Box Creation and Offensive Load, explained

An assist is a narrow accounting event. The pass has to be the last one before a made basket, so everything that happens one step earlier in the sequence goes unrecorded, and so does everything that fails to convert. A guard can bend an entire defense and get nothing in the box score for it.

Box Creation is an attempt to count the thing the assist column is a partial proxy for: scoring opportunities generated, whether or not they turn into points. That is a different question from "who made the final pass", and for most offenses it is the more useful one.

Taylor built Box Creation by hand-tracking "opportunities created" across several hundred games, then regressing that tracked count onto ordinary box-score rates. The result is a formula that reproduces the tracked number closely enough to be run on any season with a box score, including decades that have no tracking data at all. He reported a mean absolute error of 0.77 created opportunities per 100 for players with 1,000 or more possessions in the tracked sample, and 0.90 at 500 or more, and he publishes historical values at Backpicks.

The original write-up is at Nylon Calculus ("Measuring creation with the box score", August 2017) and Taylor's own sites are thinkingbasketball.net and backpicks.com. If you want the derivation and the tracking work behind it, read him first. This page explains the idea and hands you our version of the numbers.

Usage rate counts shot attempts, free throws, and turnovers. That is a reasonable read on a scorer and a poor one on a playmaker, because the work of setting up an offense mostly does not end in your own shot. Two players can carry very different responsibilities and post nearly the same usage.

Offensive Load fixes that by adding creation into the total and discounting the assists that would otherwise be counted twice. The 0.38 term removes the overlap with Box Creation; the remaining 0.75 factor discounts passes that were more relay than creation. Taylor also derived an adjusted turnover rate from it, dividing turnovers by load rather than by usage, which is fairer to players whose job is to hold the ball.

A box score records that a pass happened. It cannot see whether the pass arrived on time, on target, and on the correct side of the defender, or whether it was a hopeful lob into traffic that happened to work. Box Creation therefore measures how much creation a player is generating, not how good he is at it. Two players with identical creation figures can be very different passers.

There is a metric that addresses this. Taylor's Passer Rating estimates the quality of the looks a player generates relative to his load, and it depends on tracking inputs (things like which assists were layups, and what share of created looks were genuinely good ones) that our data feed does not carry. We could approximate it. We have decided not to, because an approximation of a quality metric built from the same box score that already produced the volume metric would mostly be re-describing the volume metric with more confidence than it earns.

Two smaller cautions. Creation is correlated with scoring volume by construction, so heavy-usage guards will always sit near the top, and it says nothing at all about defense. It is one axis of an offensive profile, not a summary of a player.

We take Taylor's published formulas and run them on per-game box stats converted to per 100 possessions using the player's season possession total. The three-point proficiency term uses the volume-gated sigmoid above, and Box Creation is floored at zero, since a negative count of created shots is not a meaningful quantity. Coverage runs across all nine seasons we reconstruct, 2017-18 through 2025-26, for both regular season and playoffs.

Box Creation and Offensive Load are not ours. Neither are BPM 2.0 and VORP, which use Daniel Myers' published coefficients, and RAPM itself is a long-established public technique that we run on our own reconstructed possessions. What is ours is the lineup reconstruction all of this sits on, IPM, and lineup chemistry. Box Creation and Offensive Load are borrowed, credited, and computed on our data. The full write-up sits on the methodology page.

Creation and load are role measures. Impact metrics answer a different question, so the interesting reads come from putting them side by side. A large offensive load next to a modest O-IPM suggests an offense leaning on a player harder than the results justify. A high creation figure on a low-usage player is usually a sign of a connector who does not get the ball enough.

Play type data is the natural next check. If Box Creation says a player generates a lot, the pick-and-roll ball handler and isolation splits tell you how, and at what efficiency. WOWY tells you what happens to the offense when he sits, which is the crudest but most direct test of whether the creation is load-bearing.