Chapter 18: Taming Intuitive Predictions

Type: Chapter 18

Category: Psychology

Difficulty: Intermediate

Reading time: 6 min

Core idea

When you predict someone’s future performance from a current impression — a job interview, a student’s essay, a startup’s pitch — you tend to predict as if the impression is a perfect indicator. Impressive interview → predicted to be an exceptional employee. Dull pitch → predicted to fail. The prediction matches the strength of the impression.

This is wrong. No single impression is a perfect predictor of future performance. All predictors are imperfect. And when predictors are imperfect, the correct prediction is less extreme than the impression — because of regression to the mean.

The intuitive prediction error is systematic and directional: people predict outcomes that are too extreme, too confident, and not sufficiently regressed toward the base rate. Kahneman provides a procedure to correct this: match percentiles in the impression to percentiles in the reference distribution, then adjust toward the mean by the actual correlation between the predictor and the outcome.

Why it matters

The structure of the prediction error

An intuitive prediction works like intensity matching: the vividness of the impression maps onto the vividness of the predicted outcome. If someone’s interview performance is in the 90th percentile of impressiveness, the intuitive prediction places them in the 90th percentile of future performance. But this is correct only if interviews perfectly predict performance — which they do not.

The actual predictive validity of job interviews is surprisingly low, typically around 0.3–0.5 correlation with job performance. A correlation of 0.4 means that about 16% of the variance in performance is explained by the interview. The remaining 84% is noise, other factors, and regression toward the mean. An intuitive prediction that anchors on the 90th-percentile impression and does not adjust for the imperfect correlation will dramatically overestimate the probability of 90th-percentile performance.

The two-step correction

Kahneman’s prescriptive procedure:

1. Start with the base rate: What is the average outcome for people in this reference class? This is the anchor — not the impression.
2. Adjust toward the impression by the correlation: if the correlation between the predictor and the outcome is 0.4, move 40% of the way from the base rate toward the impression-predicted value.

The result is a prediction that is always less extreme than the intuitive one. It accounts for both the diagnostic value of the impression and the regression that imperfect prediction implies.

Why intuition resists the correction

The corrected prediction feels wrong. It feels insufficiently responsive to evidence — like you are ignoring what you know about the person. But the feeling of “ignoring evidence” is an artifact of System 1’s intensity matching: the evidence feels more diagnostic than it is. The regression adjustment feels like giving up information; it is actually giving up false confidence.

Author’s argument: Overconfident predictions feel like good judgment. Statistically calibrated predictions feel like surrendering to averages. The feeling is wrong — the calibrated prediction is correct more often, even though it feels like a hedge.

Key takeaways

Key takeaways

• Intuitive predictions are too extreme: people predict outcomes at the same percentile as their impression of the evidence, ignoring that all predictors are imperfect and regression is inevitable.
• The intensity matching error: a vivid impression maps directly onto a vivid prediction. This is correct only if the predictor perfectly correlates with the outcome — which nothing does.
• Two-step correction: (1) anchor on the base rate for the reference class; (2) adjust toward the evidence-implied value by the actual correlation between predictor and outcome.
• The corrected prediction is always less extreme than the intuitive one — it accounts for regression toward the mean by the amount the predictor is imperfect.
• The calibration feels wrong: statistically correct predictions feel like ignoring evidence. This feeling is the hallmark of intensity matching overriding statistical reasoning.
• Application: hiring, clinical prediction, financial forecasting — all domains where impressions are mistaken for near-perfect predictors.

Mental model

click to zoom

How intuitive predictions fail to account for regression, and the two-step correction that tames them.

Read it as: Intuitive prediction anchors on the impression and matches percentile to percentile — as if the predictor were perfect. The two-step correction starts at the base rate and moves toward the impression only as far as the actual correlation warrants. A correlation of 0.4 means moving 40% of the way, not 100%. The corrected prediction is less dramatic and more accurate.

Practical application

Why this matters for hiring

The interview-performance correlation is typically 0.3–0.5. An interview that puts a candidate in the 90th percentile of impressiveness predicts — correctly — performance somewhere between the 65th and 75th percentile, not the 90th. Structured interviews and work sample tests have higher correlations (0.5–0.6) and are less affected by this problem — but all predictors require regression adjustment.

The procedure in practice:

1. Identify the reference class: What is the relevant population this person/project/investment comes from?
2. Establish the base rate: What is the average outcome in that reference class? This is your anchor.
3. Estimate the correlation: How predictive is your evidence? (Use research if available; be honest about uncertainty.)
4. Adjust proportionally: If correlation is 0.4, move 40% of the distance from base rate to the evidence-implied value.
5. Report the prediction with a range: overconfident point predictions hide the regression uncertainty.

Example

A design firm is hiring a senior designer. They interview three candidates and rate them: Candidate A at the 85th percentile of impressiveness, B at the 60th, C at the 45th.

Using a correlation of 0.5 between interviews and job performance, the corrected predictions are:

• Candidate A: 50% + 0.5×(85-50)% = 67.5th percentile predicted performance
• Candidate B: 50% + 0.5×(60-50)% = 55th percentile
• Candidate C: 50% + 0.5×(45-50)% = 47.5th percentile

The rank order is preserved, but the extremes are compressed. The firm avoids predicting that A will be exceptional (85th percentile) when the honest prediction is above average (67th percentile). This prevents the disappointment of “great in the interview, average on the job” — which would otherwise be attributed to the person, not to the interview’s predictive limitations.

Related lessons

Related

Chapter 17: Regression to the Mean The statistical foundation — why extreme scores always move toward average when measurements are imperfect. Chapter 23: The Outside View Reference class forecasting — the complementary technique for correcting overconfident predictions by anchoring on base rates. Heuristics The mental shortcuts that produce overconfident predictions — intensity matching and representativeness.