Regression to the Mean

Extreme outcomes—both amazing and terrible—tend to be followed by more average ones, simply because luck balances out over time. That unusually terrible performance was probably partly bad timing, bad luck, or an off day, so naturally you'll do closer to your usual level next time. This happens everywhere: sports teams after a winning streak, startups after explosive growth, even your mood after an especially good or bad day. Check this: Before assuming a cause, ask if the result could just be luck returning to normal.

Regression to the mean occurs whenever extreme events mix stable skill with random luck. An unusually bad performance usually involves bad luck, so the next performance—even without any intervention—will likely be better because luck averages out. The same happens in reverse: amazing results usually involve good luck, so what follows looks worse even if nothing changed.

You see this constantly. A startup grows 300% in year one (extreme luck plus talent), then only 50% in year two. Investors panic, but this is often normal regression, not failure. A pilot program shows miraculous results in one school, then modest gains elsewhere—the first school was probably just lucky on top of being good. A CEO gets praised after a record quarter, then fired when results "decline"—but that record quarter was partly lucky.

The practical trap is believing that every extreme needs an extreme explanation. When a struggling employee suddenly improves after coaching, the coach gets credit—even though the employee might have improved anyway. When a patient feels better after taking an herbal remedy, the remedy gets credit—even though symptoms often naturally improve. To avoid this trap, look at multiple data points before concluding that an intervention worked, and remember that the most extreme result is often the least repeatable.

Regression to the mean was first identified by Sir Francis Galton in the 1880s while studying heredity. He observed that tall parents tended to have children slightly shorter than themselves, while short parents tended to have children slightly taller—both groups regressing toward the average height. This statistical phenomenon appears whenever a variable is influenced by both stable factors and random variation.

• Galton (1886): Discovered that extreme parental heights produce offspring heights closer to the population average, establishing the foundational concept [1].
• Campbell (1969): Showed that pretest-posttest designs in education often mistakenly attribute regression to treatment effects, threatening experimental validity [3].

Regression to the mean only describes what happens on average—individual cases can still move further from the mean. The effect is strongest when outcomes involve substantial randomness; in highly controlled systems with little luck (like chess ratings), regression is weaker. Most importantly, regression doesn't prove that an intervention failed—it just means you can't credit the intervention without better evidence. Finally, regression can mask real effects: a genuine intervention might show weaker results than expected because luck was pulling in the opposite direction.