Regression to the Mean

An athlete has a career-best performance. Next time, they’re likely worse — not because of complacency or pressure, but because performance has a random component. Their career-best included lucky variance. Next time, variance will probably be less favorable. They regress toward their average.

Francis Galton discovered this in the 1880s studying height: tall parents had children shorter than themselves (on average), while short parents had taller children. Not because height was equalizing, but because extreme values reflect both signal (genetics) and noise (everything else). The noise component doesn’t repeat.

Regression to the mean generates persistent illusions. A manager praises an employee after exceptional work, then sees performance decline — and concludes praise backfired. Another criticizes after poor work, sees improvement — and concludes criticism works. Both are wrong. Performance would have regressed regardless. The intervention gets credit or blame for what statistics guaranteed.

The same illusion affects medicine (patients seek treatment when symptoms are worst, then improve), coaching (athletes get extra attention when slumping, then recover), and policy (interventions target problems at their peak, which then moderate). We attribute to causes what belongs to randomness.

Recognizing regression requires separating signal from signal and noise. How much of this extreme result reflects stable ability versus temporary fluctuation? If the answer is “partly noise,” expect regression.

The uncomfortable implication: exceptional performance often doesn’t last, and poor performance often doesn’t persist. The hero of this quarter may disappoint next quarter. The failure may rebound. The world is less stable than our narratives suggest, and more of what happens is luck than we want to believe.