Behavioural pitfalls

Even decision-makers who accept expected utility and Bayesian updating as normatively correct are subject to well-documented systematic biases in practice. Recognising these is as important as knowing the formal theory, because a technically correct model fed miscalibrated inputs still produces a bad decision.

Framing effects. Tversky and Kahneman (1981) showed that logically identical decisions produce different choices depending on whether outcomes are framed as gains or losses — e.g., "200 of 600 will be saved" vs. "400 of 600 will die" describe the same programme but elicit systematically different preferences (risk-averse framing toward the "saved" wording, risk-seeking toward the "die" wording). This is prospect theory's reference-dependence: people evaluate outcomes relative to a reference point, not in absolute terms, and are typically risk-averse for gains but risk-seeking for losses of the same magnitude — a reversal plain expected utility (with a single globally concave $u$) cannot represent. Practically: always state a decision problem in both frames and check the recommendation doesn't change; if it does, the framing — not the substance — is doing the work.

Overconfidence. People (including experts) systematically overestimate the precision of their own knowledge: asked for 90% confidence intervals on uncertain quantities, the true value falls outside the stated interval far more than 10% of the time in typical studies. In insurance this shows up as underestimated tail risk in loss-development or catastrophe assumptions, and in decision terms it means the prior fed to ADE is too narrow — states that "shouldn't" happen do, more often than modelled. The remedy is not a different decision rule but better-elicited, appropriately wide priors (structured elicitation protocols, reference-class forecasting, explicit adversarial "what would make me wrong" review) and stress-testing the recommendation against a deliberately wider or fatter-tailed prior.

Calibration. A forecaster is well-calibrated if, among all the times they say "70% chance," the event actually happens about 70% of the time. Calibration is checkable with historical track records (Brier scores, calibration plots) and is a necessary (not sufficient) condition for a probability estimate to be taken seriously in a decision model — a beautifully coherent but badly calibrated prior still produces bad decisions. Organisations that track forecaster calibration over time (as in some underwriting and reserving functions) can materially improve the inputs to any of the frameworks in this section.

Anchoring, availability, and the planning fallacy. Estimates are pulled toward an initially presented number (anchoring) even when it is arbitrary; the perceived probability of an event rises with how easily examples come to mind (availability), which biases catastrophe and rare-event assessment toward recently experienced events and away from equally likely but less memorable ones; and plans/estimates for one's own tasks or projects are systematically optimistic relative to comparable historical outcomes (the planning fallacy), relevant to reserving and project-cost estimation alike.

Where this leaves formal decision theory. None of these biases invalidate expected utility, Bayesian updating, or ADE — they are arguments for rigour in eliciting the inputs (priors, payoffs, likelihoods) that feed a normatively sound method, and for routinely stress-testing a recommendation's sensitivity to those inputs, exactly what ADE's sensitivity/break-even panel is for. A sound method fed a biased prior still gives a biased answer; the discipline of decision theory and the discipline of debiasing judgement are complements, not substitutes.