--- title: "How the Matchup Page Works | Held Out" description: "Where the win probabilities, projected scores, and margins on the matchup page come from." url: "https://heldout.org/methodology/matchups" --- # How the Matchup Page Works ## What's on the Matchups Page? The matchups page puts RICE to use on a Sunday. Every game in the season gets the home team's win probability, a projected score, a spread, and a range of likely margins. A played game shows the result next to what RICE expected. A game further out shows the lineup RICE expects, so a week-12 game can look different in October than it did in September. Everything on the page is display only. Nothing here feeds back into the ratings. ## What Each Card Shows #### 1\. The Price The win probability is RICE's, after one small calibration step (see Are the Probabilities Right?). The projected score and margin range come from a score model built on RICE's price. That model is calibrated on the stadium's roof and on the forecast wind and temperature at kickoff. Wind takes points off the total, and cold matters much less. Neither one changes who's favored. Games in the current week are priced with the real injury report. Games further out use the injury predictor's expected lineups (see Who Will Play). RICE-playoffs, the January version of the model, prices the playoff games. #### 2\. Injuries That Move the Line Each upcoming game shows up to three players a side whose status moves the price, ranked by how much each one changes his team's win probability. A quarterback is priced as a swap: the projected starter against the player the model expects to start if he can't. #### 3\. What a Game Did For a played game, the card shows the game's impact: how much the result moved each team's O-, D-, and ST-RICE, and RICE overall. It also shows each starting quarterback's Game BRADY, which grades how he played that day from that game's plays alone. Beside it is his career BRADY going in and coming out. That comparison is the Ability Update tab, and it shows the point I make on the BRADY page: one bad game moves the career rating much less than it moves the conversation. #### 4\. Live Win Probability During a game, the win probability updates play by play. It starts from RICE's pregame price and updates on the state of the game. In regulation, the model treats the final margin as a bell curve centered on RICE's expected margin. The game state tells it how much of that curve is still in play. Finished games get the same curve, rebuilt from the play-by-play. I started with a simpler version. Over whole games on 2003-19, this one beats it by 0.0010 in log loss (0.0005 to 0.0015). It's also ahead on 2020-25, though that difference (0.0001) is too small to resolve. The browser and the live feed run the same code, checked against the original prototype to six decimal places. ## Are the Probabilities Right? A probability should mean what it says. Sort every game by its home win probability into twenty equal groups, and compare how often the home team won in each group. A well-calibrated forecast lands close to the line, though not on it. Even a perfect forecaster shows a gap on a few thousand games, from chance alone: about 0.023 on 2000-19, and about 0.040 on the smaller 2020-25 sample. Read the model against that bar. | Seasons | Shown probabilities | A perfect forecaster | | ------- | ------------------- | -------------------- | | 2000-19 | 0.026 | 0.023 | | 2020-25 | 0.045 | 0.040 | On 2000-19, the shown probabilities pass the test I set for them in advance. On 2020-25 they narrowly miss: three of the twenty groups fall outside their bands, where two would pass. RICE's raw probabilities, before the calibration step, do better there (0.037). I kept the step because it passed the pre-registered test, and I'd rather tell you that than quietly switch. ## The Market Toggle A switch, off by default, shows a second forecast built on the betting market. This toggle is the only place on the site that reads a betting line. RICE, BRADY, and every rating never see one. The forecast averages eight small models, each anchored on the closing line and shrunk hard toward it. Some adjust the line with RICE's probability. Some learn small corrections from RICE's experts, from BRADY's reading of a starter change, and from the records the market tends to overweight (one-score games, fumble luck, and turnover margin). One blends in the spread. With only about 270 games a season, any model that strays far from the line overfits, so each one stays close to it. Walked forward from 2009, the forecast beats the strictest reading of the closing line (its midpoint) by 0.0016 in log loss a game (0.0005 to 0.0026), in 13 of 17 seasons. On 2020-25 it beats the line by 0.0017 (0.0000 to 0.0034). That's a small edge on the best public forecast there is, which is about what you should expect. For margins, nothing I tried beat the spread itself, so under the toggle, the margin shown is the spread. ## What Didn't Work - **A more flexible calibration curve.** Fitting a free-form curve to RICE's probabilities lost clearly on 2000-19, by 0.0067 (0.0015 to 0.0133). The simple version stayed. - **More complex market blends.** A free-form regression and decision trees both lost to the plain average, by 0.001 to 0.008. Theme Updated Nov 10, 2026, 9:37 PM PST © 2026, Aditya Kishore. All rights reserved. For Commercial Licensing or Questions: [Contact Held Out](/contact) Club marks are trademarks of their clubs. Past marks: [SportsLogoHistory.com](https://sportslogohistory.com/nfl-primary-logo/).