How the win probability model works
The calculator is a logistic regression. It takes the situation before a snap, turns it into twelve numbers, weights each one, adds them up and squeezes the total through the logistic function so the answer lands between 0 and 1:
The weights were fitted on 324,950 regular-season plays from 2016 to 2024 in the open nflverse play-by-play data, using only games that did not end in a tie. The label for each play is whether the team with the ball won the game.
Three ideas carry most of the weight. The score margin matters more as time runs out, so it enters three times: on its own, scaled by the square of the elapsed share of the game, and divided by the square root of the minutes left. The pregame spread is a measure of team strength, so it matters most at kickoff and fades with the clock. Field position and down describe how likely the current drive is to end in points.
Model coefficients
| Feature | Weight |
|---|---|
| Intercept | 0.5403 |
| Score margin | 0.0685 |
| Margin × (time elapsed)² | 0.0997 |
| Margin ÷ √(minutes left + 1) | 0.3316 |
| Pregame spread × time left | 0.1986 |
| Field position | 0.2407 |
| Field position × time elapsed | 0.4650 |
| 2nd down | -0.0681 |
| 3rd down | -0.1828 |
| 4th down | -0.3539 |
| log(yards to go) | -0.0518 |
| Timeout edge × time elapsed | 0.0645 |
| Offense at home | -0.1289 |
Field position is scaled so that the offense's own goal line is −1, midfield is 0 and the opponent's goal line is +1. Time elapsed runs from 0 at kickoff to 1 at the end of regulation. A positive weight raises the offense's chance of winning.
Is the win probability calibrated?
A calibrated model is right about its own confidence: plays it rates at 70 percent should be won about 70 percent of the time. The chart compares the model's predictions on the 2025 season, held out from fitting, with what actually happened.
The curve sits close to the diagonal in the middle and bends slightly flatter at the ends: when the model is very confident it is a little too confident, which is the usual price of a simple linear model on a problem where the true relationship is curved. Gradient-boosted models close most of that gap; the win probability model series walks through that trade-off.
What the calculator leaves out
- Kickoffs, extra points and two-point tries: the model prices plays from scrimmage.
- Overtime, weather and injuries.
- The last yards before the goal line. Field position enters as a straight line, so a late snap at the 2 counts little more than one at the 12: in 21 late drives since 2016 that started first to third down inside the 2 while trailing by four to six, the offense won 13 times against a model average of 27 percent.
- Team quality beyond the pregame spread. Two teams with the same line are treated as equally strong at any moment of the game.
For the step-by-step version of the idea, the win probability explainer replays one drive and prices every snap.
Frequently asked questions
How accurate is this NFL win probability calculator?
On the 2025 regular season, which the model never saw during fitting, its Brier score was 0.158 against 0.250 for a coin flip, a 37 percent improvement. The calibration chart on this page shows predicted against observed win rates in ten-point buckets.
What does win probability mean in football?
It is the share of past games, in situations like the current one, that the team with the ball went on to win. A 70 percent reading says that across many similar spots about seven in ten teams won, not that this team will win.
Why does the pregame spread matter during the game?
The spread carries information about team strength that the score alone does not. Early in a game it dominates; its weight fades toward zero as the clock runs down, which is why the model multiplies it by the share of time remaining.
Why does the estimate jump at the end of the game?
With little time left, a small margin becomes decisive. The model lets the effect of the margin grow as time elapses, and at zero seconds the result is simply the scoreboard.
Is this the same model broadcasters use?
No. Public models such as the one in nflfastR use gradient-boosted trees and more features. This calculator is a deliberately transparent logistic regression so every coefficient can be read in the table below.
Does the calculator include overtime?
No. The model is fitted on regulation plays of regular-season games, so the clock runs from 60:00 to 0:00. Ties at the end of regulation are shown as 50 percent.