How accurate is ESPN BPI?
We graded ESPN's pregame win probability on every game of the last two NBA seasons where it was on record (2,631 games), beside the market's final price, Clutch, and picking the home team.
BY UĞUR ALTUNBAŞ · OCT 23, 2026
Good at picking winners. Its probabilities trail the market's.
ESPN BPI picked the winner in 69.0% of games in 2024-25 and 68.6% of games in 2025-26. That is about as often as anything does: the market's final price picked 68.7% and 68.6%, and picking the home team every night managed 54.6% and 55.5%.
Where BPI falls behind is the quality of the number itself. On the Brier score, which grades the probability and not just the side, BPI scored 0.206 and 0.202; the market's final price scored 0.201 and 0.197, better in both seasons. A good independent model lands about where BPI does: Clutch's own number, before it looks at the market, scored 0.206 and 0.199.
2,631 games, graded the same way
| NUMBER | 2024-25 RIGHT | SCORE | 2025-26 RIGHT | SCORE |
|---|---|---|---|---|
| ESPN BPI | 69.0% | 0.206 | 68.6% | 0.202 |
| The market's final price | 68.7% | 0.201 | 68.6% | 0.197 |
| Clutch's call | 70.0% | 0.202 | 68.5% | 0.198 |
| Clutch, Model only | 67.4% | 0.206 | 68.8% | 0.199 |
| Always the home team | 54.6% | 0.248 | 55.5% | 0.247 |
RIGHT: how often the side given over 50% won. SCORE: the Brier score, lower is better; a coin flip on every game scores 0.250. Every row is graded on exactly the same 1,315 games in 2024-25 and 1,316 games in 2025-26.
Why picking the winner isn't the whole story
Most NBA games have a clear favourite, and favourites win about two games in three. Any sensible number picks those games, so every source above lands near 69% and the share of winners barely separates them. What separates them is the number. A 90% call that loses should cost more than a 55% call that loses, and a team given 70% should win about 70 times in 100.
The Brier score does exactly that. For each game, take the probability given to the team that lost, square it, and average over every game. A 55% call that misses costs 0.30; a 90% call that misses costs 0.81; a 90% call that holds costs 0.01. It rewards being confident when you're right and punishes it when you're wrong, which is what a probability is for.
Is Clutch more accurate than BPI?
By the Brier score, Clutch's call was more accurate than BPI's in both seasons (beyond the noise in 2025-26, within it in 2024-25). Most of that edge is the market's: right before tip, Clutch's call carries almost entirely the market's final price (its own model's weight is 2% there), because by then the market has priced the same injury news and more. Clutch's own number, Model only, is level with BPI: neither is ahead beyond the noise in either season.
So the plain summary: BPI is a good independent number. The market's final price is better than both, and Clutch's call matches it rather than beating it. Nobody publishing an NBA model reliably beats the market's final price, and Clutch doesn't claim to.
Early in the season
In each team's first games, every rating still leans on last season and has not seen the new roster play. Over games where either team had played fewer than 15, BPI had the worst score of the four in 2024-25 (0.213 on 236 games) and the best score of the four in 2025-26 (0.189 on 232 games). Two seasons that point in opposite directions are not a pattern.
This season, live
This season the Ledger grades both after every final, on the same games: so far Clutch 66.7% right and BPI 69.4% on 36 games. Every call is sealed in public before tip, so neither side's record can be edited afterwards. The live record, band by band →
How the test was run
The games. Every game of the 2024-25 and 2025-26 seasons with both ESPN's pregame probability and the market's final price on record: 2,631 in all. Every number is graded on the same games.
ESPN BPI. ESPN's published win probability for the game, as it stood at tip-off. Because it is the last value before tip, it already reflects late injury news, which makes the comparison fair to BPI.
The market's final price. The last price before tip, read as a fair probability: both sides adding up to 100%.
Clutch. Each month was predicted by a model fit only on games played before it, the way it would have run live. Clutch's call blends that number with the market's price; Model only is the number before the blend.
The noise. Each difference comes with a likely range: where it lands 95 times in 100 when the same games are resampled. 2024-25: Clutch's call minus BPI −0.0045 (likely range −0.0093 to +0.0006); Model only minus BPI +0.0001 (−0.0051 to +0.0052). 2025-26: Clutch's call minus BPI −0.0044 (likely range −0.0081 to −0.0008); Model only minus BPI −0.0027 (−0.0064 to +0.0009). A range that crosses zero means the gap could be luck.
Numbers from Clutch's walk-forward test, published with the model fit on Sep 29, 2026; this page is rebuilt from them, and from the Ledger, every day. ESPN and BPI are ESPN's; Clutch is not affiliated with ESPN or the NBA. How Clutch's model works →