6 ideas
When experts and statistical formulas predict the same outcome from the same information, the formula matches or beats the expert in nearly every comparison, including cases where the expert has access to extra data. Experts apply their own weighting rules inconsistently from case to case, and a formula's perfect consistency is worth more than the expert's claimed ability to spot special configurations.
A linear model that simply standardizes the relevant predictors and adds them with equal or even random-sign-corrected weights predicts nearly as well as an optimally fitted regression, and better than human judges. The skill that matters is knowing which variables to measure and in which direction; the weights matter far less than people assume, which is why these 'improper' models are robust.
A regression model built to mimic a particular expert's past judgments predicts outcomes more accurately than that same expert does. The model keeps the expert's valid policy and strips out the random noise of fatigue, mood, and inconsistency, showing that unreliability, not a lack of insight, is what degrades human prediction.
A decision is rational when its beliefs obey the rules of probability and its choices are consistent with the decider's own goals and utilities, not when it happens to turn out well. This lens separates the quality of a decision from its outcome and treats contradictions among one's own judgments, such as rating a conjunction more likely than one of its parts, as the definitive mark of irrationality.
Clinical experience often fails to improve judgment because practitioners rarely get systematic feedback, see only a biased sample of cases, and interpret outcomes through their existing beliefs. Confidence grows with years of practice while accuracy stays flat, so credentials and felt certainty are poor evidence that someone can predict well.
Intuitive judgment routinely confuses P(symptom | condition) with P(condition | symptom) and ignores how common the condition is to begin with. Correct inference requires combining the base rate with the diagnostic evidence as Bayes' theorem prescribes; a highly typical sign of a rare condition can still leave that condition improbable.