How it works

How Clutch's call is made.

In plain English, with the test it had to pass. Live performance is the Ledger's — never this backtest.

The short answers

What is Clutch?

Clutch is an NBA analytics app. It puts a win probability on every game, seals the number in a public GitHub file before tip-off, and grades it in public after the final. It is free in the browser and on iPhone, makes no calls on margins or totals, and carries no affiliate links.

What does the model use?

Form, rest, travel and the official injury report for each game, known that morning, blended with the market's price when the market has one.

Is Clutch more accurate than the market or ESPN BPI?

In the two-season test, Clutch's call has been more accurate than ESPN BPI's and level with the market's final price: as good as the market, not better. The sections below show the test and the model's limits.

Can I check that a prediction came before the game?

Yes. Every number is sealed with a SHA-256 fingerprint in a public GitHub repository before tip-off, and the Ledger grades the number in that file. Anyone can check a seal.

The test it had to pass

Two full seasons, replayed month by month: every game predicted by a model fit only on the games before it, then graded on the same games as ESPN's BPI and the market's final price before tip. The accuracy score (the Brier score) grades the probability itself — lower is better, and 0.250 is a coin flip. Each difference comes with its likely range: where it lands 95 times in 100 when the same games are resampled.

2025-261,316 games
NumberAccuracy scoreRight
Clutch's call0.19868.5%
Model only0.19968.8%
The market · final price0.19768.6%
ESPN BPI0.20268.6%
Always the home team0.24755.5%

Clutch vs ESPN −0.0044 (likely range −0.0081 to −0.0008) · Clutch vs the market +0.0005 (−0.0001 to +0.0012). Lower is better.

2024-251,315 games
NumberAccuracy scoreRight
Clutch's call0.20270.0%
Model only0.20667.4%
The market · final price0.20168.7%
ESPN BPI0.20669.0%
Always the home team0.24854.6%

Clutch vs ESPN −0.0045 (likely range −0.0093 to +0.0006) · Clutch vs the market +0.0007 (−0.0003 to +0.0017). Lower is better.

Read it plainly: Clutch's call has been more accurate than ESPN's and level with the market's final price — as good as the market, not better. Model only is level with ESPN. Nobody publishing an NBA model reliably beats the market, and Clutch doesn't claim to. The live record and its calibration →

1. What it knows, and when

Every NBA game since 2000 is replayed in date order, and each one is described using only what was known that morning: each team's margin this season (leaning on last season's until about ten games are in), a long-run Elo rating, the last ten games against the season, rest, back-to-backs and three-in-fours, travel, altitude, a neutral floor, and the league's official injury report — each listed player weighted by how likely his status keeps him out and what he has been worth. Nothing from later in a season can reach an earlier game.

2. The model

A regularised logistic regression on 18 inputs, fit on the last 10 seasons (12,839 games). It beat gradient-boosted trees in Clutch's own tests and needs no separate adjustment step. Because it is linear in log-odds, the "Why" behind every number is exact: the reasons add up to the number.

3. Clutch's call

When the market has a number for the game, Clutch's call weighs Model only against the market's price, read as a fair probability (both sides adding up to 100%). The weights were fit on three seasons of market prices: just before tip the market carries nearly everything (Model only 2%), because by then it has already priced the same injury news and more; a day out, Model only carries 48%. With no market number, Clutch's call is Model only. Both parts are graded separately on the Ledger.

4. Score and total

The projected margin is read off the probability through the market's own relation between a margin and a win chance, so the score can never name a different winner than the number. The total follows the market's when there is one (Clutch's own has not beaten it) and Clutch's scoring model otherwise. Both are in Pro.

5. Weekly refit, behind a gate

Every Monday the model is refit with the week's games. The new fit replaces the old only if, on the most recent 250 graded games, a copy trained without them scores no worse — otherwise the old model stays. The call for each game is sealed before tip-off.

6. Sealed before tip

Before each game tips, its number goes into a public file on GitHub with a SHA-256 seal; after the final, the Ledger grades the number in that file. Anyone can check a seal with tools that come with their computer.

What this model cannot do

It is level with the market, not ahead of it

Matching the market's final price is the ceiling for anyone publishing numbers, and that is where Clutch is. Where Model only disagrees with the market, the Ledger shows who turned out right.

Margins and totals are projections, not calls

Clutch's margin comes from its probability and its total follows the market's, so neither is an argument with the market. Clutch makes no calls on margins or totals.

Preseason is close to a coin flip

Across 1,122 preseason games (2012–13 to 2025–26) the regular-season model scored 0.262 — worse than a coin flip's 0.250 — because starters rest. Pulled toward 50/50 it scored 0.247. So preseason numbers stay near 50/50, carry a label, and are never graded.

Early season and playoffs are harder

In the first two weeks each team's rating still leans on last season, and roster changes it hasn't seen yet cost it — ESPN's preseason projections were better early in 2025-26. The playoffs offer far fewer games to learn from. The game page says so on the games it applies to.

What's free, what's Pro

Free shows Clutch's call, the market's number and the top three reasons for every game, who's out, and the Ledger. Pro shows all of it: every factor and how much of the call it carries, Model only — the number before the market — the projected final, and the simulation. Pricing →

Past performance does not guarantee future results. Predictions are analytics for information and entertainment, not advice to spend money on a game. Model performance varies by sample size and league conditions. Model fit Sep 29, 2026 on games through Jun 13, 2026.