How the playoff odds work

The simulation

• Each run plays every remaining game 100,000 times, then the playoffs. Ties happen about as often as in real games with similar lines (about one game in 300).
• Every simulated season is ranked with the full tiebreaking procedure: head-to-head, division and conference records, common games, strength of victory and of schedule, points rankings, net points, net touchdowns and the coin toss. Our standings engine reproduces the official division ranks (768 of 768) and every playoff seed of every season since 2002.
• Seven teams per conference make the playoffs: four division winners and three wild cards. Only the No. 1 seed has a bye; the bracket reseeds after the wild-card round.
• The odds are rerun after every game day, as soon as the final scores are published (nflverse), and the pages rebuilt.

How good the teams are

• Team strength comes from the betting market: a rating for each team fitted to the closing point spreads of every game since 2002, moved by each result (a result better than the spread raises the rating a little, as the market does).
• For the games of this week and next, the market’s posted lines are the starting point.
• Uncertainty is measured, not assumed. Nobody knows exactly how good a team is, and teams change. Each simulated season draws team strengths from that uncertainty and lets them drift from week to week, by the amounts measured on 2002–2017: how far a rating fitted today misses the lines of games one to eight weeks later. That is why a team’s odds are less extreme than its rating alone would suggest.
• Within a simulated season, results feed back into the ratings, as they do in reality: a team that starts winning in a simulation is rated higher for its later games in that simulation.
• Quarterbacks: when a starting quarterback goes on the reserve list, the team’s rating for games two or more weeks ahead drops by an amount based on the gap between him and his backup (the posted lines for the next two weeks already include it).
• Resting starters in week 18: a team whose seed is already settled before the last week plays its final game about 8 points worse than its rating (measured from the closing lines of 2002–2017).
• Other team numbers (expected points per play, points per drive, scoring margin, close-game and fumble luck, Elo) added almost nothing once the market’s view was known, so they are left out.

How good the odds were

We built and tuned the model on 2002–2017 only, then scored 2018–2025 once, as of every week of each season: 4,512 team-weeks. The dots should sit on the dashed line: when the model said 70%, teams should have made the playoffs about 70% of the time. Two versions were scored; the one used here did better on 2018–2025 than the version tuned further on 2002–2017, so its score below is a little flattering. The forecasts made during the 2026 season are the clean test.

Of the 970 forecasts of 99% or more (or 1% or less), none was wrong.

Against simpler methods

Brier score on 2018–2025: the average squared error of the probabilities (lower is better; always saying 50% scores 0.25).

Games that matter

• For every remaining game, the season is simulated again twice: once with the home team winning that game and once with the away team winning it, 20,000 times each, with the same random draws. The difference in a team’s chances is what the game itself means to it, including games between two other teams.
• We measure it that way rather than by sorting the main simulations by the game’s result. For a game weeks away, the simulated seasons in which a team won it are also the ones in which it was stronger, so sorting would credit the game with what its other games did: for the biggest game of week 18, about half again as much.
• “Playoff spots at stake” is half the sum of those differences over all 32 teams: the share of a playoff spot that moves from some teams to others.
• Because both results use the same random draws, the difference is precise: its simulation error is about a tenth of a point for most teams and at most about half a point. The pages show a team only when its swing is at least a point and at least twice that error.

Every forecast is kept

Each run’s inputs, settings and odds go into a forecast ledger that is never edited, and at the end of the season it will be scored the same way as the test above. From week 10 of 2026 on, every forecast is made before its games are played. Weeks 1–9 were run afterwards, with only the data available at the time.

What the model doesn’t know

• Injuries other than a starting quarterback on the reserve list, except through the market’s posted lines for the next two weeks.
• Coaching changes, trades and motivation beyond the week-18 resting rule.
• Exact clinching is separate from the odds: a team at “>99%” has not necessarily clinched; the marks x, y, z and e are the proofs (below).

Clinched and eliminated

• Simulations can’t prove a clinch: “in 100,000 of 100,000” is not “clinched”. So clinches and eliminations are proven separately, from the results so far, and only proven ones are marked: x a playoff spot, y the division, z the No. 1 seed, e eliminated.
• Every week: what the win totals alone settle, whatever the tiebreakers do (for example, a team that would still finish ahead of every division rival if it lost all its games and they won all theirs has clinched the division). These proofs never claim too much, but they miss clinches that rest on tiebreakers, so before the last week a mark can come a little later than the NFL announces it.
• In the last week: every combination of the remaining results (win, loss or tie; up to 3¹⁶, about 43 million) is run through the standings engine. Scores are left free, so a tie that only points could break is tried in every order. A mark then means no combination of results can change it.
• Tested on every season from 2002 to 2025, as of every week from week 6: each of the 7,203 win-total proofs agrees with how the season ended. The full last-week check was run on 2024 and 2025 (43 million combinations each, about six minutes): every label agrees with how the season ended and holds in 2,000 random finishes played with realistic scores and every tiebreaker, and the 2024 labels match that season’s picture before week 18 (Kansas City had the No. 1 seed; the NFC North and South and the AFC North were still open).

How often a tiebreaker decides

• In each simulated season, a team’s place is “decided by a tiebreaker” when it finishes level on record with a team on the other side of the playoff line (or of the division title): one gets in, the other doesn’t. The step of the NFL’s procedure that separated the two is counted too.
• Across all teams this season, the deciding steps are mostly head-to-head, conference record and division record, then common games; strength of victory and schedule rarely, points almost never. A division winner and a team from another division are never compared (division winners get in on their own), so those don’t count.

Paths and the No. 1 pick

• Paths to the playoffs (on each team’s page): how often the team finishes with each number of wins and its playoff chance with that record; for the records in between, the other teams’ games that change that chance most, each played with each result forced (20,000 seasons each way), shown only when the difference is at least five points and three times its simulation error.
• The No. 1 pick: the draft order follows the NFL’s rules: teams out of the playoffs first, the worst record first; ties go to the weaker schedule, then the division or conference tiebreakers, then the steps between conferences. Our engine’s draft order matches nflseedR’s on every real season since 2002 without a coin toss.

Data

Schedules, scores, and the closing and posted point spreads: nflverse (nflreadr), CC BY 4.0. Rosters for the quarterback rule: nflverse.