What Is the Playoff Predictor?
The Playoff Predictor answers the questions a single-game forecast can't: who makes the playoffs, who wins the division, who gets the bye, and who wins the Super Bowl.
From any week of the season, it prices every remaining regular-season game with RICE and plays the rest of the season out 2,000 times (10,000 for the current week). It seeds each simulated season with the NFL's full tiebreakers and works out every possible bracket exactly. For each team, you get its expected wins and its chances of making the playoffs, winning its division, earning a bye, reaching the conference final, winning the conference, and winning it all.
Why a Simulation, Instead of ___?
1. Games in a Season Aren't Independent
The simplest playoff model takes each game's win probability and treats every game as its own coin flip. If a team is better than we think, though, it's better in all of its remaining games at once. Coin-flip models underrate how often a team runs away with a season, or falls apart.
The predictor handles this dependence the way FiveThirtyEight's "hot" simulation did. Each simulated season draws one strength shock per team and applies it to every one of that team's remaining games. In some simulated seasons a team is hot, and in others it's cold. The shocks spread the outcomes the way real seasons do.
2. The Tiebreakers Are the Real Ones
Seeding in the NFL often comes down to tiebreakers, and a mistake in them throws off every playoff probability. The predictor applies the league's full tiebreaker procedure, with each era's own steps (strength of victory, for example, counts only from 2002). Checked against history, it reproduces 444 of 444 seeds from 1990 to 2025, apart from two the league broke on points scored.
3. Each Game Is Priced as RICE Prices It
The predictor prices every remaining game the way RICE prices a game weeks away. It uses the quarterback and lineup the injury predictor expects for that week, instead of today's injuries frozen in place. Week 18 games include each team's chance of having nothing left to play for. Each team's playoff strength also counts the injury risk it faces between now and January.
What It Learns
The predictor fits two numbers, each only on seasons before the one predicted. It fits them separately for early (weeks 0-6), middle (7-11), and late (12 on) in the season:
- How sharp RICE's prices are for games weeks away. This number comes out close to one, so RICE's probabilities at a given week are about right even for games well in the future.
- How much of a team's true strength is still unknown. It's about 0.6 early in the season, 0.5 in the middle, and 0.4 late, in log-odds. The uncertainty shrinks as results accumulate, which is what you'd hope.
How Well It Does
The test scores the predictor with log loss at every week of every season from 2002. The first benchmark plays the same RICE prices out as independent coin flips:
| Outcome | 2002-19 | 2020-25 |
|---|---|---|
| Making the playoffs | +0.006 (−0.002 to +0.014) | +0.017 (+0.002 to +0.038) |
| Winning the division | +0.000 | +0.013 (−0.008 to +0.047) |
| Winning the Super Bowl | +0.002 (−0.001 to +0.007) | +0.002 (−0.001 to +0.006) |
Positive means the predictor is better. I measured these numbers before the October 2026 switch described under What Didn't Work, which moved them by a few thousandths. The shared strength shock helps everywhere, and clearly so on the 2020-25 playoff chances.
Calibration is the predictor's strongest result. When it says 30 percent, about 30 percent of those teams should make it. Its calibration error on playoff chances is 0.012 on 2002-19, against 0.036 for the coin-flip version. On 2020-25 it's 0.020, against 0.045.
The simplest forecaster you could build uses only each team's win percentage. The predictor beats it on every outcome in both windows, by 0.038 in log loss on the 2002-19 playoff chances (0.025 to 0.052).
What Didn't Work
- A different pricing of each game. Until October 2026, each remaining game was priced from the strength each team carried into its next game. Switching to RICE's later-game pricing with the injury predictor's lineups made no difference on any outcome (a few thousandths either way). I made the switch for consistency with the rest of the site, and it didn't win.
- A separate projection model. An earlier design projected each week from the last, as its own model. It did no better than RICE's pricing, and I deleted it.
Notes
- The predictor is a view of the forecast and never an input to it. No rating reads it, so nothing on the playoff page changes a team's rating. RICE borrows from it in one place: week-18 rest (see Final-Week Rest on the RICE page).