Regression to the mean: why extreme seasons rarely repeat
The best and worst results in any season are usually part skill and part luck, so they tend to move toward average the next year. This guide explains why, how much, and how fans misread it.
A player has a career year. His shooting percentages jump, his scoring climbs, he makes his first All-Star team, and the following season he looks like a different, lesser player. Nothing is wrong with him; he has not lost a step. A team wins far more games than anyone predicted, and the next year it slides back to the pack. A struggling shooter hits a rough patch of 15 games and then, without any change in routine, drifts back toward his usual self. Observers reach for explanations: complacency, a sophomore slump, a curse, an injury nobody noticed, a coach who changed something.
Often, none of these explanations is needed. The outcome is a case of one of the most important and least appreciated ideas in statistics: regression to the mean. When a result is unusually high or low, the next result will, on average, be closer to normal. The reason is not that anything pulls the player or team back. It is that the extreme result was produced by a mixture of skill and luck, and the luck does not repeat.
This article explains regression to the mean in plain terms, shows how it works with a simple example, lists where it shows up in basketball, and points out the traps it creates for commentators, coaches and fans. It works closely with our guide to small samples: small samples produce extreme results, and extreme results regress.
The idea in one picture
Imagine that every performance is the sum of two things: the player's real, underlying ability and some amount of luck. Ability changes slowly, if at all, within a season. Luck is random and different every time, sometimes positive and sometimes negative.
Now picture the best performances in a group. The ones at the very top got there by having high ability, high luck, or both. Since high-ability players are rarer than players who got a bit of good luck, a large share of those at the top got there at least partly by being lucky. The next time those same people perform, their ability will be the same, but their luck will be new. On average, the luck will be neutral, not good. So their results fall back.
The same story runs in reverse at the bottom. The worst performances include players with low ability and players who were unlucky, and when the luck resets, the unlucky ones improve.
Nothing in this account depends on psychology, effort or aging. It is arithmetic about selection. If you pick the extremes of a noisy measurement, you are picking part luck, and luck does not follow you into the next period.
A worked example
Consider three-point shooting. Suppose the league average is about 36 percent, and consider a player who shoots 43 percent on 400 attempts in a season. We can ask what to expect from him the next year.
His 43 percent might reflect a true ability of 43 percent, or it might reflect a true ability of 38 percent with seven points of good luck, or something in between. How should we weigh these possibilities? The standard error on 400 attempts is about 2.4 percentage points, so a seven-point gap is nearly three standard errors, a large deviation. If true three-point abilities were spread across the league with a typical spread of a few points, then most players who post a 43 percent mark are better than average but not as good as 43 percent. A reasonable forecast would put his true level somewhere around 39 or 40 percent.
Here is the general recipe. The best prediction for the next result is the league average plus a fraction of the gap between the observed result and the league average. The fraction is called the reliability of the measure, and it depends on how much of the variation between players is real and how much is noise. For a high-reliability statistic such as height, the fraction is nearly one, and there is no regression. For a low-reliability statistic such as the result of a single game, the fraction is tiny, and the best prediction is close to the average. For three-point percentage over a season, the fraction is moderate, perhaps one half or a little more, which pulls our 43 percent shooter to roughly 39 or 40 percent. The numbers here are illustrative, but the pattern is dependable.
The upshot: do not predict that a hot shooter will keep shooting at the same rate. Predict that he will be good, but not that good.
Where regression shows up in basketball
Individual shooting
Three-point percentage, free-throw percentage on small totals, and shooting from a particular zone all regress strongly because they contain a lot of randomness. A player who leads the league in three-point percentage one season is unlikely to lead it again, even if he remains an excellent shooter. The same applies to the other end: the shooters with the worst percentages tend to improve. Looking at a leaderboard, the safest bet is that the extremes will soften. The cautions in how to read a leaderboard: whose numbers are meaningful apply here.
Team records
A team's win total is a mixture of its real strength and luck, especially luck in close games. Teams that win far more games than their scoring margin predicts tend to see their records fall the next year. Teams that win far fewer than their margin predicts tend to rise. This is why Pythagorean wins are a better forecast of next year's record than the record itself. It is also why a good team that finishes with a record that was boosted by luck is often described as having taken a step back when it has merely come back to its level.
Defensive numbers
Opponent shooting percentages are among the noisiest numbers in basketball. A defense that holds opponents to unusually poor three-point shooting often looks different the next year, even with the same players and scheme. The better predictor of future defensive performance is the quality of shots the defense allows, not whether opponents happened to make them. See expected points and shot quality for how analysts separate the two.
Awards and honors
Players are selected for honors based on their results, and those results include luck. The player who wins a Most Improved Player award, by definition, had a season far better than the previous one. Some of that improvement is real development; some is luck. The following year, he is likely to retain much of the improvement but not all of it. Similarly, players selected to the All-Star team based on half a season of numbers are, on average, slightly less spectacular in the second half.
Rookies and the sophomore slump
The so-called sophomore slump is partly regression. Rookies who perform exceptionally well in their first seasons, particularly those who shoot well above their college marks, are often shooting a bit above their true level. The second season returns them toward it. Real factors, such as opponents scouting a player's tendencies, also play a part, but regression is enough to produce a drop-off without any other cause.
Coaching and streaks
When a team goes through a poor stretch and the coach changes the starting lineup or calls a harsh practice, results often improve. It is tempting to credit the change. But a poor stretch, by definition, is a low point, and a low point tends to be followed by something closer to normal whether or not anything is done. The same trap catches people in the opposite direction: a coach who praises a player after a great game and sees a worse one next time may conclude that praise backfires. The psychologist Daniel Kahneman famously described how instructors in a pilot-training program drew exactly this mistaken lesson about praise and criticism. Praise followed great performances, and the next performance was, on average, worse; criticism followed poor ones, and the next was better. The pattern was regression, not the effect of the words.
A thought experiment with a gym full of shooters
Here is a way to feel regression in your bones. Picture a gym with 200 shooters, every one of them exactly a 36 percent three-point shooter, no better and no worse. Have each take 100 threes. Because of luck alone, their totals will scatter: most will land between about 27 and 45 makes, a few above 45, a few below 27. Rank them. The top ten will have shot around 45 percent, and anyone watching might declare them the best shooters in the gym.
Now have the same ten shoot again. They are still 36 percent shooters, so on average they will make 36. The group's percentage collapses from 45 to 36 without anyone getting complacent, tired or cursed. There was no skill to preserve, only a lucky draw.
Real leagues are less extreme than that gym, because real shooters differ genuinely in ability. Some are 31 percent shooters and some are 40. When the best ten shooters of a real season are ranked, the group contains true 40 percent shooters who had a normal year, true 38 percent shooters who had a hot year, and a few true 36 percent shooters who had a very hot one. On a second round, the first group stays about where it was, the second slips a little, and the third falls hard. The group average declines, but by less than in the gym, because more of the original spread was real. Regression is the sum of all those small slips.
This also shows why the amount of regression depends on the statistic. The more of the spread across players is real, the less the top group falls; the more is noise, the more it falls.
| Kind of measure | How much of the spread is real | Expected regression |
|---|---|---|
| Height, wingspan | Nearly all | Almost none |
| Rebound rate, usage rate | Most of it | Small |
| Season three-point percentage | A moderate share | Noticeable |
| Opponent three-point percentage allowed | A small share | Large |
| A single game's box score | Very little | Nearly total |
The table is qualitative, but it gives the right instinct: trust measures tied to role and physical traits, and distrust measures tied to whether a small number of shots happened to drop.
Common misunderstandings
Regression does not mean that results become average. The player who hits 43 percent and is expected to hit 40 is still above average. The claim is that the extreme moves toward the middle, not that it reaches it. Good players stay good.
Regression is not a force or a law of compensation. A player who has been lucky is not "due" for bad luck. The next shot is independent of the previous ones. Regression operates in the aggregate: a group selected for extreme results will, on average, be less extreme next time, because the group contained a lot of luck. Each individual's new luck is simply drawn fresh.
Regression does not mean nothing is real. Some extreme results represent real talent. A player who dominates for years while others do not is not a case of luck. The point is to apportion credit between skill and luck, and the way to do that is to use more data, not to dismiss performance.
Regression is not the same as decline with age. A player's skills change over time, and the typical path is covered in our look at aging curves. Regression is an additional effect that applies even when ability is constant. To tell them apart, look at the sample size and the underlying inputs: if the shots were as good as before and only the results changed, regression is the likely culprit.
Regression applies in both directions. It is often discussed as a warning about hot starts, but it is an equally good reason for optimism about a team or player who has been unlucky. A team with a strong margin and a mediocre record is a candidate to improve. A shooter with strong form and a poor percentage is likely to recover.
How to use the idea
A short guide helps in practice.
- Identify the extreme. Whenever you see a very high or very low number, ask whether it is a product of a small sample or a noisy measure.
- Estimate the reliability. Think about how much of the variation among players or teams is real. Stats that depend on rare events, such as three-point percentage or close-game record, are low in reliability.
- Pull the number toward the average. Do not go all the way. If the measure is somewhat reliable, move about half the distance; if it is very noisy, move most of it.
- Look for inputs. If the underlying inputs, such as shot quality or minutes, changed in a plausible way, the extreme may be real. If the inputs are the same, the extreme is likely to fade.
- Be careful with narratives. When someone explains a reversion with a story about confidence, pressure, effort or a curse, ask whether regression alone would have produced the same result. Usually, it would.
- Wait for more data. The more games and shots you see, the less regression you need. Check the leaders and standings pages again after a larger sample has accumulated.
What regression teaches about forecasting
Regression to the mean is a lesson in humility about prediction. The best forecast for a player or team is almost never the last result extended into the future. It is a blend of the last result and something more stable: the player's longer record, the league average or the quality of the underlying process. Teams and analysts that adopt this habit tend to avoid overpaying for last season's peak and to spot undervalued players whose results lagged their inputs.
For fans, the idea offers a different pleasure: a calmer way of watching. When a role player goes 8 for 10 from three, you can enjoy it without expecting it to continue. When a star struggles for a month, you can recognize the likelihood that the slump says more about the calendar than about him. The best response to a surprising result is neither to dismiss it nor to build a theory around it, but to ask how much of it is likely to last.
What to take away
Extreme performances contain luck, and luck does not repeat. That is the whole of regression to the mean, and it explains a wide variety of basketball patterns, from the career-year hangover to the team that overperforms and then falls back. The right response is to forecast with a blend of the recent result and a more stable baseline, to be skeptical of stories that explain a reversion without noticing the arithmetic, and to apply the idea in both directions, to the surprising successes and the unlucky failures alike.