Methodology

A win-probability curve fit on 72,809 Division I games from 2014 through 2024, anchored to current PEAR Ratings. Each rating point of difference moves the log-odds of winning by 0.37 (95% interval 0.364 to 0.379). Home field is worth 0.23 on the same scale, roughly a 6% bump (0.217 to 0.249). Per-game random rating noise has a standard deviation of about 2.0, which captures the variance you get from starting-pitcher identity. Fatigue costs 0.40 rating points per game already played in a regional or super series. Bracket sanity check: across 11 historical NCAA tournaments the host-to-super advancement rate was 61.3%, which lines up with these constants better than the 25-year literature anchor of 68.8%.

This was the 2025-26 playoff model. It is not published for the 2026-27 season; its last saved run is kept at /nba/2026-playoffs. Each team carried a regular-season Elo and a playoff-only Elo; the two were blended for postseason matchups. In-series state (1-0, 2-1, etc.) modifies win probability on each subsequent game because trailing teams play differently. Rotation ratings shifted live for injuries. The bias correction is tuned against 2024 and 2025 playoff outcomes, 166 games in all. The playoff-totals correction was refit every night from new results.

A 12-team College Football Playoff simulator. Seeding is learned from 2014–2024 CFP history, so the model picks the kind of teams the committee actually picks. First-round games are on campus, with home-field worth 2.5 points; QFs, SFs, and the final are at neutral sites. The projected bracket picks the favorite in every game. The sampled bracket adds random per-game noise, sized from the ratings, so a single sampled run has upsets in it.

The model behind every NFL projection and the site's NFL team rating. A team rating that leaves the quarterback out updates after each game from the margin of victory, in the FiveThirtyEight style. A separate rating for each quarterback travels with the player and updates from his own passing, accuracy and rushing numbers, adjusted for the defenses he faced. The two add up to the team's rating for a game, and a last layer adjusts for rest, weather, primetime, division games, home field and injuries. On 2025 games held out of training, its Brier score was 0.2191, 0.0029 lower (better) than FiveThirtyEight's published 2025 number of 0.222. The model does not beat the betting market. Full methodology, calibration, and what the May 2026 rebuild changed, at /methodology/nfl-v6.

AUDIBLE ranks players at six positions (QB, RB, WR, TE, offensive line and defense) and builds a team composite from eight position-group ratings, with weights fit by regression. Every ranking is shown next to a simple baseline, each player's average over his last eight games, so you can see where the model adds anything. Two of six positions clear that bar: quarterbacks and tight ends (per game). The other four are descriptive rankings, not forecasts. The team composite is not the site's NFL team rating; that is the game model's. Read the full breakdown at /methodology/audible and see the live table at /nfl/qb.

Before the season, each team's rating starts from last year's and is adjusted for a new quarterback or coach. That preseason rating produced the frozen win totals, and we checked it against DraftKings' 2026 win totals: 25 of 32 teams landed within 1.5 wins of the market, with a median miss of about 0.8 wins. That is a check that the ratings agree with the market, not a betting signal. The team-by-team table is public on the win totals page. During the season the site ranks NFL teams by the game model's rating (above), which updates with each week's results; see NFL team ratings.

For every game on the board we publish a band of plausible final scores with the market's posted total on the same axis, and no over/under lean. The claim is coverage and it is checkable: held out, our 50/80/90% bands covered 50.1/81.1/89.9% of NFL games and 50.5/81.1/91.4% of FBS-vs-FBS college games. We withhold a side because we measured what our number adds to the market's and it is approximately nothing: on held-out games the best blend gives ours a weight of 0.05 to 0.10, for an improvement too small to matter. The earlier version of this that did publish a lean was reproducible 78% of the time by a rule that never looked at the model. Full reasoning, including the one probability we could compute and deliberately do not publish, at /methodology/totals.

Retired on August 28, 2026. The model combined forecasts from NOAA, Open-Meteo, GFS and ECMWF, among others, corrected each one for its known bias in each city, and turned the result into probabilities for bands of the daily high and low. After a 12,000-forecast backtest showed the temperature model was overconfident above 30%, its confidence was capped there. The paper record showed an edge, but at prices almost nobody would trade; the closing note is on the calibration page.

Every published rating gets a calibration check: we bin outcomes by predicted probability and see whether the fit is y = x. Where it is not, that is a model bias we either correct or report. The results are on the calibration page, model by model.