How RICE Works

What Is RICE?

RICE (Recursive Indicator of Competitive Edge) is Held Out's team rating, and the model everything else on this site eventually feeds. Every team gets a rating after every game, and the gap between two teams' ratings is a win probability for the game between them.

Plenty of team ratings describe last season well and forecast next Sunday badly. RICE is graded on how well it prices games before they're played. Every piece of it had to improve those forecasts to stay, and most smart-sounding ideas didn't. They're in a long graveyard (see What Didn't Work).

The scale is the one BRADY uses. A 75 is the median team-season, and the best team-season on record, the 2007 Patriots, is a 99.2. Each season is re-centered so the average team sits in the same place, so a 1985 team and a 2025 team can be compared directly. Underneath, a rating is an expected margin: a team at +7 is about a touchdown better than an average opponent on a neutral field.

The RICE Family

Held Out publishes a few views of the same model:

• RICE: the team's overall rating, and the number that turns into a win probability.
• O-RICE, D-RICE, and ST-RICE: the offense, defense, and special teams separately. Each is graded on its own scale, with its median at 75 and its best unit at 98: the 2022 Chiefs' offense, the 2006 Ravens' defense, and the 2006 Bears' special teams. The three grades don't add up to RICE, but the expected margins underneath them do.
• RICE-playoffs: the version that prices January (see The Playoffs).

One display detail is easy to miss. The rating a team carries after a game counts next week's injuries and ignores its rest. A team that sits its starters in week 18 shouldn't look several points worse going into the playoffs. Before this rule, the 2007 Colts read 88.5, then 82.3 after resting everyone in the finale, then 94.6 after their first playoff game. The weekly points on the team charts are pinned at the final whistle, so a later injury report never rewrites a week that's already been played.

Why RICE, Instead of ___?

1. Cost & Thoroughness

RICE is free for non-commercial use, and it rates every team after every game back to 1977. If you want to compare the 1985 Bears' defense with the 2000 Ravens' week by week, it's all here.

2. It Knows Who's Playing

Knowing who plays is RICE's biggest advantage. Record, point differential, and Elo all treat a team as one number. When a starting quarterback is carried off the field, they keep rating the team as if he's still there. When he comes back three weeks later, they rate the team on what his backup did.

RICE keeps a team's strength net of who played. It reads each game alongside the quarterback and the lineup that were on the field. Each forecast adds back the quarterback and lineup expected for the game being priced. A team that won three straight with its starter hurt doesn't get credit for strength it didn't have. It also isn't charged twice for an injury it has already played through.

3. Performance

RICE answers to four benchmarks, in this order of priority:

• LossBench: how well RICE prices games, from one game out to twelve games out. Every game is priced from several points earlier in the season: 1, 2, 3, 5, 8, and 12 games before it, weighted toward the nearer ones. Each price uses only what was known at the time and is scored with log loss.
• MarginBench: how well RICE's expected margins predict the final score, against a simple baseline of each team's point differential so far plus home field.
• PlayoffBench: every playoff game, the Super Bowl included, priced from the end of the regular season.
• MarketBench: LossBench's games that have a closing betting line, scored against the line.

Log loss scores each predicted probability by how much of it went to the winner, averaged over games. Lower is better, and a coin flip scores 0.693. The whole range between "home field only" and the closing betting line is only about 0.076 wide, so differences in the third decimal are real.

This table compares RICE with the simple forecasts most people would build, on the same games and from the same points in the season:

The regular-season columns are LossBench, and the playoff columns are priced from the end of the regular season. Elo is the standard public version, with a 65-point home edge and a third of each rating reverting every offseason. Nothing in it is fitted.

RICE leads every column. Elo is the closest competitor. In the 2001-19 regular season the two intervals overlap, so I'd call RICE's edge there real but modest. Record and point differential trail everywhere.

On margins, the gap is bigger than you might expect. One game out, RICE's expected margin misses the final score by a mean squared error of 193.6 points², against 289.6 for point differential plus home field.

The closing line is ahead of RICE: by 0.040 in log loss on 2001-19 (0.026 to 0.053), and by 0.055 on 2020-25 (0.035 to 0.076). The comparison is lopsided by design. The line is read at kickoff, after a week of injury news and lineup leaks, while most of LossBench's prices were made weeks earlier. RICE never reads the market, which is a benchmark and never an input.

RICE's Modeling Approach

RICE is four "experts," each a rating of one part of a team. A model that learns how much to trust each one combines them into one probability.

1. Team Strength, by Unit

Each team has three running strengths: offense, defense, and special teams, all measured in PIV (Held Out's play-value metric; see the PIV page). After each game, each strength moves by how far the result surprised the model, scaled by how much the model trusts one game. One game is one measurement, with no extra week-to-week drift and no extra noise on the margin. I tested both additions, and the simpler structure won. Over an offseason, each strength regresses part of the way to the league average.

Two smaller corrections apply too. The margin a team is judged on has its fumble-recovery luck taken out, because who recovers is mostly a coin flip. Each fumble is charged its average value, whoever fell on the ball. The 2020 games, played without crowds, are all treated as neutral sites.

The strengths are kept net of who played. Each game's measurement includes the starting quarterback's BRADY and the lineup's player values, and the model learns each one's weight as it goes. In 2004, one standard deviation of starting quarterback was worth about 0.008 points per play to an offense. By 2024 it was 0.024. That trend shows up all over RICE.

2. The Quarterback

The starting quarterback's BRADY is its own expert (see the BRADY page). It's the projected starter's rating as of the week before. A quarterback taking over mid-season gets a small fitted adjustment for his first few starts. He's also priced with the men around him, so his receivers' and backs' ratings come along with him.

For a game further out, RICE uses the quarterback it expects to start that game, who isn't always this week's starter.

3. Who's Playing

The roster expert prices everyone who isn't the quarterback. It has three pieces:

• The snaps that belong to players who won't play: inactives, injured reserve, and each injury-report designation at the rate that designation keeps players out.
• How far the expected lineup is from the lineups the team's strength has already absorbed.
• The value of each missing player against the player taking his snaps.

It's one expert with fitted weights on each piece, and it has its own page, Who Will Play.

4. Final-Week Rest

In the last week of the regular season, a team whose playoff seed can't change often rests its starters, and it loses far more than its rating predicts. That one week once accounted for nearly half of RICE's gap to the betting line. RICE reads the standings with the NFL's full tiebreakers, works out which teams can no longer move, and discounts them.

For a week-18 game priced from earlier in the season, RICE can't know yet who'll be locked in. It uses the Playoff Predictor's simulated seasons to estimate each team's chance of having nothing to play for.

5. Combining the Experts

A logistic regression combines the experts. It's refitted before each season on the seasons before it, so no season is priced with knowledge of its own results. Recent seasons count more: a season nine years back counts half as much as last season. What predicts NFL games has moved (BRADY's weight has about doubled since 2000), and a model that weighs 2002 like 2024 lags behind.

A second, smaller correction for the quarterback sits on top. Out of sample, the combiner tends to under-weight BRADY for two reasons. A weight learned from the past lags the position's growing importance. And a team strength learned from results absorbs some of the quarterbacks who produced them. So for each team, RICE compares the quarterback it expects with the last 24 who played, and adjusts the probability by the difference. The correction applies only in the regular season. Tried in January, it made the playoff forecasts worse.

6. Games Further Out

A game three weeks away has no injury report yet. RICE doesn't freeze today's injuries in place or assume everyone is healthy. It uses an injury predictor that estimates each player's chance of playing each remaining game, and prices the later game with that expected lineup. Those later games get their own combiner, trained on exactly that kind of projected row.

7. The Playoffs

RICE-playoffs prices January with the same experts and the same combiner, plus two changes. The offensive strength has a separate term for each quarterback that travels with him, so a playoff team is rated net of its starter's own history. And BRADY's weight follows its trend through the season. Against RICE without these changes, the champion's log loss improves by 0.031 on 2000-19 and 0.057 on 2020-25.

The playoffs have been the hardest part of this project. I tried many January-specific corrections: recalibration, a separate playoff combiner, byes, rest, form, seeds, and a lower BRADY weight late in the year. Almost every one won on 2000-19 and then reversed on 2020-25. RICE's playoff calibration drifts by era, so a correction learned from old Januaries mostly reflects the last era.

What Didn't Work

These are a few entries from the graveyard, whose full list is long:

• Elo as an input. Once the strengths treat one game as one measurement, Elo has nothing left to add. Putting it back cost 0.0008.
• A second model blended in. A generative version of RICE ran alongside the combiner for about a month. It didn't improve the forecasts enough to keep, and there's one model again.
• More complex combining. I tried weights that move by week, by distance to the game, or by how much evidence stands behind each rating; boosted residuals; and a separate combiner for each part of the season. All were null. The only combiner changes that ever worked were training it from week 1 and weighting recent seasons more. Every other gain came from better experts.
• Box-score lineups. Pricing the lineup by box-score quality, experience, age, draft position, or the strength of a player's past teams measured nothing beyond snap shares.
• Preseason signals. I tried returning snaps, coaching changes, rookie debuts, and a dozen other preseason ideas. None beat the model on both 2000-19 and 2020-25.

Notes

• Every number here is walked forward: each season is priced only from seasons before it. I develop and test on 2000-19. The holdout is 2020-25, read once for each candidate before it ships.
• Each comparison is paired. Two forecasts are scored on the same games, and the difference is taken game by game. Intervals come from resampling whole seasons.
• Every change is checked for leaks. The check rates past seasons twice, once with the future cut off and once without, and requires the same answer. It has caught leaks that no amount of reading the code did.
• Weeks 1 and 2 are where RICE is furthest from the market, and I haven't found a preseason signal that closes the gap. Some of that gap is information the model doesn't have, such as what happens in training camp.
• The model knows only what's been reported. Until the inactive list comes out, a game is priced from the injury report, and a team with no report is treated as unknown.

Before 1999

For 1977-98, the only record is the play-by-play (from a second source) and the schedule. There are no inactive lists, rosters, injury reports, or snap counts. Those seasons are rated with team strength and the quarterback only. They're shown as history on the team and quarterback pages, and no forecast reads them. The record has no game clock for 1981, so that season's clock is estimated from each play's place in the quarter. The strike seasons of 1982 and 1987 are rated as played, replacement games included.