Draft models cannot see who runs back

Prospect evaluation is built almost entirely on what happens with the ball, which is where the least of a career actually lives.

Feed a draft model everything a college season produces and it will tell you a great deal about shooting, a fair amount about rebounding, something useful about how a player handles pressure with the ball in his hands. It will tell you close to nothing about whether he defends the eighty feet between possessions, because nothing in the box score records it and most college programmes do not have the tracking data that would.

This matters more than the models let on. A meaningful chunk of a defence's quality lives in decisions made without the ball in the frame, and the single largest one is when a player turns around. Draft evaluation is a with the ball discipline being asked to predict a mostly without the ball job.

The usual defence of the models is that these things wash out. Effort is coachable, they say, and a professional environment will fix it. I do not believe that. Recognition is not effort, and recognition is the part that decides transition defence. Knowing a shot is going to miss from the sound and the shooter's hips, turning at that instant rather than a beat later, choosing the rim over your own assignment. Some nineteen year olds already do this. Others never learn it and scouts have been saying so, in words, for decades, only to be told the words are not evidence.

So the words get discounted, and the model wins the argument, and a team drafts a scorer who will cost his defence a point of efficiency every night in ways the model was structurally unable to see.

What I want is unglamorous. Charge someone with a stopwatch and a clipboard to log one thing across twenty games: the delay between the shot leaving a prospect's opponent's hands and the prospect changing direction. It is a tenths of a second measurement and it is perfectly countable by hand. Then put that number in the model as a column, and let it fight for weight against three point percentage.

It will not win. It should not win. But a column with a small coefficient beats an attribute with no column at all, and right now half of defence is entering the process as vibes.

Scouts have been right about this for years. They just could not turn it into a number, so it did not count.