Decisions Under Risk and Uncertainty

Introduction and context

Managerial decisions are rarely made with complete information. A firm may know the alternatives available to it but not future demand, competitor behaviour, regulation, exchange rates, technology costs, or the probability of disruptive events. Decisions under risk and uncertainty provide a structured way to act when outcomes are not certain.

Risk usually refers to situations in which relevant future states and their probabilities can be estimated. Uncertainty refers to situations in which states, probabilities, or both are poorly known. The distinction is useful but not absolute: managerial judgement, data, experimentation, and market intelligence can often convert uncertainty into measurable risk.

This chapter develops a practical decision framework using payoff tables, decision trees, expected value, utility, value of information, Bayesian updating, sensitivity analysis, scenario planning, and Monte Carlo simulation. The central managerial principle is that a mathematically optimal decision is not automatically a strategically wise decision. Robust choices must also consider downside exposure, organisational capabilities, ethics, timing, reversibility, and the quality of the evidence.

Learning objectives

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

1. Distinguish decisions under certainty, risk, and uncertainty.
2. Construct and interpret payoff tables and decision trees.
3. Calculate expected monetary value and expected opportunity loss.
4. Explain risk attitudes using expected utility.
5. Evaluate the value of perfect and sample information.
6. Update beliefs using Bayes’ theorem.
7. Apply sensitivity analysis, scenario planning, and Monte Carlo simulation.
8. Recommend robust managerial actions while recognising model limitations and ethical responsibilities.

1. The structure of a decision problem

A decision problem contains five elements:

• Decision alternatives: actions controlled by the manager, such as launch, delay, outsource, or expand.
• States of nature: future conditions not controlled by the manager, such as high, medium, or low demand.
• Payoffs: profits, costs, contribution, cash flow, utility, or another measure resulting from each alternative–state combination.
• Probabilities: beliefs assigned to states when risk can be quantified.
• Decision criterion: the rule used to select an alternative.

The timing of information matters. A manager may decide before demand is observed, after a market test, or after an initial pilot. A good analysis therefore asks not only “Which option is best?” but also “What will be known when the decision is made?”

Risk, uncertainty, and ambiguity

Under certainty, one relevant state is known. Under risk, several states are recognised and probabilities are available or defensible. Under uncertainty, probabilities are unavailable, disputed, or highly unreliable. Ambiguity is uncertainty about the appropriate probability model itself—for example, uncertainty over whether a new technology will create a new market category.

Probabilities may be:

• Objective, based on repeatable frequencies or well-established data; or
• Subjective, representing a manager’s informed belief.

Subjective probabilities are legitimate analytical inputs, but they should be documented, challenged, and updated when evidence changes.

2. Payoff tables and decision criteria

A payoff table displays the result of each action under each state of nature. Consider a product launch decision with three alternatives and three demand states. Payoffs are incremental profits in thousands of currency units.

The entries must be measured on a comparable basis. If one alternative includes a capital cost and another does not, the table is misleading. Payoffs should normally be incremental, use the same time horizon, and be stated either before or after tax consistently.

Expected monetary value

The expected monetary value (EMV) of alternative a is the probability-weighted average payoff:

where pᵢ is the probability of state sᵢ, and V(a,sᵢ) is the payoff from alternative a in that state. The probabilities must be non-negative and sum to 1. EMV is appropriate when the decision maker is risk-neutral with respect to the relevant payoffs, or when the payoffs represent sufficiently diversified wealth.

For the product decision:

The EMV criterion selects the large launch, with expected profit of 320 thousand. However, the large launch also has a possible loss of 500 thousand. Expected value is an average, not a forecast or guarantee.

Other criteria under uncertainty

When probabilities are unavailable or distrusted, managers may use alternative criteria:

• Maximax: choose the alternative with the highest possible payoff. This reflects extreme optimism.
• Maximin: choose the alternative with the highest worst-case payoff. This is conservative.
• Minimax regret: choose the alternative with the smallest maximum opportunity loss.
• Laplace criterion: assign equal probabilities to states when no state is regarded as more likely than another.
• Hurwicz criterion: combine the best and worst payoffs using an optimism coefficient α, where 0 ≤ α ≤ 1.

For the Hurwicz criterion:

The coefficient α is not a probability unless explicitly interpreted as one; it represents the decision maker’s degree of optimism. These criteria do not reveal a single objectively correct answer. They make different attitudes toward ignorance explicit.

Regret and opportunity loss

Regret is the payoff forgone because the chosen alternative was not the best action after the state became known. For each state:

In the example, the regret table is:

The minimax-regret decision is the small launch, because its maximum regret is 400, lower than the alternatives’ maximum regret.

3. Decision trees and sequential decisions

A decision tree represents decisions and uncertain events in chronological order. A square conventionally denotes a decision node; a circle denotes a chance node; branches represent actions or states. Decision trees are valuable when managers can learn information and then take a later action.

The standard procedure is rollback analysis:

1. Begin at the terminal payoffs.
2. At each chance node, calculate the probability-weighted value of its branches.
3. At each decision node, retain the branch with the highest value, subject to the organisation’s criterion.
4. Work backwards to the initial decision.

For example, a firm can pay 40 thousand for a pilot test. The test reports favourable or unfavourable results. Management may then launch or abandon. The pilot is useful only if its information changes the decision sufficiently to justify its cost.

A decision tree should show:

• the decision sequence;
• the timing of payments and information;
• mutually exclusive and collectively exhaustive states;
• conditional probabilities where information changes beliefs; and
• terminal payoffs on a consistent basis.

A common error is to multiply probabilities along a path incorrectly or to treat a test result as perfectly accurate. Conditional probabilities must be used when the evidence is imperfect.

4. Expected utility and risk attitude

EMV treats each monetary unit as equally valuable regardless of wealth. In practice, many managers and organisations are risk-averse: a loss of 100 thousand may cause more distress than a gain of 100 thousand provides benefit. Expected utility addresses this issue.

Let u(x) be the utility of payoff x. Expected utility is:

The preferred alternative maximises expected utility, not necessarily expected money. Utility functions should be increasing if more wealth is preferred, and a risk-averse utility function is concave. A common illustrative function is:

This function is not universally correct and cannot be applied to negative payoffs without modification. It is useful only as an illustration when payoffs are non-negative and measured on an appropriate wealth scale.

Risk aversion can also be described locally by the coefficient of absolute risk aversion:

where u′(x) and u″(x) are the first and second derivatives of utility. A positive value indicates local risk aversion. In managerial practice, utility should reflect the organisation’s financial capacity, strategic flexibility, stakeholder obligations, and ability to absorb losses—not merely an executive’s personal preference.

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

The difference between EMV and CE is the risk premium for a risk-averse decision maker, provided both are expressed in comparable monetary units.

5. Value of information

Information has economic value when it improves decisions. The expected value of perfect information (EVPI) is the maximum amount a decision maker should pay to know the future state before choosing an action.

For the launch table, the best payoff in each state is 900, 300, and 50. Therefore:

The best EMV without information is 320, so:

The firm should not pay more than 105 thousand for perfect information, assuming the information arrives before the decision, payoffs remain unchanged, and the probabilities and payoffs are credible.

Expected value of sample information (EVSI) applies to imperfect research such as surveys, pilots, or forecasts:

The net value is:

Information can have strategic value beyond immediate EMV, including learning, option creation, capability development, and risk reduction. Conversely, information may create delay, confidentiality risk, false precision, or decision paralysis.

6. Bayesian updating

Bayesian analysis updates prior beliefs after observing evidence. For a state S and evidence E:

where:

• P(S) is the prior probability;
• P(E|S) is the likelihood of observing evidence if the state is true;
• P(E) is the overall probability of the evidence; and
• P(S|E) is the posterior probability.

Suppose high demand has prior probability 0.30 and a pilot is favourable with probability 0.80 if demand is high and 0.20 if demand is not high. Then:

The favourable result raises the probability of high demand from 0.30 to approximately 0.632. The result is not certainty. Bayesian updating is only as reliable as the prior, likelihood estimates, independence assumptions, and measurement process.

For multiple mutually exclusive states, posterior probabilities are calculated separately and must sum to 1. Conditional dependence between evidence sources must be modelled; otherwise, managers may double-count the same information.

7. Sensitivity analysis and robust choice

Sensitivity analysis tests how a recommendation changes when assumptions vary. Relevant variables include demand, price, variable cost, probability estimates, discount rate, project delay, and salvage value.

A one-way analysis changes one input while holding others constant. A two-way analysis changes two inputs jointly. A break-even value identifies the point at which two alternatives have equal value. If the launch and small-launch alternatives have EMVs based on high-demand probability p, with medium and low probabilities adjusted separately, the analyst should state exactly how probabilities change; otherwise the sensitivity calculation is not well defined.

Useful outputs include:

• a tornado chart ranking assumptions by impact;
• a break-even table;
• downside and upside ranges;
• the probability that value falls below zero; and
• the conditions under which the recommended option changes.

A robust decision is not necessarily the one with the highest central estimate. It may be the alternative that performs acceptably across a broad range of plausible assumptions, can be staged, or can be reversed at low cost.

Scenario planning

Scenario planning develops coherent stories about alternative futures rather than assigning a single forecast. A strong scenario is internally consistent and combines drivers such as economic growth, regulation, technology, customer behaviour, and competitor response. Scenarios are not predictions. They are disciplined tests of strategy.

A practical process is:

1. Define the strategic decision and time horizon.
2. Identify trends, uncertainties, and predetermined elements.
3. Select two or more high-impact, high-uncertainty dimensions.
4. Construct internally consistent scenarios.
5. Test current strategy in each scenario.
6. Identify no-regret moves, contingent actions, and early-warning indicators.

Scenario planning complements probability analysis. It is particularly useful when probabilities are unstable or when structural change makes historical frequencies unreliable.

8. Monte Carlo simulation

Monte Carlo simulation represents uncertain inputs with probability distributions and repeatedly samples from them to generate a distribution of outcomes. Unlike a single-point estimate, it shows ranges, percentiles, correlations, and the probability of adverse results.

A basic procedure is:

1. Define the outcome equation, such as profit = revenue − variable cost − fixed cost.
2. Specify distributions for uncertain inputs.
3. Specify dependencies and correlations.
4. Draw random values for all inputs.
5. Calculate the outcome.
6. Repeat many times.
7. Summarise the simulated distribution and test assumptions.

For a simulation with N trials and outcome Yⱼ in trial j, the estimated mean is:

The simulated probability of exceeding a threshold c is:

Simulation outputs are estimates and contain sampling error. More trials generally improve precision, but additional trials cannot correct a badly specified model. Distributions must have defensible bounds and units. For example, demand cannot be negative, and a triangular distribution requires a minimum, most likely value, and maximum.

Correlations are essential. Advertising spend and demand, exchange rates and imported costs, or interest rates and refinancing costs may move together. Treating correlated inputs as independent can materially understate risk. Simulation should therefore be combined with sensitivity analysis and expert review.

9. Illustrative case: the electric delivery van decision

Background

The fictional MetroFresh Foods operates a regional delivery fleet. It must decide whether to purchase 40 electric vans, lease a smaller pilot fleet, or continue using diesel vans. The decision horizon is three years. Management is uncertain about battery prices, charging reliability, fuel prices, and customer willingness to pay for low-emission delivery.

Problem

The purchase alternative has the highest projected average contribution but requires an immediate capital commitment and may be exposed to charging infrastructure delays. The pilot has lower expected contribution but creates learning and preserves flexibility.

Analysis

The finance team first constructs a payoff table for high, moderate, and low adoption of green delivery services. It then:

• calculates EMV and downside loss;
• models a charging pilot as sample information;
• updates adoption probabilities after pilot results;
• runs sensitivity analysis on energy prices and utilisation; and
• conducts a Monte Carlo simulation using distributions for kilometres travelled, energy cost, maintenance cost, and customer demand.

The simulation shows that purchasing has a mean contribution 8% higher than the pilot, but a 22% probability of producing a negative incremental contribution. The pilot has a 7% probability of a negative result and provides an option to expand after six months. Its expected contribution is lower, but its worst plausible outcomes are substantially less severe.

Approach and solution

The board adopts a staged strategy: lease ten vans, install modular charging capacity, and define expansion triggers. Expansion occurs if utilisation exceeds 75%, charging downtime remains below 3%, and customer retention improves by at least 2 percentage points. A review committee monitors whether the assumptions remain valid.

Outcome

After six months, favourable operating evidence raises the posterior probability of strong adoption. The firm expands gradually rather than committing all capital initially. This approach does not maximise the original point estimate, but it limits irreversible exposure and creates information before a larger commitment.

Learning points

The case illustrates that:

• EMV alone may favour a high-exposure option;
• information has value when it changes subsequent action;
• staging can be a real option;
• scenario triggers convert analysis into governance; and
• model outputs should be treated as conditional evidence, not objective truth.

10. Governance, ethics, and implementation

Decision models can create ethical risks. Managers may manipulate probabilities to justify a preferred project, omit inconvenient scenarios, or present a simulation’s average as a guaranteed result. Governance should require transparent assumptions, version control, independent challenge, and clear separation between observed data and judgement.

Important controls include:

• documenting data sources and probability elicitation;
• reporting ranges and downside probabilities, not only expected values;
• conducting pre-mortems that ask why the project might fail;
• testing model outputs against historical or external benchmarks;
• identifying who bears losses and who receives benefits;
• protecting confidential and personal data used in forecasting; and
• assigning accountability for monitoring and updating assumptions.

AI-supported forecasting can process large datasets and generate scenarios, but it may reproduce historical bias, conceal uncertainty, or produce confident-looking unsupported probabilities. Human oversight should review data quality, explainability, distributional effects, cybersecurity, and whether automated recommendations are appropriate for high-impact decisions. A model should support managerial judgement, not replace responsibility.

Best-practice checklist

Before approving a major decision under risk or uncertainty, ask:

1. Are the alternatives genuinely feasible and comparable?
2. Are states of nature exhaustive and mutually exclusive?
3. Are payoffs incremental, consistently timed, and correctly measured?
4. Are probabilities evidence-based, elicited transparently, and updated when necessary?
5. Has the analysis included downside exposure and not just EMV?
6. Could a pilot, staged investment, or contractual flexibility create valuable information?
7. Which assumptions drive the recommendation?
8. How does the decision perform across credible scenarios?
9. Are dependencies and correlations represented in simulation?
10. What triggers will cause management to revise, pause, or terminate the decision?

Chapter Summary

Decisions under risk and uncertainty require managers to combine structured analysis with judgement. Payoff tables organise alternatives, states, and consequences. Expected monetary value provides a probability-weighted benchmark, while minimax, minimax regret, Laplace, and Hurwicz criteria make different attitudes toward uncertainty explicit. Decision trees support sequential choices and rollback analysis.

Expected utility recognises that managers may value gains and losses differently. EVPI and EVSI quantify the economic value of information, while Bayesian updating revises beliefs when evidence arrives. Sensitivity analysis identifies influential assumptions and break-even conditions. Scenario planning explores coherent alternative futures when probabilities are unstable. Monte Carlo simulation estimates distributions of outcomes but cannot rescue a poorly specified model.

Strategically robust choices often combine staged commitment, information acquisition, monitoring triggers, and downside protection. Ethical governance requires transparent assumptions, appropriate human oversight, and honest communication of uncertainty.

Key Terms

Risk: A situation in which possible states and defensible probabilities are available.

Uncertainty: A situation in which states, probabilities, or both are poorly known.

Ambiguity: Uncertainty about the appropriate model or probability distribution.

Payoff: The consequence of an alternative under a particular state of nature.

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

Expected monetary value: The probability-weighted average payoff of an alternative.

Regret: The payoff forgone relative to the best action after the state is known.

Maximax: A criterion selecting the highest possible payoff.

Maximin: A criterion selecting the best among the worst possible payoffs.

Minimax regret: A criterion minimising the largest possible regret.

Expected utility: The probability-weighted utility of outcomes.

Risk aversion: Preference for a certain outcome over a fair risky outcome.

Certainty equivalent: The sure amount with utility equal to a risky prospect.

EVPI: Expected value of perfect information.

EVSI: Expected value of sample information.

Bayesian updating: Revising prior beliefs using observed evidence.

Sensitivity analysis: Testing how results change when assumptions vary.

Scenario planning: Developing coherent alternative future environments.

Monte Carlo simulation: Repeated random sampling to estimate an outcome distribution.

Real option: Flexibility to delay, expand, contract, or abandon an investment.

Review Questions

Short Answer Questions

1. Distinguish risk from uncertainty and ambiguity.
2. What assumptions are required for EMV analysis?
3. Define regret and explain how a regret table is constructed.
4. What is the difference between EVPI and EVSI?
5. Why are correlations important in Monte Carlo simulation?

Long Answer Questions

1. Explain how a decision tree supports sequential managerial decisions, including rollback analysis.
2. Compare EMV, expected utility, minimax regret, and scenario planning as decision approaches.
3. Explain Bayesian updating and discuss the practical challenges of estimating priors and likelihoods.
4. Describe a robust process for applying sensitivity analysis and Monte Carlo simulation to an investment decision.
5. Discuss the ethical and governance risks associated with AI-supported decisions under uncertainty.

Analytical / Application Questions

1. Using the product-launch payoff table in this chapter, calculate EMV for each alternative and EVPI. Explain why the EMV and minimax-regret recommendations differ.
2. A manager estimates that a favourable pilot result occurs with probability 0.80 under high demand and 0.20 otherwise. If the prior probability of high demand is 0.30, calculate the posterior probability after a favourable result and interpret it.
3. Design a staged decision strategy for a capital project whose expected value is high but whose downside probability is material. Specify information milestones, expansion or abandonment triggers, and the assumptions you would test through simulation.

Multiple Choice Questions with Answers

Question 1

Which condition most directly supports the use of expected monetary value as the primary decision criterion?

A. The decision maker is risk-neutral for the relevant payoffs.

B. All possible outcomes are negative.

C. No probabilities can be assigned to states.

D. The decision is irreversible.

Correct Answer: A. The decision maker is risk-neutral for the relevant payoffs.

Question 2

Using the chapter’s product-launch table, what is the EMV of the small launch?

A. 250 thousand

B. 285 thousand

C. 320 thousand

D. 425 thousand

Correct Answer: B. 285 thousand

Question 3

What does EVPI measure?

A. The cost of collecting historical data

B. The maximum value of knowing the future state before deciding

C. The variance of project returns

D. The probability of making a correct forecast

Correct Answer: B. The maximum value of knowing the future state before deciding

Question 4

In Bayesian analysis, P(S|E) represents:

A. The probability of evidence given the state

B. The prior probability of the state

C. The probability of the state given the evidence

D. The probability that evidence is independent

Correct Answer: C. The probability of the state given the evidence

Question 5

Which criterion selects the alternative with the smallest maximum regret?

A. Maximax

B. Maximin

C. Laplace

D. Minimax regret

Correct Answer: D. Minimax regret

Question 6

A risk-averse utility function is generally:

A. Concave over the relevant wealth range

B. Constant for all wealth levels

C. Decreasing in wealth

D. Undefined for every positive payoff

Correct Answer: A. Concave over the relevant wealth range

Question 7

What is a principal purpose of sensitivity analysis?

A. To guarantee the forecast is correct

B. To identify how changes in assumptions affect the recommendation

C. To eliminate all uncertainty

D. To replace managerial judgement with arithmetic

Correct Answer: B. To identify how changes in assumptions affect the recommendation

Question 8

Monte Carlo simulation primarily produces:

A. A single certain forecast

B. A distribution of possible outcomes

C. A legal contract

D. A perfect-information decision

Correct Answer: B. A distribution of possible outcomes

Question 9

If the prior probability of high demand is 0.30, and a favourable result has likelihoods 0.80 under high demand and 0.20 otherwise, what is P(High|Favourable), approximately?

A. 0.300

B. 0.500

C. 0.632

D. 0.800

Correct Answer: C. 0.632

Question 10

Which practice best improves the governance of simulation-based decisions?

A. Reporting only the mean outcome

B. Hiding assumptions to avoid confusion

C. Documenting assumptions and reporting downside probabilities

D. Treating model output as objective fact

Correct Answer: C. Documenting assumptions and reporting downside probabilities