How NFL Picks & Ratings Work
How the BBMI NFL rating is built — and why, after treating the NFL as analytics-only, we now publish and grade NFL picks.
Why we publish NFL picks now
The NFL is the most efficient, most heavily-bet market in sports — and we treated it as analytics-only for exactly that reason: we wouldn’t publish a pick we couldn’t stand behind. That changed. Our current simulator, walk-forward tested across the 2023–2025 seasons, shows a durable edge on totals and positive closing-line value— so NFL is now a published, graded pick product.
We still show the work: every pick graded, wins and losses. Our edge is in totals and in beating the closing line — not in a raw against-the-spread win rate — so spread picks that carry more information risk are flagged, not hidden. See the public NFL record.
The BBMI NFL Rating
The BBMI NFL rating is a single opponent-adjusted efficiency number: how many points above or below average a team is on a neutral field. It is built in three steps:
- Play-by-play EPA — Expected Points Added on every play, from nflverse. EPA measures the value of each snap independent of the final score, stripping out turnover luck, garbage time, and field-position luck.
- Opponent adjustment (ridge regression) — every pass and rush in the league is solved simultaneously for each team’s offense and defense, so a unit is credited for who it played. An aggressive ridge penalty shrinks small samples toward average, so a few hot games don’t mint an elite rating.
- Market win-total prior — a de-vig consensus of multiple sportsbooks’ preseason win totals. Early in the season the prior carries most of the weight; as games accrue, observed opponent-adjusted EPA takes over (blend ≈ 5 / (5 + games)).
The blended rating is converted to points per game by a fitted scale, then feeds the game simulator. The rating updates weekly during the NFL season.
What is EPA?
EPA (Expected Points Added) measures the value of each play in terms of expected points. A first down on your own 20-yard line is worth about 0.5 expected points. A touchdown from the 5-yard line is worth about 6.3. Every play either adds or subtracts expected points based on the down, distance, yard line, and result.
EPA/play tells you how efficient a team is per snap. A team with +0.10 offensive EPA/play gains about 0.10 expected points more than average on every play — over 65 plays per game, that’s about 6.5 points better than average. It’s the most granular publicly available measure of team quality.
The inputs
Offensive EPA
Expected points added per play, adjusted for the quality of defenses faced. Passing and rushing are weighted separately.
Defensive EPA
The same, from the defense’s side — how many expected points a unit prevents versus an average offense.
Strength of schedule
Every game is weighted by opponent quality, so a strong record against a hard slate outranks one padded against weak teams.
Situational context
Home field, rest (short weeks, byes), and key-player injury impact adjust the projected margin for a given game.
From ratings to picks & projections
For any matchup, the opponent-adjusted rating gap plus home field and QB adjustments give a projected margin, total, and win probability. To project the full season — playoff odds, seeding, Super Bowl probabilities — the model simulates the remaining schedule 20,000 times, resolving tiebreakers and playing out the bracket in each run. Those simulations power the Season dashboard and Playoff Pulse.
Data sources
| Source | Data | Update |
|---|---|---|
| nflverse | Play-by-play with EPA, CPOE, WPA | Weekly |
| Odds API | Sportsbook spreads and totals — feeds the win-total prior and grades closing-line value | Pre-game |
| Open-Meteo / NWS | Game-day weather for outdoor stadiums (totals arm) | Game-day |
The record
On a 2023–2025 walk-forward, the totals model gets stronger the more it disagrees with the market: near breakeven at an edge of 2 points, then clearly past it as the edge grows — 53% at edge ≥ 3, 57% at edge ≥ 4, and 60% at edge ≥ 6. Both markets also show positive closing-line value: spread +1.25 points (t = 5.2) and totals +0.66 (t = 3.2) — the strongest evidence a model is genuinely ahead of the market. Straight-up winner accuracy is 66.7% across the graded set.
Totals is the flagship pick product. Spread picks are published with an edge cap — above the threshold they’re flagged “high-edge, information risk” (backup QBs, resting starters, week-18 seeding) and shown, not deleted. The NFL does not have a raw against-the-spread edge and we don’t claim one; the validated signal is totals and closing-line value. Everything is graded publicly, wins and losses, on the NFL record.
NFL vs. our other sports
BBMI’s NCAA basketball, football, baseball, and MLB models earn their edge in thin markets — 130+ teams, less sharp money, and information asymmetry a model can exploit across the whole slate. The NFL is different: 32 teams and the sharpest lines in sports, so the edge is narrower and specific — totals and closing-line value, not a broad against-the-spread advantage. Different markets require different honest assessments.