An attempt to measure the hardest job in basketball to quantify
Working on this website trying to analyze different aspects of a players game, it takes time to get ideas on what to probe at next. While thinking of another idea, and watching Game 4 of the 2026 NBA Finals seeing some questionable decisions I set on the task of trying to quantify coaching skill. While this task is nearly impossible because of all the noise surrounding coaching I tried using a variety of a different metrics to determine which coaches are the best. It goes without saying that this is not a definitive list but rather a fun attempt to use data to point us in the right direction.
Five signals go into the score. Each one is converted to a league percentile (where does this coach rank among the 30?) and then weighted by how much we trust the underlying measurement (or frankly care about the metric). The weights matter: coach RAPM gets twice the weight of challenges because four seasons of every possession is a stronger signal than less impactful coach challenges.
When a coach calls a timeout, the possession immediately after is one they drew up. For every timeout in the dataset, the offensive rating on that play is compared to the team's season baseline ORTG. The difference (ATO Diff) is how many points per 100 possessions the coach's drawn plays generate above or below their team's normal offense. A coach averaging +4.0 is generating four extra points per 100 possessions every time they touch the whiteboard.
The same concept applied to sideline out-of-bounds plays, possessions inbounded from the sideline, usually late in quarters or after stoppages. SL Diff is the ORTG on those plays minus the team's baseline. Coaches who design clever sideline sets score well here.
Simply the percentage of challenges the coach wins. It carries less weight in the composite relative to the other components. Challenge success, while a real decision, is the least indicative signal of overall coaching quality in this model.
Using Jeremias Engelmann method, Coach RAPM is a Regularized Adjusted Plus-Minus treating the coach as a sixth player on the floor for every possession. A ridge regression over four seasons of possession data estimates each coach's point impact per possession while controlling for every player on the court on both sides. Heavy regularization (α = 10,000) and 30 bootstrap resamples keep the estimates stable. Displayed per 100 possessions, like player RAPM. This is the strongest signal in the composite and its weight reflects that. But four seasons is still a limited sample for isolating a coaching contribution from roster quality, so treat individual values as directional, not precise.
The same coach-as-sixth-player RAPM, restricted to clutch possessions: fourth quarter, under five minutes, score within ten. The idea is to capture late-game decision-making, when to call timeouts, which lineups to trust, what plays to run when it matters most. Even across four seasons, clutch possessions are a small fraction of total data, so the confidence intervals are wide and this is the noisiest component in the composite.
Each component is converted to a percentile among the 30 coaches, then combined with weights reflecting confidence in each signal:
Read the score as a structured attempt, not a definitive ranking. The honest problems: