Extreme outcomes tend to be followed by less extreme ones. This single statistical fact is responsible for more misallocated resources, more false confidence, and more broken strategies than any other concept in probability.
Regression to the mean describes the phenomenon whereby a variable that is extreme on its first measurement will tend to be closer to the average on its second measurement — not because of any causal force pulling it back, but because extreme values are, by definition, partly the product of luck, and luck does not persist. The tall parent's child is still likely tall, but less extremely so. The fund manager who crushed the market last year will probably beat it again, but by a smaller margin. The team that won fourteen consecutive games will probably keep winning, but not fourteen more in a row. The mathematics are not optional. They are a consequence of any system where outcomes are influenced by a combination of skill and randomness.
Francis Galton discovered the phenomenon in the 1880s while studying hereditary height. He measured the heights of 928 adult children and their parents and found that children of unusually tall parents were, on average, shorter than their parents — and children of unusually short parents were taller. Galton initially called it "regression toward mediocrity," a phrase that inadvertently captured both the statistical truth and the psychological resistance it provokes. The finding unsettled him. It seemed to imply that humanity was converging toward a grey uniformity. It was not. The distribution of heights stayed the same across generations — the variance was preserved. What changed was which individuals occupied the extremes. The extreme position was not heritable because it was partly the product of chance factors that did not repeat.
The mathematics are straightforward. Any measured outcome can be decomposed into a stable component (skill, genetics, structural advantage) and a variable component (luck, noise, random error). When the variable component contributes to an extreme measurement, it is overwhelmingly likely that the variable component was favourable — because unfavourable randomness combined with the stable component would have produced a less extreme result. On the next measurement, the variable component is drawn independently. It might be favourable again, but the expected value of any random variable is its mean, not its previous realisation. So the expected second measurement is closer to the overall average than the first. The regression is not toward the individual's true ability. It is toward the population mean, because the noise that inflated the first measurement has nowhere to go but back.
The concept sounds elementary, and it is — mathematically. The difficulty is entirely psychological. Humans are narrative creatures. When a CEO delivers an extraordinary quarter, the board constructs a causal story: the new strategy worked, the team executed brilliantly, the product resonated. When the next quarter is merely good, another causal story emerges: the market shifted, execution slipped, focus was lost. Both stories may contain elements of truth. But the most parsimonious explanation — that an extreme outcome was followed by a less extreme one because that is what extreme outcomes do — never makes it into the boardroom. Regression to the mean has no protagonist, no villain, no narrative arc. It is statistically inevitable and psychologically invisible.
Daniel Kahneman identified regression to the mean as one of the most consequential cognitive blind spots in decision-making. In Thinking, Fast and Slow, he recounted a pivotal moment from his work with Israeli Air Force flight instructors. The instructors told Kahneman that praise after an exceptionally good landing was typically followed by a worse landing, while criticism after a poor landing was followed by an improvement. They concluded — reasonably, by their experience — that criticism was more effective than praise. Kahneman recognised the error instantly: the pattern was pure regression to the mean. An unusually good landing would be followed by a more typical one regardless of what the instructor said, and an unusually bad landing would likewise be followed by improvement. The instructors' feedback was irrelevant. But the temporal coincidence between punishment and improvement created a powerful and entirely false causal belief that punishment works and praise backfires. The instructors were not stupid. They were victims of a statistical phenomenon that mimics causation so convincingly that it fools experts in every domain.
The phenomenon operates everywhere outcomes combine ability with randomness — which is to say, everywhere. Sports performance, academic test scores, investment returns, business metrics, medical treatments, crime rates, employee evaluations, student grades, weather patterns, and political approval ratings all exhibit regression to the mean. The phenomenon does not require any change in underlying quality. It does not imply that the system is deteriorating or improving. It is a mathematical artefact of measurement in the presence of noise, and it will appear in any dataset where the correlation between successive measurements is less than perfect.
The consequences for decision-making are severe. Companies fire managers after a bad year that was partially driven by noise and hire replacements who benefit from regression and get credit for the improvement. Sports teams trade players after a slump and the replacement performs better — not because of superior talent, but because the original player's slump was an extreme measurement that would have regressed anyway. Governments implement policies after a crisis — a spike in crime, a disease outbreak, a market crash — and claim credit when conditions improve, when the improvement was baked in by the same statistical logic that guaranteed the extreme would not persist.
Consider the scale of the distortion in medicine alone. Before the widespread adoption of randomised controlled trials, most medical treatments were evaluated by administering them to patients who were at their worst — because that is when patients seek treatment — and observing whether they improved. They almost always improved. Not because the treatments worked, but because extreme symptoms regress toward the patient's baseline regardless of intervention. Bloodletting persisted for two thousand years partly because physicians observed that patients who were bled tended to get better afterwards. They did get better. They would have gotten better without the bleeding. The regression was invisible because there was no control group — no untreated patients whose identical improvement would have revealed the treatment as inert. The randomised controlled trial, developed by Ronald Fisher and applied to medicine by Austin Bradford Hill in the 1948 streptomycin trial, was specifically designed to neutralise regression to the mean by ensuring that both treated and untreated groups regressed equally, so that any remaining difference could be attributed to the treatment. The entire architecture of modern evidence-based medicine is, at its core, a system for distinguishing genuine treatment effects from regression artifacts.
The financial markets provide an equally consequential theatre. The mutual fund industry's most reliable pattern is not alpha generation — it is regression to the mean in fund performance. S&P Global's SPIVA scorecards consistently show that funds in the top quartile over one period are no more likely to remain in the top quartile over the next period than chance would predict. The "hot hand" in fund management is almost entirely a regression illusion: the fund that outperformed was partly skilled and partly lucky, and the lucky component does not persist. Yet the entire fund marketing apparatus — the five-star ratings, the performance advertisements, the "top fund" lists — is built on presenting extreme recent performance as if it were a reliable predictor. Investors allocate billions of dollars annually to funds at the peak of their noise component, and then experience the inevitable regression as a personal loss.
The Israeli flight instructor story is not an anecdote. It is a template for how regression to the mean systematically distorts cause-and-effect reasoning in every institution that evaluates performance, allocates resources, or intervenes after extreme outcomes. The cost is not merely intellectual. It is operational: organisations repeatedly punish people for bad luck, reward people for good luck, and build their entire incentive systems around a statistical illusion.
The depth of the illusion is hard to overstate. Kahneman himself called regression to the mean "the most important lesson I have ever learned" — a remarkable statement from the psychologist who also identified loss aversion, anchoring, and the availability heuristic. He ranked regression above all of them because of its unique combination of ubiquity and invisibility: it operates in every domain, affects every decision that follows an extreme outcome, and is almost never recognised in real time by the people it is affecting.
Section 2
How to See It
Regression to the mean is operating whenever an extreme measurement is followed by a less extreme one and someone attributes the change to a cause other than statistics. The signal is not the change itself — it is the confident causal explanation offered for a change that required no explanation at all.
The pattern is remarkably consistent across domains. Something extreme happens — a spectacular quarter, a terrible season, a viral product launch, a health crisis. An intervention follows. Conditions move back toward normal. The intervention receives credit. The statistical inevitability receives none. Once you learn to see this pattern, you will find it operating in virtually every evaluation system you encounter:
Investing
You're seeing Regression to the Mean when a fund manager who returned 40% last year returns 12% this year, and the financial press writes articles about "what went wrong." Nothing went wrong. A 40% return in a market with 10% average annual returns contains roughly 30 percentage points of performance that must be explained by either exceptional skill or exceptional luck. If the manager's true skill is 15% alpha, the remaining 15% was noise — and noise does not repeat. The expected second-year return was always going to be closer to the mean. Jim Simons's Medallion Fund was the rare exception that sustained extreme returns across decades — and even Simons acknowledged that separating signal from regression noise was the central challenge of quantitative investing.
Technology
You're seeing Regression to the Mean when a product feature launches to extraordinary engagement metrics in Week 1, the team celebrates, and engagement drops 30% by Week 4. The team diagnoses the decline: novelty wore off, the algorithm deprioritised it, competitors copied it. The simpler explanation: the early adopters who discovered the feature in Week 1 were the most engaged segment of the user base. Their behaviour was extreme relative to the overall population. As the feature reached a broader audience, engagement regressed toward the population mean. The feature did not deteriorate. The measurement moved from a self-selected extreme sample to a representative one.
Sports
You're seeing Regression to the Mean when a rookie athlete has a spectacular first season, appears on magazine covers, signs a massive extension, and then "disappoints" in Year 2. The Sports Illustrated cover jinx — the folk belief that appearing on the cover predicts a decline — is regression to the mean wearing a superstitious costume. Athletes appear on covers because they just did something extraordinary. Extraordinary performance, by definition, contains a larger-than-usual luck component. The luck component regresses. The cover did not cause the decline. The selection criterion guaranteed it.
Personal life
You're seeing Regression to the Mean when you try a new productivity system during a week where your output is unusually low, and the following week is better, confirming the system "works." Your unusually low week was extreme — it would have been followed by a more typical week regardless of intervention. The productivity system gets credit for regression. This is the mechanism behind most self-help testimonials: people adopt new practices at their lowest point (because that is when they are motivated to change), and the subsequent improvement that was statistically guaranteed gets attributed to the practice.
Section 3
How to Use It
Decision filter
"Before attributing a change in performance to any cause — a new strategy, a personnel decision, an intervention, a punishment, a reward — ask: was the previous measurement extreme? If yes, the most likely explanation for the change is regression to the mean, not whatever cause is being claimed. The burden of proof is on the causal story, not on the statistics."
As a founder
Your metrics will regress. The question is whether you build systems that account for regression or systems that panic every time it happens. A product launch that generates exceptional first-week metrics is not evidence that the product is exceptional — it is evidence that the first-week sample is non-representative. Build evaluation windows long enough that regression has time to operate before drawing conclusions. The founder who changes strategy after one bad month and changes again after one good month is not responding to signal. They are chasing regression artifacts across a noisy landscape.
The hiring implication is equally important. The candidate who performed spectacularly in an interview may be genuinely exceptional — or may have had an exceptional day. The candidate who stumbled may have had a bad day. If your hiring process consists of a single high-stakes evaluation, regression to the mean guarantees that your best hires will disappoint and your rejected candidates include people who would have been outstanding. The fix is multiple independent measurements: more interviews, work trials, reference checks. Each additional measurement reduces the noise component and makes the signal more visible.
As an investor
Regression to the mean is the mathematical foundation of value investing. When a company's earnings are temporarily depressed — by a product cycle trough, a one-time charge, or a cyclical downturn — its stock price falls to reflect the depressed earnings. If the depression is temporary (driven by noise rather than permanent impairment), earnings will regress toward the company's long-run average, and the stock price will follow. Warren Buffett's career is substantially a career of buying regression: identifying companies whose current performance is well below their sustainable mean and waiting for the inevitable reversion.
The inverse error is equally costly. Extrapolating an exceptional quarter or year into a permanent improvement — and paying a premium multiple for it — is a bet against regression. The company that grew revenue 60% last year because of a pandemic tailwind, a competitor's stumble, and a one-time contract is not a 60% grower. The mean to which it will regress is the sustainable growth rate, and paying a price that assumes the extreme persists is the precise error that regression to the mean punishes most severely.
As a decision-maker
The most important operational application is in performance evaluation. Any evaluation system based on a single period of measurement — annual reviews, quarterly earnings, monthly dashboards — is contaminated by regression to the mean. The employee who had the best quarter may have the same underlying ability as the employee who had the third-best quarter, with the difference explained entirely by noise. Ranking and rewarding based on single-period performance systematically overpays luck and underpays skill.
The correction is structural: evaluate over longer time horizons, weight multiple independent observations, and compare performance to a baseline that accounts for contextual factors. The manager who says "your numbers were down last quarter — what happened?" is asking a question that may not have a meaningful answer. The numbers might have been down because last quarter's numbers were up, and the up was the anomaly, not the down.
Common misapplication: Assuming everything regresses. Regression to the mean operates on the noise component of an outcome, not the signal. A company that has built genuine structural advantages — network effects, switching costs, regulatory moats — may sustain extreme performance indefinitely because the extreme outcome is driven by skill, not luck. The discipline is distinguishing between extreme outcomes with high noise content (which will regress) and extreme outcomes with high signal content (which may persist). Compounding effects in particular can sustain and even amplify extreme performance, creating a genuine tension with regression logic that must be resolved case by case. The question to ask is not "will this regress?" but "what fraction of this extreme outcome is noise?" If the answer is high, expect regression. If the answer is low — because the outcome is driven by identifiable, durable structural advantages — regression may never come.
A second trap: using regression to the mean as an excuse for inaction. "It will regress on its own" is not a strategy when the extreme outcome is driven by a genuine causal change — a new competitor, a regulatory shift, a technological disruption. Regression operates on noise. It does not operate on structural breaks. The skill is knowing which is which — and that skill requires domain knowledge, not just statistical sophistication.
A third trap: applying regression logic to single observations. Regression to the mean is a statement about the expected value across many observations, not a guarantee about any individual case. The fund manager who outperformed massively last year might be genuinely extraordinary. The regression prediction says that most managers who outperform massively are partly lucky, and on average they will do less well next year. But "on average" and "this specific person" are different claims. The discipline is maintaining appropriate uncertainty about individual cases while recognising the strong base rate of regression across the population.
Section 4
The Mechanism
Section 5
Founders & Leaders in Action
Regression to the mean is the invisible statistical current that runs through every evaluation, every investment, and every performance review. The leaders below did not merely understand the phenomenon intellectually — they built systems, portfolios, and decision frameworks that accounted for it structurally. Each recognised that in domains where outcomes combine skill with noise, the single-period measurement is a dangerously unreliable guide to underlying quality, and that the institutions that mistake regression for causation will systematically misallocate their most valuable resources.
The common pattern is a refusal to overreact to extreme outcomes in either direction — a statistical discipline that looks like patience from the outside but is actually a rigorous application of probability theory to real-world decision-making. What unifies these cases is not passivity. Each of these leaders acted decisively — Buffett bought aggressively during panics, Simons deployed capital into validated patterns, Thorp sized bets with mathematical precision. The discipline was in acting on the signal while ignoring the noise, and in building evaluation systems long enough that regression had time to reveal what was real and what was artefact.
Buffett's entire investment philosophy is, at its mathematical core, a regression-to-the-mean strategy. Value investing — buying assets whose price has fallen below intrinsic value — is a bet that the current price reflects temporarily depressed earnings or sentiment that will regress toward the company's long-run economic average. When Buffett bought American Express in 1964 after the salad oil scandal, the stock had dropped 50% because of a fraud committed by a single client. The scandal was noise — it had no bearing on American Express's structural competitive advantages: the charge card network, the merchant relationships, the brand trust. Buffett bought the regression. The stock recovered and then some.
The discipline operates in reverse as well. Buffett has consistently refused to buy companies whose current performance is extreme on the upside unless the structural advantages justify sustained outperformance. His avoidance of technology stocks during the dot-com bubble was not technophobia — it was a recognition that the extreme revenue growth rates of 1999 contained an enormous noise component (speculative spending, unsustainable customer acquisition costs, accounting inflation) that would regress brutally. The companies whose revenue was 90% signal and 10% noise — like the ones Buffett held — regressed modestly. The ones whose revenue was 10% signal and 90% noise regressed to zero.
Buffett's annual letters repeatedly emphasise multi-year evaluation windows rather than single-quarter assessments — a structural accommodation of regression to the mean. "We've long felt that the only value of stock forecasters is to make fortune tellers look good," he wrote. The joke is aimed at extrapolation, which is the precise cognitive error that regression to the mean punishes: assuming that extreme recent performance is a reliable guide to future performance.
Jim SimonsFounder, Renaissance Technologies, 1988–2020
The central problem of quantitative trading is distinguishing genuine statistical patterns from regression artifacts. A strategy that backtests well may have captured a real signal — or it may have identified a period of extreme performance in a noisy dataset that will regress when capital is deployed. Simons built Renaissance Technologies' Medallion Fund on the discipline of separating signal from regression noise with a rigour that had no precedent in finance.
The process was specific: Renaissance reportedly demanded that any identified pattern be tested across multiple independent time periods and market regimes. A pattern that appeared only in one dataset was treated as a probable regression artifact — an extreme value in the noise that happened to look like signal. Patterns that persisted across independent samples were more likely to reflect genuine statistical structure. The methodology was essentially a regression-to-the-mean filter: by testing on multiple samples, Renaissance ensured that only the signal component — the part that would not regress — survived to receive capital allocation.
The deeper insight was in position sizing. Even validated patterns contain a noise component that will produce periods of extreme positive and negative performance. Simons sized positions so that the inevitable regression periods — the drawdowns when a genuine pattern produced temporarily negative returns — would not impair the fund's ability to keep trading. The discipline was not about avoiding regression to the mean. It was about structuring the portfolio so that regression to the mean could not be fatal. The fund's 66% average gross annual return over three decades reflects a system that extracted signal systematically while allowing noise to regress without consequence.
Thorp understood regression to the mean from first principles in a way that most finance practitioners never achieve — because he learned it first in a domain where the mathematics are transparent: blackjack. In Beat the Dealer (1962), Thorp's card-counting system exploited the fact that the composition of the deck shifts the house edge toward the player under specific conditions. But the edge was small — typically 1–2% — and the variance was enormous. A skilled card counter could lose money for weeks despite having a genuine mathematical advantage, because the short-term results were dominated by noise that would regress.
The lesson transferred directly to financial markets. When Thorp launched Princeton Newport Partners in 1969, he applied the same framework: identify genuine edges with mathematical precision, size positions so that the noise component (short-term losses) could not destroy the portfolio, and wait for the signal to dominate over a sufficient number of trials. Thorp's convertible arbitrage strategies had edges of a few percentage points — modest by the standards of most investors' aspirations. But because Thorp understood that returns are a combination of edge and noise, and that the noise component regresses while the edge component compounds, he could extract extraordinary long-term returns from apparently modest advantages.
The key insight was that most investors abandon strategies during regression periods — they sell after drawdowns, switch approaches after underperformance, fire managers after bad years. Each abandonment is a response to regression that forfeits the long-term edge. Thorp's mathematical training made him regression-immune: he sized positions small enough that regression periods were survivable and held strategies long enough that the signal emerged from the noise. Princeton Newport Partners delivered 19% annualised returns over nearly two decades with minimal drawdowns — a record built not on avoiding regression but on structuring the portfolio so that regression was irrelevant to long-term outcomes.
Charlie MungerVice Chairman, Berkshire Hathaway, 1978–2023
Munger's contribution to regression-to-the-mean thinking was primarily diagnostic: he used the concept as a filter for separating genuine competitive advantage from temporary outperformance that would regress. When evaluating a business, Munger asked a question that most analysts skip: is this level of performance sustainable, or is it extreme and likely to regress?
The distinction was operationally critical. A company earning 30% return on capital might be genuinely exceptional — possessing structural advantages (brand, network effects, switching costs) that sustain supranormal returns — or temporarily lucky: benefiting from a cyclical peak, a competitor's stumble, or a one-time demand shock. Munger's analytical framework demanded evidence of the structural mechanism before accepting that extreme performance reflected skill rather than noise. "It's not enough that the numbers look good," he told shareholders. "You have to understand why the numbers are good and whether the reasons will persist."
Munger extended the logic to human evaluation with characteristic bluntness. He frequently warned against promoting employees based on one exceptional year: "If someone has one great year, the question is whether the year was great because the person was great or because the circumstances were great. If you promote people based on one great year, you will fill your organisation with people who were lucky once and then regress to exactly the mediocrity they always were." The advice was pure regression to the mean, expressed in the language of organisational design rather than statistics. Munger's insistence on long track records — he and Buffett preferred CEOs with decades of performance data — was a structural accommodation of the mathematical reality that single-period measurements are contaminated by noise that will not persist.
Section 6
Visual Explanation
Regression to the Mean — Why extreme outcomes are followed by less extreme ones, and why the change requires no causal explanation
Section 7
Connected Models
Regression to the mean intersects with models that either amplify its effects, create productive tension with its implications, or represent the strategic consequences of understanding it properly. Some models reinforce regression thinking by providing complementary frameworks for separating signal from noise. Others create tension by describing forces — compounding, narrative construction — that seem to defy or obscure the statistical reality. And some represent the natural next steps: the investment and portfolio strategies that become available once regression is properly understood.
The six connections below map how regression awareness propagates through adjacent decision-making frameworks. Two models sharpen regression thinking by explaining why the phenomenon is so pervasive and so often misinterpreted. Two create genuine intellectual friction by describing dynamics that can override or mask regression. Two translate regression awareness into actionable strategy — investment approaches and portfolio structures that exploit the phenomenon rather than falling victim to it:
Reinforces
Correlation vs Causation
Regression to the mean is one of the most prolific generators of false causal attribution. When an intervention follows an extreme outcome and the outcome regresses, the intervention receives causal credit that it did not earn. Speed cameras installed after a spike in accidents "cause" the subsequent decline. A new CEO hired after a terrible year "causes" the subsequent recovery. A medical treatment administered after a flare-up "causes" the improvement. In each case, the temporal coincidence between intervention and regression creates a correlation that is mistaken for causation. Understanding regression to the mean is therefore a prerequisite for accurate causal reasoning — it eliminates the single largest category of false causal claims in performance evaluation, policy analysis, and medical treatment assessment.
Reinforces
Goodhart's Law
Goodhart's Law — "when a measure becomes a target, it ceases to be a good measure" — intersects with regression to the mean through the dynamics of selection. When organisations select individuals or teams based on a single extreme metric, they are selecting partly on signal and partly on noise. The noise component will regress on subsequent measurements, making the selected group appear to deteriorate — which prompts management to intensify the metric-targeting, which increases gaming, which further contaminates the signal. Regression to the mean and Goodhart's Law together explain why performance management systems that reward top-quartile results in one period systematically produce disappointment in the next: the top quartile was partly populated by luck, and the luck does not repeat.
Tension
Compounding
Section 8
One Key Quote
"Success = talent + luck. Great success = a little more talent + a lot of luck."
— Daniel Kahneman, Thinking, Fast and Slow (2011)
Section 9
Analyst's Take
Faster Than Normal — Editorial View
Regression to the mean is the most important statistical concept that almost nobody applies correctly in practice. The idea is taught in every introductory statistics course and promptly forgotten in every boardroom, investment committee, and performance review. The gap between knowing the concept and applying it is wider for regression to the mean than for almost any other mental model, because the human need for causal narratives is so powerful that it overrides statistical literacy in real time.
The most expensive manifestation is in CEO hiring and firing. Research by Markus Jenter and Fadi Kanaan found that CEO turnover is strongly predicted by stock price performance — but the stock price decline that triggers a firing is substantially driven by industry-wide and market-wide factors that no CEO controls. Companies fire CEOs for bear markets, then hire replacements who receive credit for the subsequent recovery that was statistically inevitable. The new CEO implements a "turnaround plan" — which is typically a collection of changes that would have happened anyway — and the board congratulates itself on the leadership change. The entire cycle is regression to the mean dressed in a narrative costume. The cost is not merely the severance package and search fees. It is the organisational disruption, strategic discontinuity, and institutional knowledge loss that accompany every unnecessary leadership transition.
The second most expensive manifestation is in government policy evaluation. Policies are implemented in response to crises — crime spikes, disease outbreaks, economic downturns — which are, by definition, extreme values in noisy time series. The policy is followed by improvement. The improvement is attributed to the policy. The policy becomes permanent. The phenomenon is so well-documented in criminology that it has its own name: the "regression fallacy in crime statistics." A city experiences a spike in violent crime, implements a policing initiative, and violent crime falls. The initiative gets credit. But the spike was extreme — it would have been followed by a decline regardless of intervention — and without a control group (a comparable city that didn't implement the initiative), the causal claim is pure regression attribution.
The subtlest version operates in personal performance evaluation. Every time you have an exceptionally productive week, the odds are strong that the following week will be less productive — not because you lost motivation or discipline, but because the exceptional week contained a favourable configuration of circumstances (fewer interruptions, easier tasks, higher energy) that will not recur identically. If you attribute the subsequent decline to a personal failing, you will attempt a correction (new routine, new tools, more discipline) that addresses a problem that does not exist. The correction may temporarily coincide with another uptick — regression again, in the other direction — and you'll credit the correction. The entire self-optimisation loop is built on regression artifacts that you are systematically misinterpreting as causal evidence.
Section 10
Test Yourself
Regression to the mean operates most deceptively when an intervention coincides with a naturally regressing extreme — making the intervention appear effective when it was irrelevant. These scenarios test your ability to distinguish genuine causal improvement from statistical regression, and to recognise when an apparent change is simply the mathematics of noise reasserting itself.
The key diagnostic skill is asking: was the prior measurement extreme, and was the selection or intervention triggered by that extremity? If yes, regression is the default explanation and any causal claim must clear a higher bar of evidence. If no — if the measurement was typical and the change was large — genuine causation becomes more plausible.
Is this mental model at work here?
Scenario 1
A pharmaceutical company runs a randomised controlled trial with 10,000 participants. Patients are randomly assigned to receive either the drug or a placebo. After 12 months, the drug group shows a 23% reduction in symptoms compared to the placebo group, with p < 0.001. The company concludes the drug is effective.
Scenario 2
A city's crime rate spikes 35% in one year. The mayor implements a new policing strategy. The following year, crime drops 20%. The mayor holds a press conference crediting the strategy and expands it citywide.
Scenario 3
A venture capital fund's top-performing analyst picked the fund's best investment last year — a company that returned 12x. The managing partner promotes the analyst to partner and gives them a larger allocation for this year's investments. This year's picks return an average of 1.8x.
Scenario 4
A basketball player averages 28 points per game over a full 82-game NBA season — up from 21 points per game the prior season. His coach implemented a new offensive scheme in the off-season that features the player more prominently.
Section 11
Top Resources
The essential reading spans statistics, psychology, investing, and experimental design. Kahneman provides the definitive treatment of regression's psychological invisibility. Galton's original paper is short, clear, and surprisingly modern. Secrist provides a cautionary tale of how even trained statisticians can misunderstand the phenomenon. Taleb translates the concept into the language of risk and randomness. Senn provides the technical toolkit for anyone designing experiments where regression is a confounder. Start with Kahneman for the cognitive framework, then read the empirical applications in investing and policy evaluation.
The intellectual progression matters: start with Kahneman for the psychology of why regression is invisible, move to Galton for the historical origin, read Secrist as a warning of how experts get it wrong, and finish with Taleb and Senn for the practical applications in finance and experimental design. Each resource illuminates a different facet: Kahneman explains why the error persists, Galton explains the mathematics, Secrist demonstrates the error in action, Taleb connects it to risk management, and Senn provides the technical defence.
Kahneman's Chapter 17 on regression to the mean is the single best treatment of the topic for a general audience. The Israeli flight instructor anecdote, the analysis of the Sports Illustrated cover jinx, and the systematic exploration of why regression is psychologically invisible are essential reading for anyone who evaluates performance, allocates resources, or makes decisions based on outcomes that combine skill with luck — which is everyone.
The paper that started it all. Galton's presentation to the Royal Anthropological Institute is remarkably readable for a 140-year-old statistical paper. The data on parent-child height relationships is presented with clarity, and the conceptual leap from empirical observation to statistical principle is visible on the page. Short, foundational, and a reminder that the most consequential ideas in statistics began as careful observation rather than mathematical abstraction.
The most famous misunderstanding of regression to the mean in the history of statistics. Secrist, an economics professor, spent years documenting that companies with extreme performance — both extremely good and extremely bad — tended to converge toward average performance over time. He concluded that American business was declining toward mediocrity. The statistician Harold Hotelling's devastating review pointed out that Secrist had simply rediscovered regression to the mean and mistaken it for an economic finding. The book is valuable precisely because the error is made by a credentialled expert with extensive data — proof that understanding the mathematics does not automatically produce correct interpretation.
Taleb's exploration of how humans systematically misinterpret random outcomes as skill — and how regression to the mean punishes the misinterpretation. The book's treatment of survivorship bias, alternative histories, and the role of luck in financial success is built on the same statistical foundation as regression to the mean, and Taleb's prose makes the concepts viscerally memorable. Essential for investors and founders who operate in domains where noise dominates signal in the short run.
Senn's paper in the International Journal of Epidemiology provides the most rigorous modern treatment of regression to the mean in clinical and policy contexts. The paper explains why regression to the mean contaminates before-and-after studies, why randomised controlled trials neutralise the effect, and how to estimate the magnitude of regression in any given dataset. Senn also addresses the common misconceptions — including the belief that regression implies convergence toward uniformity — with mathematical precision. Technical but indispensable for anyone designing experiments, evaluating interventions, or making resource decisions based on extreme performance measurements.
Compounding appears to defy regression to the mean — and in specific circumstances, it does. A company with genuine structural advantages (network effects, switching costs, brand equity) can sustain extreme performance not because it is lucky but because the advantages compound: each year of market dominance strengthens the position that produces the next year of dominance. In these cases, extreme performance is driven primarily by signal, and the noise component — which is the only part that regresses — is small relative to the total. The tension is real: regression to the mean says extreme outcomes are unsustainable, compounding says they can accelerate. The resolution is analytical: decompose the extreme outcome into its signal and noise components. If the signal component (structural advantage) dominates, compounding will sustain or amplify the extreme. If the noise component (luck, timing, one-time events) dominates, regression will prevail.
Tension
[Narrative](/mental-models/narrative) Fallacy
The narrative fallacy — the compulsion to construct coherent causal stories from random data — is the primary psychological mechanism through which regression to the mean becomes invisible. When performance regresses after an extreme outcome, the brain refuses to accept "it was statistically inevitable" as an explanation. System 1 demands a story: the CEO lost focus, the market turned, the team got complacent. The story feels explanatory and is almost always unnecessary. The narrative fallacy does not merely obscure regression — it actively fights against recognising it, because a world where extreme changes require no explanation is a world that feels dangerously random and uncontrollable. The tension between regression (which says many changes need no explanation) and the narrative fallacy (which demands an explanation for every change) is one of the deepest conflicts in human cognition.
Leads-to
Margin of Safety
Understanding regression to the mean directly enables the margin-of-safety framework. If you know that extreme negative outcomes tend to regress toward the mean, then buying assets during periods of extreme pessimism — when prices are well below intrinsic value — builds in a statistical tailwind: the regression that is likely to occur provides a margin of safety independent of the analytical precision of your valuation. Benjamin Graham's insistence on buying below intrinsic value is, mathematically, a regression-to-the-mean bet: the current price reflects extreme negative sentiment that will, on average, moderate. The margin of safety converts a statistical tendency into an investment discipline.
Leads-to
Barbell Strategy
Regression to the mean informs the barbell strategy by clarifying where to expect regression and where to expect persistence. The barbell — Nassim Taleb's recommendation to combine extreme safety with extreme risk while avoiding the middle — is partly built on regression logic. Middle-risk, middle-return positions are vulnerable to regression from both directions: their moderate outperformance may regress to mediocrity, and their moderate safety may prove insufficient in a crisis. The extremes of the barbell are chosen precisely because they have different regression properties: the safe portion is designed to be insensitive to regression (Treasury bills do not regress), while the risky portion is designed to benefit from the positive asymmetry when extreme positive outcomes occur. Regression to the mean provides the analytical foundation for understanding why the middle of the distribution is the worst place to be in many domains.
The investment implications are the ones I find most consistently underappreciated by sophisticated practitioners. Morningstar's research has repeatedly shown that mutual fund investors systematically underperform the funds they invest in because they buy after strong performance (near the peak of the noise component) and sell after poor performance (near the trough). The gap between time-weighted returns (the fund's return) and dollar-weighted returns (the investor's return) is approximately 1–2% annually — a direct tax levied by regression to the mean on investors who chase extreme recent performance. The investors are not irrational. They are doing exactly what every financial advertisement, every media story, and every natural cognitive impulse tells them to do: identify what is working and allocate toward it. The problem is that "what is working" is partly signal and partly noise, and the noise component — which is what made the recent performance extreme — will regress.
The interaction with sample size is critical and almost universally ignored. Small samples produce more extreme values than large samples — this is a mathematical fact that follows directly from the central limit theorem. The school with the highest test scores in a state is disproportionately likely to be a small school, because small samples have higher variance. The fund with the best three-year record is disproportionately likely to be a small fund, for the same reason. When resources are allocated based on extreme performance from small samples — promoting the teacher whose small class scored highest, investing in the fund with the best short track record — the subsequent regression is brutal because the noise component in small samples is enormous.
The AI and machine learning era amplifies the regression problem rather than solving it. Modern ML systems are trained on historical data where extreme outcomes are present. A model trained to predict which employees will be top performers, using last year's top performers as training data, will learn the characteristics associated with extreme outcomes — but those characteristics include the noise component that will regress. The model replicates the regression error at scale: it systematically overweights the features of people who were lucky and underweights the features of people who were skilled but unlucky. The predictions look precise. The regression is baked into every one of them. The organisations that understand this use ML predictions as one signal among many and demand multiple independent measurements before making high-stakes decisions. The organisations that don't will automate the regression fallacy at a speed and scale that no human decision-maker could achieve.
My operational recommendation: build a regression-to-the-mean checklist for any decision triggered by extreme performance. Before acting on an extreme outcome — positive or negative — ask: (1) How extreme is this relative to the historical distribution? (2) What is the estimated noise-to-signal ratio in this measurement? (3) What would you expect the next measurement to be if no intervention occurred? If the expected next measurement is substantially closer to the mean than the current one, any intervention you undertake will receive false credit for the regression that was going to happen anyway. This does not mean you should never intervene. It means you should evaluate interventions against a regression-adjusted baseline, not against the extreme value that triggered the intervention.
The ultimate test of regression awareness is emotional, not intellectual. Everyone nods when the concept is explained. Almost nobody applies it when the extreme outcome is personal — when it is their team that overperformed, their investment that crushed it, their strategy that delivered the breakthrough quarter. Regression to the mean requires accepting that some of your best outcomes were partly luck and some of your worst outcomes were partly bad luck. That acceptance is threatening to the ego in a way that no statistical argument can fully overcome. The leaders who internalise it — Buffett, Simons, Thorp, Munger — built systems that enforce regression awareness structurally, through long evaluation windows, multiple measurement points, and explicit decomposition of outcomes into skill and noise. They did not rely on willpower to override the narrative instinct. They built institutions that did it for them.
The concept has a beautiful irony at its core. Regression to the mean is named after Galton's observation that extreme parents have less extreme children. But Galton's own intellectual legacy — founding statistics, correlation, and regression analysis — has not regressed at all. It has compounded across a century and a half, becoming more central to every domain of human decision-making with each passing decade. The concept he discovered explains why most extreme outcomes fade. His own extreme contribution is the exception that proves the structural point: when the extreme is driven by genuine signal — a real intellectual breakthrough, a true competitive advantage, an authentic structural change — regression has nothing to operate on. The skill is knowing the difference.
Scenario 5
A school district identifies the ten lowest-performing schools based on standardised test scores and enrols them in an intensive improvement programme with additional funding, coaching, and curriculum changes. After one year, eight of the ten schools show improved scores. The district superintendent presents the results as evidence the programme works and requests funding to expand it.
