Skip to main content
4DFourth & Data
Fri 01:15SEA53%DENSun 14:30HOUJAX64%Sun 18:00BAL71%CLESun 18:00CAR57%PHISun 18:00CHI53%ATLSun 18:00NONYG60%Sun 18:00NYJNE82%Sun 18:00PIT51%TBSun 18:00TENIND70%Sun 21:05ARILA69%Sun 21:25BUF70%LVSun 21:25LACKC87%Mon 01:20DAL53%GBTue 01:15WASSF85%
Method

How the model works, and what it can't do

We'd rather you understood the game than followed a tip. So here's the working.

What this is

Fourth and Data is an analysis publication with a model inside it. We take a small number of games apart properly (protection, who got open, the calls, the drives, the defence) and we publish a forecast for every game, graded against the result and against the betting market.

It's written by a Steelers fan, which is why Pittsburgh gets a deep dive every week. The numbers don't get to be fans: Steelers games are broken down exactly the way anyone else's would be.

The model

The win-probability model is an ensemble trained on NFL play-by-play since 2010, calibrated so that its percentages mean what they say. Features are built from efficiency (expected points added, success rate, drive outcomes), recent form, schedule strength, rest and the injury report. Every feature is lagged: the model only ever sees what was known before kickoff.

Separate models project the margin and total points, and every game gets an F&D score out of 100 for how clear-cut the call is. The model is retrained every Wednesday once the week's games are final, and forecasts are refreshed daily as a game approaches, picking up the latest injury news and odds.

How we grade ourselves

Every pick is published before kickoff and stored as a timestamped, unchangeable snapshot. Each game is graded from the last forecast published before kickoff, never a later one. We grade against the betting market as well as the result, we show the weeks the market did better, and we publish calibration: when the model says 70%, you can check how often that actually happened.

In 2025 the model went 140-80 (63.6%) picking winners with a Brier score of 0.230. The live record, the market's record and every individual pick are on The Record.

What it can't do

Week 1

Early-season forecasts lean entirely on last season. Roster turnover, new coaches and new quarterbacks take weeks to show up in the data, so week 1 is where we’re weakest.

Week 18

Teams rest starters once seeding is locked. The model has no concept of motivation, and across five seasons it has been worse than the market in week 18. We publish the line but treat it as a no-bet week.

Late news

A quarterback ruled out after our last run isn’t in the forecast. The market often knows first.

Variance

A 65% favourite loses about one time in three. Single games prove very little either way; seasons do.

On betting

We are not a tipster. Forecasting well and betting profitably are different things: the bookmaker's margin is a real cost, and our own testing at real, bettable prices says we don't beat the market. Where the model disagrees with the market, treat it as a question worth reading into, not a bet.

If you do bet, set limits you're comfortable losing. 18+ only. Responsible gambling has help and resources for the UK and US.

Data

  • Play-by-play: nflverse (nflfastR / nflreadr data), historical and live.
  • Charting and tracking: FTN charting, Next Gen Stats and PFR advanced stats, used for analysis only, never as model inputs.
  • Injuries: the official injury report via the Sleeper API.
  • Odds: The Odds API, de-vigged to fair probabilities.

Questions or corrections: support@fourthanddata.com.