Our baseline is an Elo model. Every team starts at 1500 before the 2010 season. At the start of each later season, ratings move one third of the way back toward 1500.
Before kickoff
The rating difference includes 48 points for home field, except at a neutral venue, and 3 points for each extra day of rest, capped at seven days. The home win probability is 1 / (1 + 10^(−difference / 400)). The estimated home winning margin is the rating difference divided by 25.
After the game
Ratings change by 20 × (result − predicted probability). A win scores 1, a tie 0.5 and a loss 0. Parameters were fixed before evaluating the 2023–2025 seasons. We do not optimize them on those test results.
Limitations
The model does not account for injuries, starting quarterbacks, weather or roster changes. It predicts a two-outcome equivalent and does not separately model ties. These estimates should be read alongside available team and injury information. The market was more accurate in our historical test.
Market quotes
When available, NFL moneyline equivalents come from the Polymarket US and Kalshi public APIs. We match both teams and the scheduled kickoff, then read the current buy quotes for each side. Decimal odds equal 1 divided by the contract price; American odds are another representation of that price. Each platform is listed separately with its own check time. These figures exclude fees and do not come from the Elo model. Different settlement rules mean these are not a guarantee of the best final return.
We check for new quotes about every two minutes. While this page is open, displayed prices refresh automatically once a minute. Quotes expire after five minutes or at kickoff, whichever comes first. Closed, suspended, unmatched and stale markets are omitted. Overtime counts; a tie pays $0.50 per contract, rather than refunding its purchase price. The platform’s current rules and available quantities determine execution.
Sources and timing
Schedules and results come from nflverse/nfldata. Supplementary data comes from nflverse-data under its CC BY 4.0 notice, subject to the original providers’ rights. GidStats calculates team summaries and model estimates. Data is provided as available, without a warranty of completeness.
All displayed game dates and times use Eastern Time, with daylight saving time applied. Historical model figures are marked as backtests. Future predictions are stored with their real capture time; no historical capture date is invented.