The pattern behind every striking matchup stat

Striking numbers are not distributed evenly across a dataset. They cluster where the sample is smallest, because that is the only place a couple of possessions can move a rate to an extreme. This is not a quirk of basketball data, it is arithmetic, and it is measurable here.

Every column tells the same story. Spread collapses as evidence accumulates. The maximum observed rate falls from 4.00 to 0.54, not because thick-sample scorers are worse, but because no one sustains 4.00 over fifty possessions, and anyone can over three.

The scoreless problem

The last column is the one that should change how you read matchup claims. In 55.2% of pairings under 15 possessions, the scorer did not score at all. Given that a third of tracked rows contain no shot attempt whatsoever, that is unsurprising, and it is fatal to a particular kind of statement.

When a graphic says a defender held a scorer scoreless across a matchup, the overwhelmingly likely explanation is that the two were barely matched. It is a fact about opportunity, not about defense. In the thickest band, scoreless pairings are 1.8%, because over fifty possessions a scorer will eventually score.

Why this compounds when someone goes looking

The distribution alone would be manageable. What makes it dangerous is that nobody browses matchup data at random. People search for a specific pairing, scan for the striking line, and stop when they find one. That process selects for extremes, and extremes live almost exclusively in the thinnest samples.

So the search itself produces the bias. A reader who looks at ten matchups and reports the most dramatic has not found the strongest signal, they have found the smallest sample. This applies with equal force to anything published here.

Three habits that defend against it

• Set the sample threshold before you look, not after. Deciding what counts as enough once you have seen the number is how the number decides for you.
• Treat the possession count as part of the statistic. A rate without it is not a weaker statistic, it is not one.
• Expect regression toward the middle. A pairing at an extreme is far more likely to be mismeasured than to be a genuine outlier, and the thinner it is the more that holds.