The model, in plain English
How the model works
Every prediction on this site comes from the same short recipe, and you can audit it on every matchup page. This page walks through the recipe once, slowly, with no jargon. If you want the constants and the tuning story, that's the methodology page.
Step 1
One number per team
Each team gets a single Elo rating — one number that says how good it is. Average is 1500. Above it, better than a typical NFL team; below it, worse. Roughly speaking, 1600 is a contender and 1400 is a team picking early in the draft.
Ratings move after every game: the winner takes points from the loser, and beating a strong team is worth more than beating a weak one. How fast they move is set by the K-factor (here K = 28, tuned on 2025 results), and big wins move ratings a little more than nail-biters. Your team's 2026 starting rating is a blend of its 2025 Elo, its scoring strength, and its late-season form — the full audit trail is on every team page.
Example. Seattle at 1650 hosts New England at 1550. Before anything else is considered, the model sees a 100-point gap in Seattle's favour.
Step 2
Home field is worth real points
Playing at home helps, so the home team's rating gets a bonus before any probability is computed: +25 points of home field advantage. That number isn't a textbook default — it was fitted to the 2025 season, when home teams went 142-123 and the classic 55-point figure turned out to overprice the host. At a neutral site the bonus is zero.
Example, continued. Seattle's effective rating for this game: 1650 + 25 = 1675. The gap over New England grows from 100 to 125 points. If the game were in London or at a neutral venue, it would stay at 100.
Step 3
Rest matters, a little
A team coming off a bye or a Thursday game has had more recovery time than one on a short week. The model prices that rest edge at 5 points per extra day of rest, capped at ±25 points so it can nudge a game but never decide one. Week 1 has no rest adjustment at all — every team is coming off the same offseason.
Step 4
One curve turns the gap into a probability
Add it up: home rating, plus home field, plus rest, minus away rating. That final number is the rating edge, Δ. The win probability is a single logistic curve of Δ:
P(home wins) = 1 / (1 + 10^(−Δ / 400))That's the entire formula — there is no hidden layer after it. The curve is deliberately smooth: a few extra rating points near even barely move the probability, and once a team is a huge favourite, more points move it less and less, because no NFL game is ever truly certain. The same Δ divided by 25 gives the model spread in points.
| Rating edge | Win probability | Roughly |
|---|---|---|
| Even | 50.0% | A true coin flip |
| +25 points | 53.6% | Wins about 4 of every 7 |
| +50 points | 57.1% | A field-goal favourite |
| +75 points | 60.6% | Wins 3 of every 5 |
| +100 points | 64.0% | Wins almost 2 of every 3 |
| +150 points | 70.3% | Wins 7 of every 10 |
| +200 points | 76.0% | Wins 3 of every 4 |
| +300 points | 84.9% | A genuine mismatch |
Example, finished. Δ = 1675 − 1550 = 125 points, so P(Seattle) = 1 / (1 + 10^(−125/400)) ≈ 67%, and the model spread is 5 points in Seattle's favour. Every matchup page shows this exact arithmetic in its own calculation ledger.
See it on a real game
Every game on the schedule links to a matchup page with the actual ratings, each adjustment as its own line, and the same logistic math done in front of you.