts from a blend of preseason and in-season signal. The preseason side leans on Elo, SP+, FPI, Talent Composite, and returning-production/transfer-portal/recruiting data — each smoothed with a recency-weighted blend across recent seasons rather than a single-year snapshot, since any one preseason rating is a noisy read on its own. The in-season side is a set of learned team ratings, one each for EPA per play, success rate, explosiveness, rushing, passing and line play, on both offense and defense. They're fit on every play so far and adjusted for who each team has played, and each one starts from a preseason estimate (last season, returning production including at quarterback, portal moves by position, coaching changes) that the season's results pull away from, faster the more a team plays. So a team's rating shifts as its season actually unfolds rather than staying anchored to where it started in August.

How it predicts

Two models: point margin and combined total points. Margin and total are fit with linear regression over the standardized feature set above (the margin model on the NIL/transfer-portal era, 2021 on, which tested more accurate than training on older seasons) — kept deliberately simple, since more complex versions were tested empirically and didn't hold up out of sample. Win probability comes straight from the predicted margin, calibrated on past games, so a team's win chance, its predicted score and our straight-up pick always agree. The one exception is Moneyline Value: those underdog picks come from a separate model tuned specifically for spotting live underdogs, which tested far better for that job; it never changes the win probabilities or scores shown anywhere else.

It keeps learning, every week

After every week of games, the model checks itself against what actually happened and adjusts: how much home field is really worth right now, whether its totals have been running high or low, and which kinds of spread picks have been winning. It only ever learns from games already played, and the track record is graded the same way, one week at a time.

All Picks vs. Best Bets

• All Picks: every spread and total where our number beats the line by the minimum edge. This is the full record of what we recommend.
• Best Bets: a featured shortlist of six picks each week: the three spread picks the model rates most likely to win, and the three totals with the biggest edge.

How it's tested

The numbers on the Track Record tab come from a genuine blind holdout. Every completed season since 2021 is graded by the same model retrained from scratch without that season, so it never saw a single game it's being graded on. The current season is graded by the live model, which only trains on completed seasons. That's the same discipline as testing a forecaster by asking "how did you do on the games you hadn't seen the result of yet," not "how did you do on games you already knew the answer to."

• Blind, not in-sample. A model's fit to its own training data is close to meaningless on its own — this only reports performance on seasons excluded from training.
• Graded at the threshold we actually bet at. Picks only count once the model's edge against the market clears a fixed minimum — the same bar used for real recommendations. Every season shown is graded under today's rules; we raised the spread and total minimums in October 2026 after testing them on past seasons.
• Every graded pick is counted. No excluding cold streaks, no reshuffling the pick list after the fact.

We'd rather show a thinner, honest number than a rounder, flattering one. If a season goes cold, it stays in the record.

Two different ATS numbers

• Spread Picks (ATS) card: only the games we actually pick — where our edge against the spread clears the minimum. This is the record of our recommendations.
• All-Games KPIs: a pick on every game with a line, no edge required, including the coin flips. Usually lower, and the number that compares directly with public model leaderboards.

All-Games KPI definitions

• SUP / ATS: straight-up winner and against the spread, every lined game (pushes excluded).
• wSUP, pSUP, wATS, pATS: the same, weighted by the size of our projected margin (w) or of the actual result (p).
• MAE / DAE: mean / median points off the final margin. RSE: average squared error (the source's own formula).
• GRD: one composite grade, higher is better: [(2·SUP + wSUP + pSUP + 2·ATS + wATS + pATS) / 8] − [(2·MAE + √RSE + DAE) / 4].

Definitions match CFB Picker's glossary so the numbers compare directly.