Game Prediction Model

Lower log loss is better. It rewards being confident and right, and punishes being confident and wrong. The market wins on both measures, and the page does not pretend otherwise. What the model does have is calibration: its stated probabilities land within about 2.1 points of what actually happens, so a 70% here means close to 70%. That is the property worth using, and it improved sharply when the training window widened to 2006.

1. 01Team-game panelEvery team-game since 1999 is assembled with its own offense and, mirrored from the opponent, its defense.
2. 02Rolling formEach metric is smoothed with a 10-game exponentially weighted average so recent games matter more than old ones.
3. 03Leakage guardEvery rolling feature is shifted one game before smoothing, so a prediction for a game can only ever use games played before it.
4. 04Matchup interactionsA team's offensive EPA is paired against the specific defense it faces, rather than compared to league average.
5. 05Situational contextRest days, bye weeks, injuries, venue, weather and the opponent's base defensive front are encoded as categorical inputs.
6. 06Walk-forward validationThe model is only ever tested on seasons later than the ones it trained on, never a random split.

EPA / Play

Expected Points Added per play

Pass EPA / Dropback

Passing efficiency

Rush EPA / Carry

Rushing efficiency

CPOE

Completion % Over Expected

First Downs / Play

Drive-sustaining rate

Sack Rate

Sacks taken per dropback

Turnover Rate

Giveaways per play

Injury Impact

Position-weighted availability score

Explosive Pass Rate

20+ yard completions per attempt

The injury score has no values for an upcoming slate, because reports are not published this far out, which is why that metric shows an empty state on each game. These are completed games, kept so the measure means something concrete. They are history, not predictions, and are not part of any forecast on this site.

Three games is an illustration, not evidence. The weights behind the score are fitted across thousands: a quarterback absence costs roughly 2.6 points of margin and a skill-position absence 0.7, while line, linebacker and secondary absences came back indistinguishable from zero.

Margins are not a smooth curve. A field goal and a touchdown are the units the game is scored in, so games pile up on those numbers and thin out between them. Three points alone accounts for 14.6% of all games, more than three times either neighbor, while nine points, a margin no combination of scores lands on naturally, happens in 1.7%.

• 2 pts4.3%
• 3 pts14.6%
• 4 pts5.1%
• 6 pts6.5%
• 7 pts8.8%
• 8 pts3.8%
• 9 pts1.7%
• 10 pts5.2%

This explains why so many games finish close. It does not, on its own, help predict who wins them.

The head-to-head chart on a game page shows nine metrics, which can read as though nine numbers decide the forecast. They do not. Here is the whole input list, grouped.

Every feature the shipped model fits on, grouped. The head-to-head chart on a game page shows a readable subset; this is the whole input list. Wider feature sets were tested and scored worse: 202 features reached 61.2% and 455 reached 62.8%, against 63.2% for this one.

• Point differential
• Overall matchup
• Team win rate
• Offensive EPA per play
• Pass matchup
• Offensive EPA per dropback
• Offensive success rate
• INJ impact
• Team QBR
• Defensive pressure rate
• Home INJ QB out
• Offensive sack rate

Point differential and win rate dominate, which is worth knowing: most of what the model does is notice which team has been better lately. The efficiency terms refine that rather than replace it. Only the twelve heaviest are published, so a metric missing from this list is unranked rather than proven unimportant.