Updated Sep 7, 2026 · Week 1 board.
The formula, plotted against forecast probability
The definition is short. A Brier score is the squared difference between a forecast probability and the outcome: BS = (p − o)². The p is the probability you assigned. The o is 1 if the thing happened and 0 if it didn't.
Lower is better. Zero is perfect, one is as wrong as a forecast gets. The chart plots both curves across every value of p from 0 to 1. Blue is your score if the pick lands: (1−p)².
Red is your score if it doesn't: p². The eight dots on each curve are the eight picks in the table below, sitting at their own published confidence. It updates whenever our picks do. Today it shows probabilities and nothing else, because there's no outcome yet to plot against.
Eight house picks, priced as forecast probabilities
The confidence on each published pick is read here as p, the forecast probability that the picked side is right. The correct-branch and incorrect-branch columns show what BS = (p−o)² comes to under each possible ending, computed straight from p. Correct branch is (1−p)². Incorrect branch is p².
Neither column claims anything about what happened. You get both, because those are the same eight dots plotted above and the ending isn't known.
| Game | Market | Side | Confidence (p) | If correct | If incorrect |
|---|---|---|---|---|---|
| Carolina @ New Orleans | moneyline | home | 0.5505 | 0.2021 | 0.3031 |
| Denver @ Pittsburgh | moneyline | away | 0.5507 | 0.2019 | 0.3033 |
| Cleveland @ Tampa Bay | moneyline | home | 0.6101 | 0.1520 | 0.3722 |
| Atlanta @ Green Bay | total | under | 0.6114 | 0.1510 | 0.3738 |
| Cleveland @ Jacksonville | moneyline | home | 0.7187 | 0.0791 | 0.5165 |
| New York Giants @ Los Angeles Rams | moneyline | home | 0.7193 | 0.0788 | 0.5174 |
| Cincinnati @ Miami | spread | home | 0.9189 | 0.0066 | 0.8444 |
| Cincinnati @ Atlanta | spread | home | 0.9116 | 0.0078 | 0.8310 |
Read the shape before you read the numbers. As p climbs from 0.55 toward 0.92, the correct-branch score drops from 0.2021 to 0.0066. The incorrect-branch score climbs from 0.3031 to 0.8444. That's the blue and the red pulling apart at the right edge of the chart.
Confidence pays enormously when you're right and costs enormously when you're wrong. Near p = 0.55 a coin-flip forecast barely moves either way, which is why the two curves nearly touch on the left. That asymmetry is the entire thing a Brier score exists to measure, and you can see it here without a single settled game.
Why the score cannot be computed yet
Of our 6,120 published picks, exactly 1 has a real result, and it was a miss. The other 6,119 are unresolved. Effectively nothing in the whole history has landed. Our season-to-date line for these picks shows zero resolved picks: no win count, no return figure, no Brier average.
A separate model-level count shows 2,234 picks logged with zero wins and zero losses. None of the eight picks above is that resolved 1 out of 6,120. Averaging their hypothetical branches into one number would manufacture a Brier score nothing supports, so you're getting the state of things instead of a figure.
Two reference points worth knowing before the first Week 1 score exists
Two reference points to hold onto. A forecaster who always says p = 0.5, whatever the game, scores exactly 0.25 on every graded pick, because (1−0.5)² = 0.25 and 0.5² = 0.25. The outcome makes no difference. On the chart that's the one place blue and red cross.
That 0.25 is the no-skill line every real Brier score gets measured against. A mean below 0.25 across many graded picks says the forecaster is adding something past a coin flip. A mean at or above 0.25 says he isn't. Second point: always certain and always right scores 0, always certain and always wrong scores 1.
Every real forecast lives between those, all eight above included, since none of their confidence values is exactly 0 or 1. One graded pick proves nothing about calibration in either direction. Wait for a big graded sample rather than reacting to the first result, because both branches above stay plausible on any single game.
Our live confidence and edge figures on every published pick are at /picks/nfl. Settled grades, when they exist, land on /track-record, and that's where a real Brier score for our model will show up. Watch that page over the next few weeks as picks grade out. The three-book market prices behind these same games are at /odds, and the expected-value arithmetic this metric feeds is at /desk/tools/ev.
The Receipts Drawer
Nothing on this page proves our model is any good, and publishing it now is exactly the point. The curves and the eight dots are the measuring stick, set out before the first game so nobody gets to move it afterward. What settles it is real outcomes landing on real picks. So watch the track-record page as picks start resolving, and check whether our mean score sits under 0.25, because that's the line between adding information and flipping a coin.
