Decision Mathematics under Risk and Uncertainty

Introduction

Business decisions are often made before managers know which future state will occur. A retailer must choose inventory before demand is observed; a manufacturer must select capacity before market growth is clear; and a firm may invest in a project whose returns depend on regulation, technology, and competition. Decision mathematics provides a structured way to compare alternatives when outcomes are uncertain.

This chapter distinguishes risk, where probabilities can reasonably be assigned to future states, from uncertainty, where probabilities are unavailable, disputed, or highly subjective. It develops payoff tables, decision trees, expected monetary value, opportunity loss, utility, value of information, scenario analysis, sensitivity analysis, and robust decision-making. The objective is not to replace managerial judgment, but to make assumptions visible, calculations reproducible, and trade-offs explicit.

Learning Objectives

By the end of this chapter, you should be able to:

1. Construct and interpret payoff tables and decision trees.
2. Distinguish decisions under certainty, risk, and uncertainty.
3. Calculate expected monetary value and expected opportunity loss.
4. Use utility functions when monetary outcomes do not adequately represent preferences.
5. Evaluate the expected value of perfect and sample information.
6. Conduct scenario and sensitivity analyses.
7. Compare alternatives using robust reasoning when probabilities are unreliable.
8. Recognize ethical, governance, and practical limitations of quantitative decision models.

1. The Structure of a Decision Problem

A decision problem contains four basic elements:

• Decision alternatives: actions controlled by the decision-maker.
• States of nature: relevant future conditions not controlled by the decision-maker.
• Payoffs: financial or non-financial consequences of each action–state combination.
• Information: current evidence and any information that may be obtained before acting.

A payoff table has alternatives as rows and states of nature as columns. Let a_i denote alternative i, s_j denote state j, and v_ij denote the payoff when alternative a_i is selected and state s_j occurs.

The entries may represent profit in thousands of currency units. A negative payoff represents a loss. The table is complete only if the alternatives are mutually exclusive and the states cover the important possibilities.

Certainty, Risk, and Uncertainty

• Under certainty, the relevant state is known.
• Under risk, probabilities p_j can be assigned to states, with 0 ≤ p_j ≤ 1 and ∑p_j = 1.
• Under uncertainty, probabilities are unavailable or not sufficiently credible.

This classification concerns the decision-maker’s knowledge, not merely the objective randomness of the environment. Two firms may face the same market but reasonably use different probability assessments because their information differs.

2. Decision Trees

A decision tree represents the temporal order of decisions and uncertain events. A square is conventionally used for a decision node, a circle for a chance node, and terminal branches for final outcomes. The tree is evaluated from right to left: first calculate chance-node expected values, then select the best branch at each decision node.

Figure: A decision tree separates controllable decisions from uncertain states and terminal payoffs.

A tree is especially useful when information arrives in stages, when later decisions depend on earlier outcomes, or when a decision-maker can abandon, expand, or revise a project.

3. Expected Monetary Value

When monetary payoffs are appropriate and probabilities are credible, the expected monetary value (EMV) of alternative a_i is the probability-weighted average of its possible payoffs:

Here, m is the number of states, p_j is the probability of state j, and v_ij is the payoff from alternative i in that state. The formula assumes a common monetary unit, mutually exclusive states, collectively exhaustive states, and probabilities that represent the decision-maker’s beliefs. EMV is a long-run average criterion; it does not predict the payoff of one particular decision.

Worked Example: Capacity Choice

Using the table above and probabilities of high demand 0.6 and low demand 0.4:

The large-capacity alternative has the highest EMV: 84 thousand currency units. This recommendation is not automatically correct if the firm has limited cash, severe downside aversion, or unreliable probabilities.

4. Opportunity Loss and Regret

Opportunity loss, also called regret, measures the payoff forgone because the chosen alternative was not the best alternative for the state that actually occurred. For state s_j, the opportunity loss of alternative a_i is:

The best payoff in each column is identified first. For the capacity example, the high-demand maximum is 180 and the low-demand maximum is 40.

The expected opportunity loss (EOL) is:

Thus, EOL(Large) = 0.6(0) + 0.4(100) = 40, EOL(Small) = 0.6(80) + 0.4(0) = 48, and EOL(Outsource) = 0.6(110) + 0.4(10) = 70. The lowest EOL is 40, again selecting large capacity.

For a complete and correctly constructed payoff table:

5. Value of Information

Information has economic value only if it changes decisions enough to improve expected outcomes after its cost is considered.

Expected Value of Perfect Information

With perfect information, the decision-maker knows the state before choosing an alternative. The expected value with perfect information is:

The expected value of perfect information (EVPI) is:

For the example:

Therefore, EVPI = 124 − 84 = 40 thousand currency units. The firm should not pay more than 40 thousand for perfect information, before considering implementation costs and risk preferences.

Sample Information and Bayes’ Rule

Market research is usually imperfect. Suppose a research report can be positive or negative. Let [[MATH{T}]] denote the true state and [[MATH{R}]] the report. Posterior probabilities are calculated using Bayes’ rule:

The expected value with sample information is found by calculating the best conditional EMV after each report outcome, weighted by the probability of that outcome:

The expected value of sample information is EVSI = EV_SI − max_i EMV(a_i). It cannot exceed EVPI. A report should be purchased only when its EVSI exceeds its price and relevant implementation costs.

6. Utility and Risk Attitudes

EMV treats one monetary unit as equally valuable in all circumstances. In reality, a loss may matter more to a financially constrained start-up than to a large corporation, and decision-makers may prefer a secure outcome to a risky outcome with the same mean.

A utility function u(x) converts a monetary payoff x into a preference value. The expected utility of alternative a_i is:

The alternative with the highest expected utility is preferred under the specified utility model. A risk-averse decision-maker has a concave utility function; a risk-neutral decision-maker has linear utility; and a risk-seeking decision-maker has a convex utility function.

One simple risk-averse function, applicable only when its domain is positive, is u(x)=√x. If outcomes may be negative, a shifted function such as u(x)=√(x+c) requires x+c≥0 and a justified constant c. Utility functions should be elicited carefully rather than selected merely because they are mathematically convenient.

The certainty equivalent is the sure amount producing the same utility as a risky prospect. If utility is invertible:

The risk premium is the expected monetary value minus the certainty equivalent. It represents the maximum amount the decision-maker would theoretically surrender to eliminate the risk, subject to the validity of the utility model.

7. Decisions under Uncertainty

When probabilities are not credible, several criteria can be used. They express different attitudes toward ambiguity and should not be presented as universally optimal.

For Hurwicz’s criterion, with 0≤α≤1:

These criteria can yield different choices because they encode different views about ambiguity, resilience, and managerial priorities. A responsible analysis reports the criterion and tests how conclusions change.

8. Scenario, Sensitivity, and Robust Analysis

Scenario Analysis

Scenario analysis develops coherent combinations of assumptions, such as recession, base case, and expansion. It differs from changing one variable at a time: a scenario changes several related drivers consistently.

A scenario-weighted value can be calculated when scenario probabilities are defensible:

Here, V_r is the value under scenario r, w_r is its weight, and ∑w_r=1. If the weights are merely subjective labels, they should not be presented as objective probabilities.

Sensitivity Analysis

Sensitivity analysis changes one or more assumptions to examine their influence on the recommendation. For a two-state alternative:

The probability at which alternatives a and b are equally attractive solves:

For the large and small capacity choices:

This gives 240p−60 = 60p+40, hence p=0.5. Large capacity is preferred when the probability of high demand exceeds 0.5; small capacity is preferred when it is below 0.5. This break-even analysis identifies the assumption deserving the greatest managerial attention.

Useful sensitivity techniques include tornado diagrams, one-way and two-way data tables, threshold analysis, and simulation. Sensitivity is not a forecast: it reveals dependence on assumptions.

Robust Reasoning

A robust decision performs acceptably across a broad range of plausible states and model specifications. Robustness may involve:

• limiting severe downside outcomes;
• preserving flexibility through staged investment or real options;
• diversifying suppliers, technologies, or markets;
• using conservative assumptions where errors are costly;
• monitoring leading indicators and defining trigger points;
• selecting an alternative that remains reasonable under several probability sets.

Robustness is especially valuable under deep uncertainty, where precise probabilities may create false confidence. It may justify a lower-EMV option if the EMV advantage depends on fragile assumptions.

9. Illustrative Case: Regional Distribution Centre

Background and Problem

The fictional Northstar Foods is considering a regional distribution centre. It can build a large centre, build a small centre, or contract with a logistics provider. Demand over the next planning period may be high or low. Payoffs, in thousands, are those in the opening table. Management currently estimates a 0.6 probability of high demand.

Analysis

The EMVs are 84, 76, and 54 respectively, so large capacity is the EMV choice. Its worst payoff is −60, however, while the small centre guarantees at least 40. The EVPI is 40, setting an upper limit on the value of perfect demand information.

Suppose a consulting firm offers a demand study for 18 thousand. Historical validation suggests:

• positive report probability given high demand: 0.80;
• positive report probability given low demand: 0.25.

The probability of a positive report is 0.8(0.6)+0.25(0.4)=0.58. The posterior probability of high demand after a positive report is 0.48/0.58≈0.828. After a negative report it is 0.12/0.42≈0.286.

After a positive report, large capacity has conditional EMV approximately 0.828(180)+0.172(−60)=89.3, while small capacity has approximately 0.828(100)+0.172(40)=89.7. Thus, small capacity is marginally preferred after a positive report. After a negative report, small capacity is clearly preferred: 0.286(100)+0.714(40)=57.1, compared with large capacity’s approximately −3.4. Therefore:

The EVSI is approximately 76.0−84=−8.0, which indicates an inconsistency: information cannot have negative value when the decision-maker can ignore it. The issue is that the positive-report calculation selected small capacity, while the unconditional best action is large; the conditional expected value must be compared with the option of retaining the original action, and arithmetic should be checked carefully. Recomputing the positive branch: large gives 137.9, not 89.3, because 0.828(180)+0.172(−60)=138.0. Small gives 89.7. Thus large is selected after a positive report. The negative branch selects small. Consequently:

The EVSI is approximately 20 thousand. Since the study costs 18 thousand, its net expected value is about 2 thousand, before considering delays and operational costs. The result is economically attractive but fragile: errors in test accuracy, payoff estimates, or implementation could eliminate the gain.

Outcome and Learning Points

Northstar should purchase the study only if the research assumptions are credible and the organization can act on the report. It should also define a downside limit, review the recommendation under lower high-demand probabilities, and retain an operational fallback. The case demonstrates the importance of decision-tree timing, Bayes’ rule, arithmetic verification, and the distinction between gross and net information value.

10. Governance, Ethics, and Technology

Decision models can produce precise-looking recommendations from weak assumptions. Governance should therefore require documented data sources, probability elicitation procedures, model ownership, version control, validation, and independent review. Analysts should disclose omitted alternatives, correlation between assumptions, confidence limits, and the consequences of model error.

Ethically, decision criteria must not conceal distributional effects. A project with high expected profit may impose costs on workers, communities, or the environment. Payoffs should include material externalities where possible, and non-monetary impacts should be reported rather than silently converted into arbitrary monetary values.

Spreadsheets, optimization software, Monte Carlo simulation, and AI-assisted forecasting can improve speed and scenario coverage. They do not establish causality or make uncertain probabilities objective. AI systems may reproduce historical bias, obscure assumptions, or encourage automation bias. Human accountability, audit trails, sensitivity testing, privacy protection, and governance approval remain essential.

Best-Practice Workflow

1. Define the decision, timing, objective, and constraints.
2. Identify alternatives and mutually exclusive, collectively exhaustive states.
3. Build and audit the payoff table.
4. Elicit probabilities only where evidence supports them.
5. Calculate EMV, EOL, and relevant utility measures.
6. Assess perfect and sample information value.
7. Test scenarios, thresholds, and parameter sensitivity.
8. Examine downside exposure and robust alternatives.
9. Include information, implementation, ethical, and external costs.
10. Document assumptions, monitor outcomes, and update the model.

Chapter Summary

Decision mathematics organizes choices involving uncertain future states. Payoff tables provide a compact representation of alternatives, states, and consequences, while decision trees represent timing and sequential information. Under credible probabilities and suitable monetary outcomes, EMV compares probability-weighted payoffs; opportunity loss and EOL measure regret. EVPI and EVSI establish rational upper limits on information expenditure. Utility extends analysis to risk attitudes and produces expected-utility and certainty-equivalent comparisons. When probabilities are unavailable, maximin, maximax, minimax regret, Laplace, and Hurwicz criteria offer contrasting perspectives. Scenario analysis, sensitivity analysis, threshold calculations, and robust reasoning reveal whether a recommendation is stable. Good practice combines quantitative discipline with governance, ethical evaluation, managerial judgment, and continuing review.

Key Terms

Decision alternative: An action available to the decision-maker.

State of nature: A future condition not controlled by the decision-maker.

Payoff: The consequence associated with an alternative and state.

Payoff table: A matrix of payoffs by alternatives and states.

Decision tree: A graphical representation of sequential decisions and uncertain events.

Risk: A situation in which credible probabilities can be assigned.

Uncertainty: A situation in which probabilities are unavailable or unreliable.

Expected monetary value: Probability-weighted average monetary payoff.

Opportunity loss: Payoff forgone relative to the best action for the realized state.

Expected opportunity loss: Probability-weighted average opportunity loss.

Perfect information: Information that reveals the future state with certainty.

Sample information: Imperfect information obtained from a forecast, test, or study.

Utility: A numerical representation of preference for outcomes.

Certainty equivalent: A sure amount equally preferred to a risky prospect.

Risk premium: EMV minus the certainty equivalent.

Scenario analysis: Evaluation of coherent alternative future environments.

Sensitivity analysis: Examination of how results change when assumptions vary.

Robust decision: A decision performing acceptably across many plausible conditions.

Model governance: Controls for assumptions, validation, accountability, and use.

Review Questions

5 Short Answer Questions

1. Distinguish risk from uncertainty.
2. What are the four basic elements of a decision problem?
3. Define expected monetary value.
4. Why can EVPI be regarded as an upper limit on information expenditure?
5. What is the purpose of sensitivity analysis?

5 Long Answer Questions

1. Explain how a decision tree is constructed and evaluated, including the roles of decision and chance nodes.
2. Compare EMV, expected opportunity loss, and expected utility as decision criteria.
3. Explain Bayes’ rule and describe how sample information can be valued.
4. Compare maximin, maximax, minimax regret, Laplace, and Hurwicz criteria.
5. Discuss why governance and ethical review are necessary when using quantitative decision models.

3 Analytical / Application Questions

1. An alternative returns 150 if demand is high and −50 if demand is low. A second returns 90 if demand is high and 30 if demand is low. Determine the probability of high demand at which their EMVs are equal, and interpret the result.
2. Construct a payoff table for a product launch with three launch sizes and three demand states. Calculate EMV, opportunity loss, and minimax regret using probabilities that you justify.
3. A manager can buy a market report before selecting capacity. Design a decision tree, specify the conditional probabilities required, calculate EVSI, and recommend whether the report should be purchased after including its cost.

Multiple Choice Questions with Answers

Question 1

Which condition characterizes decision-making under risk?

A. The future state is known with certainty.

B. Credible probabilities can be assigned to future states.

C. No alternatives are available.

D. All payoffs are identical.

Correct Answer: B. Credible probabilities can be assigned to future states.

Question 2

Using probabilities 0.6 and 0.4, what is the EMV of payoffs 180 and −60?

A. 48

B. 72

C. 84

D. 120

Correct Answer: C. 84

Question 3

Opportunity loss for an alternative in a state is calculated as:

A. Its payoff minus the state maximum

B. The state maximum payoff minus its payoff

C. Its payoff multiplied by its probability

D. The average of all row payoffs

Correct Answer: B. The state maximum payoff minus its payoff

Question 4

For the capacity example, the EVPI is:

A. 18

B. 40

C. 76

D. 124

Correct Answer: B. 40

Question 5

A concave utility function generally represents:

A. Risk aversion

B. Risk neutrality

C. Risk seeking

D. Perfect information

Correct Answer: A. Risk aversion

Question 6

Which criterion chooses the alternative with the largest worst-case payoff?

A. Maximax

B. Laplace

C. Maximin

D. Minimax regret

Correct Answer: C. Maximin

Question 7

If high-demand probability is p, the EMV of payoffs 150 and −50 is:

A. 150p − 50(1−p)

B. 150 + 50p

C. 50p − 150(1−p)

D. 200p

Correct Answer: A. 150p − 50(1−p)

Question 8

The expected value of sample information should be compared with:

A. The largest possible payoff only

B. The cost of obtaining and implementing the information

C. The worst payoff only

D. The number of alternatives

Correct Answer: B. The cost of obtaining and implementing the information

Question 9

What is a central purpose of sensitivity analysis?

A. To guarantee the forecast is correct

B. To eliminate all uncertainty

C. To identify how assumptions affect a recommendation

D. To replace managerial judgment entirely

Correct Answer: C. To identify how assumptions affect a recommendation

Question 10

Which statement best describes a robust decision?

A. It is optimal only under one precise forecast.

B. It performs acceptably across many plausible conditions.

C. It always has the highest EMV.

D. It ignores downside outcomes.

Correct Answer: B. It performs acceptably across many plausible conditions.