Regression to the Mean
An unusually great or unusually bad result is often followed by one closer to average, simply because chance played a role the first time. Mistaking this natural drift for real improvement or real decline is one of the most common errors in judging performance.
When part of a result depends on chance, an extreme outcome is usually followed by one closer to the average, whether or not anything changed.
Was this change real, or is it the ordinary drift that extreme results show on their own?
Performance reviews, coaching programs, sports analytics, quality control, hiring, investing, and checking claims about what caused an improvement.
Fast overview
Francis Galton first described this pattern in 1886 while studying the heights of parents and their children, and called it “regression toward mediocrity.” Whenever a measurement mixes a stable, real factor with a random or noisy one, the most extreme readings are the least likely to repeat exactly. Selecting a case because it is extreme all but guarantees the next measurement will look less extreme.
This is not a claim that nothing ever really changes. It is a caution: before crediting a program, a punishment, a coaching session, or a lucky streak with causing a shift, check how much of that shift regression alone would predict.
Mental Model Compass
A result stands out sharply from a group average or from someone's own usual performance.
Small sample sizes, one-time extreme scores, and claims that a single event explains a change.
How much of this extreme result could be chance, and what would the next measurement look like with no intervention at all?
Assuming a bounce-back proves an intervention worked, or that a decline proves a real, lasting slump.
Learn it progressively
Spot a result that is unusually high or low compared to the group, or to the person's own history.
Check how many observations produced that result, and how reliable this kind of measurement usually is.
Notice if the case was chosen precisely because it was extreme. That alone guarantees some regression on remeasurement.
Estimate how far back toward average the next result would likely move even with no changes at all.
Compare the actual change to the predicted drift. Only the difference beyond that drift is a candidate for a real cause, and a comparison group makes the comparison far more trustworthy.
Diverse examples
A branch's worst quarter
A store posts its lowest sales quarter ever. Leadership sends in a turnaround consultant, and next quarter improves. Some of that recovery would likely have happened anyway.
The sophomore slump
A rookie has a spectacular debut season built partly on a hot streak. A more ordinary second season is not proof of declining skill.
Coaching after the worst review
An employee with the lowest score on a noisy annual review gets extra coaching. A better score next year is expected even without it.
Tutoring the lowest scorers
Students who scored worst on one test, partly due to a bad day, tend to score higher on a retest regardless of tutoring.
Chasing last year's top fund
A fund manager with an exceptional year attracts new money. Performance often looks more ordinary the following year.
Enrolling patients at their worst
Patients often join a study when symptoms are most severe, so some natural improvement is expected even with a placebo.
Worked application
Five reps had a terrible first month. What happened next?
Five new sales reps posted the lowest call-quality scores on the team in month one, averaging 42 against a team average of 62. No special coaching was introduced. Here is what month two looked like.
Five reps posted the lowest month-one scores on the team.
Team average was 62. The bottom five averaged 42 in month one.
All five rose 13 to 17 points in month two, with no special program.
Crediting a manager's pep talk or a script change alone for the recovery.
A weak first month is a noisy signal. A rise back toward the team average is expected even with no intervention.
Compare a coached group with an uncoached group of similarly low first-month scorers before claiming coaching worked.
History and accuracy
Francis Galton coined the phrase while studying how the heights of children related to the heights of their parents. Unusually tall parents tended to have somewhat shorter children, and unusually short parents tended to have somewhat taller children, both drifting toward the population average. Galton's original term, “regression toward mediocrity,” described the statistical pattern, not a judgment about the people involved, and it later softened into the more neutral phrase used today.
Is this the same as mean reversion in investing?
They are related but distinct. Regression to the mean is a general statistical property of any measurement that mixes a stable and a noisy component. Mean reversion in markets is a specific, debated empirical claim about how asset prices behave over time, and it should not be treated as automatic proof of the broader statistical pattern.
Does this mean nothing ever really improves?
No. Real, lasting change happens constantly through skill development, better systems, and effective interventions. The model only asks that a change be checked against how much drift would have happened anyway, so that real causes get credited accurately instead of being confused with ordinary statistical noise.
How large does the effect tend to be?
It depends on how reliable the original measurement is. A noisier measurement, such as a single sales month or a single test, shows more regression. A more reliable measurement built from many observations, such as a full-year track record, shows less, because less of the original result was due to chance in the first place.
Common mistakes
The illusion of decline
Praising a rookie's hot streak, then treating an ordinary second season as proof of a real slump.
The illusion of improvement
Crediting a coaching session or a punishment for a bounce-back that regression alone would predict.
Ignoring sample size
Extreme results are far more likely when they come from a small number of observations.
Assuming full reversal
Regression predicts partial movement toward the average, never a swing to the opposite extreme.
Calling every extreme noise
Some extreme results reflect a real, durable skill or a real, ongoing problem, not chance.
Skipping the comparison group
Without an untreated group that was equally extreme, a claimed cause cannot be verified.
When not to use it
Use direct measurement when…
a stable, well-documented trait explains the result and the measurement carries little noise.
Use causal testing when…
you need to prove an intervention works, which requires an actual control or comparison group.
Use trend analysis when…
you are looking at a long, consistent pattern rather than a single standout data point.
Related models
Probabilistic Thinking
Estimate how much of an extreme result is likely chance before judging whether a change is real.
Survivorship Bias
Check whether the extreme cases you are studying were already selected for being extreme in the first place.
Second-Order Thinking
After accounting for natural drift, ask what the remaining real change implies going forward.
Try it yourself
Your AI-ready prompt
Save as skill
Turn this mental model into a reusable AI skill
Save as Skill converts this model into portable Markdown instructions that teach an AI when and how to apply Regression to the Mean, what questions to ask, what evidence to seek, what mistakes to avoid, and how to structure its analysis.
Save the skills you find useful and organize them into your own reusable library. The Markdown is intentionally model-neutral, so it can be used wherever an AI system accepts uploaded files, project knowledge, custom instructions, or reusable text context.