Every number on the site, first in a sentence and then precisely.
The Two Models
LOWFI/1
The LOWFI model that sees points. It rates every offense and defense in points per game against an average opponent, and it's never told who won.
Target: end-of-season offensive and defensive BLUPs from a crossed random-effects model of points scored, fitted on every game since 2004 and centered within season. Estimator: GAMs on in-season scoring features, with a separate week 1 model built only from preseason inputs, then recalibrated per week out of fold. Preseason prior: de-vigged market win totals, split into offense and defense using last season.
LOWFI/0
The original LOWFI model. It learns only from who won each game - not the score, not the opponent's quality - and it keeps one number per team.
A Beta distribution over each team’s true strength. Preseason: last season’s win rate regressed toward .500, with prior width from the variance of season win rates after removing binomial sampling noise. Games: log5. Updating: rejection sampling — draw strengths, keep the draws consistent with each actual result, refit a Beta by method of moments.
Ratings
Net rating
How many points per game a team would beat an average team by, on a neutral field. Plus is better than average.
LOWFI/1 offense minus LOWFI/1 defense (raw coefficient), in points per game against an average opponent on a neutral field. The in-season number is the estimator’s prediction of the end-of-season rating, not the season to date.
Offense and defense
The two halves of net rating. Offense is points scored above average, and defense is points kept off the board, so on this site higher is better for both.
Offense: points scored per game above an average offense, against an average defense. Defense is shown as points prevented, so higher is better; the data release keeps the raw coefficient, where more negative is better. Defense carries less separable signal than offense, so its bands are wider.
True strength
How often a team would beat an average opponent if they played forever. A .600 team is better than average, and .500 is average.
LOWFI/0 only. The posterior mean α / (α + β) of the team’s Beta distribution: its long-run probability of beating an average opponent on a neutral field.
Bands and ±
What LOWFI doesn't know yet - not how much a team swings from week to week. Bands are wide in September and narrow as the games pile up.
LOWFI/1 bands are ±1.645 standard errors (a 90% band). The standard error blends the estimator’s own uncertainty with historical rating dispersion, inflated early in the season by a factor that decays with games played. LOWFI/0’s ± is the standard deviation of its Beta distribution; its trajectory band is ±1.645 of those.
Rank
Where a team's number sits among all 32. It always ranks the number printed right next to it.
The rank of the number printed next to it, 1 to 32. A rank always ranks the value beside it.
Projections
Projected wins
The average win total across every simulated season.
The average win total across 10,000 simulated seasons, run through nflseedR with full seeding and tiebreaker rules. Each simulation draws one rating per team from its band and holds it for the season, then samples every remaining game.
Playoffs and division
How often a team makes the playoffs, or wins its division, across all the simulated seasons.
The share of the 10,000 simulated seasons in which the team made the playoffs, or won its division.
Pythagorean pace
What a team's points scored and allowed say its record should be over 17 games. It isn't LOWFI - it's the classic yardstick LOWFI gets measured against.
Points for2.37 / (points for2.37 + points against2.37), times 17. Not a LOWFI model: a reference point that uses scores but ignores opponents.
Schedule ahead and played
How good a team's opponents are by LOWFI/1, for the games still to come and the games already played. Plus means tougher.
The average current LOWFI/1 net rating of a team’s remaining opponents (ahead) or past opponents (played). Positive means a harder schedule.
Games
Expected margin
How many points a model expects a team to win or lose by. On team pages it always reads from that team's side, so plus means favored.
LOWFI/1: an OLS regression of home margin on both teams’ offensive and defensive ratings; the intercept is home field. LOWFI/0: its win probability mapped to a margin by regression on the log5 logit, fitted on training seasons only — a calibration, not a forecast of points. Market: the posted spread. On team pages every margin reads from that team’s side, so + means favored.
Win probability
The chance a team wins the game, from the model's full distribution of possible margins.
From the margin distribution: an ordered probit over integer margins from −49 to +59, located by the expected margin, so 3 and 7 are as common as football makes them. A tie counts half. LOWFI/0’s game probabilities on the site run its mapped margin through the same distribution.
Home cover and push
LOWFI/1's chance that the home team beats the spread, and the chance the game lands exactly on it.
LOWFI/1’s probability that the home team beats the posted spread, and that the game lands exactly on it, both from the same margin distribution.
Market
What the betting market thought. For live forecasts it's the price on file that Tuesday, and for retrodictions it's the closing line.
Live: the market win probability on file with that Tuesday’s forecasts, de-vigged, and the spread posted at the time. Retrodiction scoreboard: the closing spread run through the same margin distribution the models use.
No-vig
Betting prices with the bookmaker's cut taken out, so the two sides' chances add up to 100%.
Implied probabilities from both moneylines, divided by their sum so they add to one. Proportional de-vigging, which leaves the favorite-to-underdog ratio untouched.
Stake
How much LOWFI/1 would put on a game, in units per hundred - and zero when it would pass. It's there as a comparison with the market, not advice.
Quarter Kelly sizing in units per hundred, capped at 2 units, and zero wherever the expected value is under 2%. A comparison with the market, not advice.
Scoring and Records
Scoreboard points
Points from FiveThirtyEight's old forecasting game. A coin flip scores nothing, being certain and right is worth 25, and being certain and wrong costs 75.
Points = 25 − 100 × (P − R)2 per regular-season game, where P is the home win probability and R is 1 if the home team won, 0 if it lost. Ties score zero. From FiveThirtyEight’s NFL forecasting game, as recreated by Lee Sharpe.
Live and retrodiction
Live numbers existed before the games. Retrodictions are today's model reading seasons that already happened - useful, but not a track record.
Live: produced by that Tuesday’s run and archived before the week’s first game, then read back unchanged. Retrodiction: the current model applied to a past season revealed week by week. Seasons 2023 on are in the production model’s training data.
As of
When a number was produced. A forecast only counts if it existed before the week's first game.
When a number was produced: the date the week’s run wrote its archive. A forecast counts only if that date is on or before the week’s first game.