Build a fair comparison with the same objectives, horizon, constraints, cost conventions and market states.
Build a fair comparison with the same objectives, horizon, constraints, cost conventions and market states. Then calculate expected value, floor, dispersion and reversibility to expose trade-offs. Minimax regret is one possible criterion only; its formulation belongs to robust optimisation. Reject any comparison whose assumptions or conventions differ, then select a scenario that meets the floor and preselected risk criterion.
Method.
Define genuinely executable, distinct allocations—cautious, central, offensive—and document shared constraints. Build 3–5 coherent market states from material uncertainty, dependencies and available evidence; assign probabilities only when defensible. For every cell calculate contribution using identical conventions, then derive expectation, minimum, dispersion, implementation cost and review delay. Choose criterion by exposure and reversibility, and define signals for moving to another scenario.
Worked example.
A, B and C yield 180/270/330 k€, 110/300/400 k€, 210/240/280 k€ in low, central and high states; probabilities are 30%, 50%, 20%. Expectations: A 255 k€, B 263 k€, C 239 k€. Floors: A 180 k€, B 110 k€, C 210 k€. B gains 8 k€ expectation over A but lowers floor 70 k€; C raises floor 30 k€ but gives up 16 k€ expectation. A is compromise if approved minimum is 175 k€; B is reversible learning option; C is preferable if required floor becomes 200 k€.
Checks and limits.
Allocation and states are coherent and executable; expectation, minimum, dispersion and reversibility appear together; choice criterion explicitly depends on reversibility; signals enable scenario change. Matrix depends on retained states. No single criterion replaces agreement on floor and reversibility. Subjective probabilities do not become objective by calculation.
Resources and sources.
Download the comparison matrix. Related: optimise under uncertainty, formalise constraints, compare evidence. Sources: Schoemaker (1995); Goodwin and Wright (2001); Keeney and Raiffa (1976).