Attributing a natural return-to-average to whatever intervention happened to coincide with it.
Imagine you have a really, really bad day where everything goes wrong. Your mom gives you a lucky penny, and the next day is pretty normal. You might think the penny fixed things, but really, a terrible day is just unusual — most days are closer to normal, so the next day was probably going to be more normal anyway, penny or not.
The regression fallacy occurs when people observe an extreme outcome followed by a more moderate one and mistakenly attribute the return toward average to whatever action they took in between, rather than recognizing the statistical inevitability of regression to the mean. Because people are most motivated to intervene when things are at their worst (or best), any subsequent normalization feels like proof that the intervention worked. This creates a systematic illusion of causal efficacy for treatments, punishments, rituals, and policies that may have had no real effect. The fallacy is especially insidious because it can sustain belief in ineffective or even harmful practices for generations, as the natural ebb and flow of variable phenomena continually reinforces the illusion.
The same glitch looks different depending on the terrain. Finance, medicine, a relationship, a team — same mechanism, different costume.
Investors observe that stocks which performed extremely poorly in one period often rebound in the next and attribute this to their own 'buy the dip' strategy rather than recognizing statistical mean reversion. Similarly, firing a fund manager after a bad quarter and seeing improvement afterward reinforces the false belief that the personnel change drove the recovery.
Patients typically seek treatment when symptoms are at their peak. Because symptoms naturally fluctuate and often improve from extreme levels, patients and clinicians attribute improvement to whatever treatment was administered — sustaining belief in ineffective remedies ranging from historical bloodletting to modern unproven supplements. Clinical trials without control groups are especially vulnerable to this distortion.
Teachers who punish students after unusually poor performance observe improvement, while those who praise students after unusually strong performance see decline. This creates a systematic illusion that punishment is more effective than reward, biasing pedagogical approaches toward punitive methods and away from positive reinforcement.
When conflicts in a relationship reach a peak and a partner makes a dramatic gesture — an apology, a gift, or an ultimatum — the natural de-escalation of tension is attributed to the gesture. This can reinforce unhealthy confrontation patterns or grand-gesture dynamics as the perceived 'solution' to relationship problems.
Product teams often deploy emergency fixes or A/B test changes after metrics hit unusual lows. If metrics rebound (as they statistically tend to), the change is credited as effective. This leads to accumulating features and tweaks that may have had no real impact, cluttering the codebase and misleading product strategy.
Managers who intervene harshly after a team's worst quarter and then see performance recover develop an authoritarian management style. Performance improvement plans initiated at the nadir of an employee's output appear to work simply because extreme underperformance naturally reverts, reinforcing the belief that formal intervention was necessary.
Policymakers introduce new legislation or programs in response to crisis peaks — crime waves, economic downturns, or public health spikes. When conditions naturally improve, politicians claim credit for the recovery, and voters reward them. This makes it extremely difficult to evaluate whether policies actually work without rigorous counterfactual analysis.
Francis Galton first described the statistical phenomenon of regression to the mean in 1886 in 'Regression Towards Mediocrity in Hereditary Stature.' The recognition of the regression fallacy as a cognitive bias was formalized by Daniel Kahneman and Amos Tversky in their 1973 paper 'On the Psychology of Prediction' and their 1974 Science paper 'Judgment under Uncertainty: Heuristics and Biases.'
In ancestral environments, rapidly detecting cause-and-effect relationships between actions and outcomes was critical for survival — if you ate a berry and got sick, assuming causation was safer than waiting for statistical proof. This hair-trigger causal detection system was adaptive because most action-outcome sequences in simple environments genuinely were causal. The cost of occasionally attributing natural fluctuation to an intervention was trivially small compared to the cost of missing a real causal threat.
Machine learning models trained on data selected at extreme values — such as training a recommendation system on users who just churned or optimizing ad spend after a campaign's worst week — can encode regression artifacts as learned patterns. If a model is evaluated only on outcomes following extreme inputs without proper control baselines, its apparent predictive accuracy is inflated by regression to the mean, leading to overconfidence in model performance.
50 cards are free to preview. Buyers unlock the rest of the deck plus the interactive training — Spot-the-Bias Quiz unlimited, Swipe Deck with spaced repetition, My Blindspots, Decision Pre-Flight, the Printable Deck + Cheat Sheets, and the Field Guide e-book. $29.50$59.
