Applications across domains

The point of a single unifying decision engine is that the same seven-parameter machine ($\rho,\lambda/\kappa,\varepsilon,\omega,Z$, prior, evidence) produces a defensible recommendation in wildly different domains, simply by changing the states, actions, payoffs, and — where the professional context calls for it — the parameter settings. Each preset in the Decision Studio is a runnable version of the sketches below.

Insurance — reinsurance purchase. States: no/one/two large losses this treaty year. Actions: buy an excess-of-loss layer for a fixed premium, or retain. Because a single large loss can be existential for capital, this is a natural case for moderate risk aversion (finite $\rho$ set near the firm's risk tolerance in dollars) and non-trivial credibility blending ($Z<1$) between the firm's own thin loss history and industry benchmark severity curves. The EVPI calculation quantifies, in dollars, what a perfect catastrophe forecast would be worth — a natural budget ceiling for catastrophe-modelling spend.

Actuarial reserving — point estimate selection. States: favourable / as-expected / adverse development. Actions: book a low, central, or prudent (high) reserve. Regulatory and rating-agency scrutiny of understatement is typically asymmetric with the capital cost of overstatement, which the payoff matrix should encode directly (larger penalty for the adverse-development branch under a low booking) rather than via an artificial risk-aversion parameter — a good discipline generally: model asymmetric consequences in the payoffs first, and reserve $\rho$ for genuine risk-aversion-over-money effects.

Finance — asset allocation. States: macro regimes (recession/muted growth/boom). Actions: equities/bonds/cash. This is the textbook case for comparing EMV (which favours equities) against expected-utility and ambiguity-robust variants (which pull toward bonds/cash as $\rho$ falls or $\varepsilon$ rises) — a clean illustration of how the "same" data support different rational recommendations depending on disclosed risk and ambiguity attitudes.

Health — treatment choice. States: responder / partial responder / non-responder. Payoffs in QALYs. Health decisions often justify a non-zero $\varepsilon$: the evidence base for how a specific patient will respond is frequently thinner and more heterogeneous than a population-level clinical trial number, so an ambiguity-robust (Gilboa–Schmeidler-flavoured) evaluation alongside plain expected QALYs gives clinicians a defensible "even in the worst plausible case, here is what we'd expect" view, consistent with modern shared-decision-making practice.

Education — exam-sitting choice. States of preparedness are highly self-assessed and prone to overconfidence (see "Behavioural pitfalls"); ADE here is most useful for making the prior over one's own preparedness explicit and then checking how sensitive the sit-vs-delay recommendation is to that (likely biased) self-assessment.

Economics — capacity expansion. States: demand growth scenarios. Actions: expand now / expand later (an embedded real option) / hold. Comparing "expand now" and "expand later" values directly under ADE quantifies the real-option value of waiting for information, connecting this section back to "Decision trees and sequential decisions."

Everyday — the umbrella problem. States: rain / no rain. Actions: take / leave the umbrella. Trivial in scale, but pedagogically exact: it is the cleanest possible illustration that even "take an umbrella" is a Bayesian decision under a payoff asymmetry (soaking is worse than mild inconvenience), and the same EMV/EU/maximin/regret table used for a nine-figure reinsurance decision applies unchanged.

The common thread. In every domain, the analyst must (1) define states and actions honestly, (2) elicit or estimate payoffs in comparable units, (3) form a defensible prior (data, benchmark, or judgement — disclosed), and (4) disclose the risk-aversion, distortion, ambiguity, and regret parameters used, rather than silently picking the classical method that happens to favour a preferred answer. ADE's contribution is not new mathematics in any one domain — it is making the same auditable dial-settings do the work everywhere, and showing, side by side, exactly which classical rule each setting reproduces.