Calibration and the Brier score, explained
Two ways to grade anyone who puts a number on a game: does it come true as often as it says, and was it confident when it mattered? With real NBA calls, scored one game at a time.
BY UĞUR ALTUNBAŞ · OCT 23, 2026
Does 70% happen 70% of the time, and how confident was it when it mattered?
Calibration asks whether a forecaster's numbers come true as often as they say: of all the games called at 70%, the favourite should win about 70 in 100. The Brier score turns calibration and confidence into one number per forecaster. Lower is better and a coin flip on every game scores 0.250. Over the last two NBA seasons the market's final price scored 0.201 and 0.197, ESPN BPI 0.206 and 0.202, and picking the home team every night 0.248 and 0.247.
Calibration
Put a forecaster's games in bands by the number it gave the favourite: 50–60%, 60–70%, and so on. In each band, count how often the favourite won. A calibrated forecaster's 60s win about 60% of the time and its 80s about 80%. If its 80s win only 65%, the 80s are overconfident, however many games it gets right.
Small samples lie. Eight games in a band can easily come out 5 of 8 or 7 of 8 by luck, so a band means little until it holds a few dozen games. Clutch's bands this season, so far: 50–60%, 10 of 18; 60–70%, 3 of 7; 70–80%, 6 of 6; 80–100%, 5 of 5. Every band, updated each morning →
The Brier score, one game at a time
For each game, take the number given to the side that lost and square it; average over every game. A favourite at 88% that loses costs 0.774; the same favourite winning costs 0.014. Here are Clutch's graded calls from Oct 24, scored:
| GAME | CLUTCH | RESULT | COST |
|---|---|---|---|
| GSW at PHX | PHX 77% | RIGHT | 0.052 |
| NOP at DEN | DEN 80% | RIGHT | 0.039 |
| HOU at ATL | HOU 51% | RIGHT | 0.238 |
| WAS at CHI | CHI 53% | MISSED | 0.277 |
| SAS at DAL | SAS 88% | RIGHT | 0.014 |
| MIL at PHI | PHI 84% | RIGHT | 0.027 |
| BKN at CLE | CLE 81% | RIGHT | 0.037 |
| MIA at CHA | MIA 60% | MISSED | 0.365 |
The night's average: 0.131 on 8 games. One night says little; the seasons below are 1,300 games each.
What a good score looks like
A coin flip on every game scores 0.250. Picking the home team at its usual rate gets you to about 0.248 and 0.247. The serious NBA forecasts sit close together: ESPN BPI scored 0.206 and 0.202 over 2024-25 and 2025-26, Clutch's call 0.202 and 0.198, and the market's final price 0.201 and 0.197.
Those gaps look tiny, and over a single night they are noise. Over 1,300 games, a gap of 0.004 can be real, which is why every comparison Clutch prints comes with a likely range: the span the gap falls in 95 times in 100 when the same games are resampled. The full comparison with ESPN BPI →
Why both matter
Calibration alone can be gamed by never saying much: a forecaster that gives every home team 56% is close to calibrated and tells you almost nothing. The Brier score catches that, because it rewards numbers that are far from 50% and right. A forecaster that is bold and wrong loses on both counts. Good forecasts are calibrated and as confident as the evidence allows.
The night and the bands are rebuilt from Clutch's Ledger every day; the season scores from its walk-forward test. More on reading the numbers →