statistical
Regression to the Mean
Also known as: Mean Reversion, Regression
Regression to the mean is the statistical tendency for extreme observations to be followed by less extreme ones, moving closer to the long-run average. Hot streaks cool and cold streaks warm, in any process driven partly by chance.
Regression to the mean is one of the most misunderstood concepts in sports betting. Extreme performances in either direction are usually a mix of true skill and randomness, and the random component does not persist. Across a longer sample, observed performance moves toward the underlying true rate.
The key insight: regression is not a force that “corrects” past results. It is a property of how repeated random sampling works. A hot streak doesn’t make a cold streak more likely (that’s the gambler’s fallacy); it just means the next sample is drawn from the player’s true distribution, which has a mean below their recent hot output.
Worked example. An NBA player has a career true scoring projection of 18 points per game. Across his last 5 games, he’s averaged 28. The naive read is “he’s hot, ride him.” The disciplined read: his last 5 games are some mix of (a) genuine skill change (small), (b) favorable matchups (medium), (c) variance (large). The over-projection for his next game should be closer to 20-21 than to 28. Recreational money chases the hot streak; sharp money fades it when the line has moved past the regressed projection.
Elite Picks builds regression into every projection through Bayesian shrinkage. Short-window stats (last 5 games) are blended with longer-window priors (career or season averages) in proportion to the reliability of each. The smaller the recent sample, the more it gets shrunk toward the prior. This single technique is responsible for a meaningful share of the model’s edge against recreational bettors who anchor on whatever just happened.
Example
A player hits 50% from three across his first 10 games against a career 36% true rate. Over his next 50 games, his expected three-point percentage is closer to 36% than to 50%, because the 10-game sample is mostly noise on top of his true skill.