How the ratings work
No polls. No opinions.
The ratings only look at final scores and who each team played. Rankings from the AP poll, the playoff committee or anyone else play no part.
Scores and opponent strength
Every game says something like "Team A was 10 points better than Team B at Team A's stadium." We find the set of team numbers that best explains all of those results at once. Beating a strong team counts for more than beating a weak one, the way the old Sagarin ratings worked.
Last season as a starting point
Every team starts the year with part of last season's final rating, 50% in the NFL and 70% in college. That number fades as this season's games pile up. Replaying 2026 week by week, it cut the NFL's average miss from 10.4 to 9.1 points (91 games) and college's from 14.4 to 12.7 (471 games). It's a starting point that fades, not a poll: it comes only from last season's scores.
Home field
Half a season isn't enough games to measure home field, so each season's estimate starts from the long-run average and moves only when the season's results push it. In the NFL that average is 1.7 points: from 2015 to 2025 it ranged from about 0 (2020, no fans) to 2.7. Current estimates: 4.4 points in college and 0.6 in the NFL.
NFL: touchdown tiers, and blowouts get MORE credit
In the NFL we count the margin in touchdowns. Winning by 1 to 7 points is a one-touchdown win, 8 to 14 is two touchdowns, 15 to 21 is three, 22 to 28 is four, and 29 or more is five-plus. Every tier counts, and each extra touchdown tier is worth 2.5% more than the one before it (7, then 7.18, 7.35, 7.53, 7.7 points...). That way a big blowout gains extra power instead of being cut down. Inside a tier, a team gets partial credit, so a 10-point win counts for more than an 8-point win.
| Win by | TD tier | Counts as |
|---|---|---|
| 3 pts | 1 | 3.0 |
| 7 pts | 1 | 7.0 |
| 10 pts | 2 | 10.1 |
| 14 pts | 2 | 14.2 |
| 21 pts | 3 | 21.5 |
| 35 pts | 5+ | 36.8 |
We replayed the 2017-2024 NFL seasons week by week, rating teams only with games already played (2,111 games). The 2.5% escalation predicted best of the tier versions (10.38 points average miss vs 10.40 at 5% and 10.44 at 10%). A curve that shrinks blowouts did slightly better (10.32), but it breaks our rule that blowouts count, so we don't use it.
College: full margin, every point counts
In college the full final margin counts, with no cap and no diminishing returns, so a 49-point win counts as 49. We replayed ten college seasons (2016-2024, 5,347 games, each game predicted only from earlier ones). The full margin predicted better than every curve that shrinks big wins: an average miss of 12.8 points, against 13.5 for the old log curve. On 2025, held back until the end, it missed by 12.2 against 13.1. Replaying 2026, with last season as a fading starting point, it missed by 12.7 points against 14.3 for the old setup.
Conferences: no polls, only results
Polls play no part, anywhere. A conference's strength comes only from its teams' scores against teams from other conferences, plus how they did against the same opponents. Early in the season each team starts near its conference's average and moves away from it as its own results pile up. That conference average is also learned only from cross-conference scores. In our ten-season test a lighter version of this pull predicted best.
The FCS Test
An FBS team playing an FCS (smaller-division) team should win by at least 21 points, three touchdowns. Any shortfall is a penalty. Win by 14 and you fell 7 short, win by 3 and you fell 18 short, and a loss is the full shortfall times 1.5 (lose by 3 = 24 × 1.5 = 36 points short). Each shortfall point takes 0.05 off the team's Power rating, so a 3-point loss to an FCS team costs about 1.8 points of rating. Because conference ratings average their teams, these penalties also count against the conference. In our backtest the penalty cost a little accuracy (about 0.05 to 0.15 points per game), and we keep it because it's one of our rules.
Two ratings: Power and Résumé
Like Jeff Sagarin, who published a margin-based "Predictor" next to a wins-and-losses-only "ELO" rating, we publish two numbers. Power uses the margin curve above and is the one behind our lines, projections and offense/defense split. Résumé only looks at who you beat and who you lost to, adjusted for how good those opponents are. Margin and the order of games don't matter, and it's shown in the same point units. Power is for betting. Résumé is for ranking and playoff arguments.
Offense and defense
Each rating splits into an offense part (points scored versus expected) and a defense part (points allowed versus expected). Together they add up to the overall rating.
Our line: base plus adjustments
The base line is the difference between the two teams' Power ratings, plus home field. Then come the weekly adjustments. They move only our line and never change Power or Résumé. Each one has its own column on the Our Line vs Vegas page. An adjustment gets points only if it made our predictions better when we replayed 2016-2024 week by week. Anything that didn't still shows as a flag worth 0 points.
Quarterback. We read the ESPN injury report. If a team's starter (the QB with the most starts this season, or a starter we set by hand when injuries make that misleading, as with Lamar Jackson in Baltimore) is listed Out, Doubtful or on injured reserve, the team loses 3 to 7 points. It's 3 points plus 2 for every yard per pass the starter averages above his backup, capped at 7. We never guess at a status. Questionable counts as 0, and no listing means 0. In the history, a change at quarterback was worth about 2.7 points on average (850 games). It was the only adjustment that clearly helped: it cut the average miss from 10.27 to 10.20 points.
Recent form. We also built a version of Power where recent games count more (a game three to eight weeks old counts half). Adding any of it made predictions worse, so it's 0 points for now.
Yards per play. We built a rating from yards gained and allowed per play, adjusted for opponents, and turned it into points. Mixing in even 10% made predictions slightly worse, so it's 0 points.
Spots. A letdown (coming off a win over a top-10 or better team), a lookahead (a weak opponent now with a top-10 team next week) and a bounce-back (coming off a blowout loss). Each one was within noise in the history (letdown -0.2, bounce +0.3, lookahead -1.1 ± 0.8 points), so they show as flags only.
Travel and rest. Trips of 2,000+ miles, time zones crossed, West Coast teams in 1 PM ET starts, East Coast teams in late West Coast games, short weeks, extra rest after a bye, and back-to-back road games. All are measured from the ESPN venue and kickoff time, shown in ET. Every one was under about one point and within noise (time zones about -0.8 ± 0.5 per zone, a bye +0.4 ± 0.8), and together they made predictions worse. They're flags only.
When our line is 3+ points from the market, the game is flagged on the Our Line vs Vegas page.
Simulations
We simulate every remaining game thousands of times using the ratings. That gives projected records, division and playoff odds, #1 seed odds in the NFL and bowl odds in college. This week: 165 NFL and 322 college games simulated, from 107 NFL and 570 college finals.
Updated every Tuesday
After the weekend, the ratings rerun from the new scores and last week's lines get graded.
Limits
Early-season ratings are noisy, college has about 130 teams with uneven schedules, the market already prices in most of what we measure (our historical NFL lines beat the spread about 50% of the time, before the bookmaker's cut), and there's no historical college odds data to test college lines against the spread. Use it as one input, not a guarantee.