A student scored in the 90th percentile on the first exam. What will she score on the second? If you predict "90th percentile again," you're ignoring regression to the mean. If you predict "exactly average," you're overcorrecting. Kahneman's four-step method threads the needle.
The Framework
The four-step regression correction is Kahneman's method for making unbiased predictions by explicitly accounting for regression to the mean. The steps: (1) Establish the baseline prediction — the average outcome for the relevant reference class. (2) Form an intuitive impression of the specific case — your gut assessment of how far above or below average this case seems. (3) Estimate the correlation between your evidence and the outcome — how predictive is the information you have? (4) Move from the baseline toward your intuitive impression, but only by a fraction equal to the correlation.
If the correlation between exam 1 and exam 2 scores is 0.5, and the student scored 20 percentile points above average on exam 1, the corrected prediction is 10 points above average (0.5 × 20). Your intuitive impression (90th percentile again) is adjusted halfway back toward the mean. The lower the correlation, the more you regress toward the baseline. With zero correlation, predict the mean. With perfect correlation, predict the intuitive impression. Everything else falls between.
Where It Comes From
Chapter 18 of Thinking, Fast and Slow presents the four-step method as the practical antidote to the intuitive predictions that dominate human judgment. Without the correction, people consistently make extreme predictions — predicting that an excellent student will be excellent again, that a bad quarter will be followed by another bad quarter — ignoring the statistical reality that extreme observations contain a large luck component that won't persist.
> "Unbiased predictions leave the determination of an individual case to the statistics of the reference class." — Thinking, Fast and Slow, Ch 18
The Implementation Playbook
Sales Forecasting: A salesperson who crushed quota last quarter (150% attainment) should not be forecasted at 150% next quarter. Step 1: baseline = 100% (average quota attainment). Step 2: intuitive impression = 150% (last quarter was exceptional). Step 3: correlation between consecutive quarters = maybe 0.4. Step 4: corrected prediction = 100% + 0.4 × 50% = 120%.
Hiring: A candidate who aced the interview (top 10% impression) should not be predicted to perform at the top 10%. The correlation between interview performance and job performance is roughly 0.3-0.5 for structured interviews. Correct the prediction: predict a strong performer, but closer to average than the interview impression suggests.
Investment: A fund that returned 25% last year (vs. 10% market average) should be predicted to return roughly 10% + (correlation × 15%). Since the correlation of fund returns year-to-year is near zero for most active managers, the corrected prediction is close to the market average.
Academic Performance: A student who scored in the 95th percentile on the SAT will likely perform well in college — but not in the 95th percentile. The SAT-GPA correlation is roughly 0.5, so predict the 75th percentile (halfway between the 50th percentile baseline and the 95th percentile SAT).
Key Takeaway
The four-step regression correction replaces the natural (and always too extreme) intuitive prediction with a calibrated estimate. It acknowledges both the evidence (your intuitive impression) and the reality (regression to the mean), weighting them by the correlation between evidence and outcome. It will always feel too conservative — and it will almost always be more accurate than the uncorrected gut prediction.
Continue Exploring
[[Regression to the Mean]] — The statistical phenomenon the four-step method accounts for
[[Planning Fallacy]] — The prediction error that results from skipping regression correction
[[Reference Class Forecasting]] — The complementary outside-view method for project-level predictions
📚 From Thinking, Fast and Slow by Daniel Kahneman — Get the book