Decisions & optimisation

Robust decisions: expected value or downside protection?

How much should you order when demand is uncertain? The same numbers produce different quantities under different decision rules. This calculator makes explicit which uncertainty and which kind of loss each rule addresses.

  • 4 min read
  • Updated on
  • smartlytics editorial team

The short answer

Robust decisions examine how an action performs across possible states. Maximin protects the worst payoff; minimax regret limits the largest missed opportunity. Expected value instead weights states by their probabilities.

Try the free interactive calculatorFree · in your browser · no sign-up

At a glance

  1. A best decision requires an explicit evaluation criterion.
  2. Maximin and minimax regret need scenarios here, but do not use scenario probabilities.
  3. Robustness applies only to the states that the model actually includes.

How much should we order before demand is known?

A retailer orders either 20, 40 or 60 units before a sales day. Each unit costs €6 and sells for €10. Unsold stock initially has a salvage value of €2 per unit. Demand can be 20, 40 or 60 units. The decision must be made before sales begin, and the order cannot be changed afterwards.

For an order qq and demand dd, the payoff is 10min⁡(q,d)+smax⁡(q−d,0)−6q10\min(q,d)+s\max(q-d,0)-6q, where ss denotes salvage value. A stockout means missed sales, but no additional contractual penalty. This simple formula gives nine inspectable outcomes. It is not a demand forecast: it describes what a particular decision would mean under each assumed demand state.

Three rules ask three different questions

With a €2 salvage value, ordering 20 units produces €80 in every state. Ordering 40 produces zero when demand is low and €160 otherwise. Ordering 60 gives payoffs of minus €80, €80 and €240. The larger possible gain therefore comes with a weaker downside outcome.

The starting probabilities are 30% for low demand, 30% for medium demand and 40% for high demand. Expected value is then €80 for an order of 20, €112 for 40 and €96 for 60. Maximising this weighted average selects 40 units. That does not promise an actual payoff of €112: none of the three states produces that amount.

Maximin compares only the worst payoffs: €80, zero and minus €80. It selects 20 units. Minimax regret instead measures the missed opportunity relative to a decision made with perfect advance knowledge. The best payoffs in the three demand states are €80, €160 and €240. The maximum regrets are therefore €160, €80 and €160 for the three order quantities. This rule selects 40 units.

Experiment with probabilities and salvage value

Interactive

Robust decisions – interactive calculator

With demand probabilities of 30%, 30% and 40%, expected value and minimax regret select 40 units; maximin selects 20. Unsold stock initially has a salvage value of €2 per unit.

First increase only the probability of high demand. The remaining probability is split equally between low and medium demand. Expected payoffs change, but individual entries in the payoff matrix do not. The expected-value choice and ranking can therefore change while the maximin and minimax-regret choices stay the same.

Next change salvage value. This changes the consequences of ordering too much, so it also affects the robust criteria. Selecting a decision rule highlights its choice without hiding the other results. When alternatives tie, the calculator prefers the smaller order. This is an explicit simplifying convention rather than an additional optimisation objective.

What robustness means here

Robust optimisation handles uncertain model parameters by explicitly considering a set of possible values (Bertsimas, D.; Brown, D. B.; Caramanis, C. 2010). Our example uses three finite demand states. Maximin and minimax regret remain distinct criteria: one protects the absolute worst payoff, while the other limits the largest gap from the decision that would have been best in hindsight.

Robustness can have a price in the form of lower expected performance or a smaller upside. Research explicitly studies this trade-off between protection and performance (Bertsimas, D.; Sim, M. 2004). In our example, maximin gives up some expected payoff to improve the worst outcome. That makes it neither universally better nor universally worse. What matters is which consequences the business can actually absorb and which objective its decision makers endorse.

Which states might be missing?

The scenario set determines the claim. Zero demand, a late supplier, perishable stock with no salvage value and additional fixed costs are all absent. A decision can be robust within these three states and fragile outside them. The probabilities are also adjustable learning assumptions, not a forecast estimated from customer data.

For a real decision, the people responsible should first verify costs, revenues and practical limits on the available actions. Then agree on a criterion and examine sensitivity: at which probability or salvage value does the preferred action change? If that threshold lies in a plausible range, better information may be worthwhile. The decision theory article places these criteria within a broader decision process.

Before using a robust choice as policy, also record who can revise it and when. A rule chosen for a one-day order may be unsuitable for a repeated process that can learn from sales. Keeping the decision horizon explicit prevents a conservative local choice from silently becoming a permanent business rule.

Frequently asked questions

Is maximin always the safest decision?

Maximin maximises the worst modelled payoff. Whether that is the relevant form of safety depends on the objective and whether the scenarios cover important risks.

What does regret mean?

Regret is the gap from the best action if the realised state had been known in advance. It measures a missed opportunity, not necessarily a financial loss.

Sources and further reading

  1. OnlineBertsimas, D.; Brown, D. B.; Caramanis, C. (2010): Theory and Applications of Robust Optimization.
  2. OnlineBertsimas, D.; Sim, M. (2004): The Price of Robustness.

Content last reviewed on 14 November 2026. Please let us know about any errors: impressum@nyxai.com.

From NyxAI

smartlytics is the learning hub of NyxAI GmbH.

Explore the methods we work with, explained openly, with sources and calculations you can reproduce.

Decision Core

NyxAI is developing a Decision Core for Decision Intelligence that combines optimisation, simulation, statistics and reinforcement learning: compare options, make uncertainty visible and learn from outcomes, with people in control.

DeepSquare

Explainable chess AI with human playing styles: NyxAI’s testing ground for understandable decisions under time pressure.