Decision Analysis, Risk and Simulation

Introduction

International business decisions are rarely made with complete information. Firms must choose production capacity, market-entry modes, sourcing locations and currency strategies before exchange rates, demand, tariffs or competitor actions are known. Decision analysis provides a disciplined way to structure such choices.

A decision model does not eliminate uncertainty. It makes assumptions, alternatives, consequences and managerial preferences explicit. This improves consistency, reveals important information gaps and allows decision makers to examine whether a recommendation remains defensible when assumptions change.

This chapter distinguishes risk—where probabilities can reasonably be assigned—from ambiguity, where probabilities are incomplete or disputed. It then develops payoff tables, decision trees, expected utility, sensitivity and scenario analysis, Monte Carlo simulation, and downside-risk measures.

Learning Objectives

After completing this chapter, students should be able to:

1. Construct and interpret payoff and regret tables.
2. Apply decision criteria under risk and ambiguity.
3. calculate expected monetary value and the value of information.
4. Roll back a decision tree using conditional probabilities.
5. Incorporate risk preferences through expected utility and certainty equivalents.
6. conduct sensitivity, break-even and scenario analysis.
7. Design and interpret a Monte Carlo simulation.
8. Calculate and evaluate variance, value at risk and expected shortfall.
9. Recognize governance, ethical and model-risk issues in analytics-based decisions.

1. Foundations of Decision Analysis

A decision problem normally contains four elements:

• Decision alternatives: actions under managerial control.
• States of nature: uncertain external conditions.
• Consequences: outcomes produced by each action–state combination.
• Preferences: the decision maker's valuation of those outcomes.

A payoff is a consequence expressed in a relevant unit, such as profit in euros, market share, emissions or a utility score. Alternatives should be mutually distinguishable, and states of nature should be mutually exclusive and collectively exhaustive for the model's intended scope.

1.1 Certainty, risk and ambiguity

Probabilities may be estimated from historical data, market research, forecasting models or expert judgement. Historical frequencies are not automatically appropriate when structural breaks, geopolitical changes or new technologies alter the business environment.

2. Payoff Tables and Decision Criteria

Let a_i denote alternative i, s_j state j, and x_ij the monetary payoff from selecting a_i when s_j occurs. A payoff table displays all modeled values x_ij.

2.1 Expected monetary value

When state probabilities are credible and the decision maker is risk-neutral over the relevant payoff range, the expected monetary value of alternative a_i is:

Here, p_j is the probability of state s_j, p_j ≥ 0, and ∑p_j = 1. Because probability is dimensionless, EMV has the same unit as the payoff—for example, € million.

EMV is a probability-weighted long-run average. It is not a prediction that the expected value will occur in a single decision.

2.2 Criteria under ambiguity

When probabilities cannot be justified, several criteria may be reported rather than concealing uncertainty behind arbitrary weights.

• Maximax: choose the alternative with the highest possible payoff. It is strongly opportunity-seeking.
• Maximin: choose the alternative with the highest worst-case payoff. It emphasizes protection against the worst modeled state.
• Laplace: assign equal probabilities to all states. This is appropriate only when equal weighting is defensible, not merely convenient.
• Hurwicz: combine each alternative's best and worst payoffs using an optimism coefficient.
• Minimax regret: minimize the largest opportunity loss relative to the best action in each state.

For a profit problem, regret is:

Regret is nonnegative and has the same monetary unit as the payoffs. Minimax regret chooses the alternative with the smallest value of max_j(r_ij).

For the Hurwicz criterion:

The coefficient α lies between 0 and 1. A larger α places more emphasis on the best outcome. This formula assumes that payoffs are being maximized; a cost problem requires suitable modification.

3. Decision Trees and Bayesian Updating

A decision tree represents sequential decisions and uncertain events. A square node denotes a decision; a circle denotes a chance event. Branches from a chance node must have probabilities summing to one.

The tree is evaluated from right to left by rollback:

1. At each terminal branch, record the payoff.
2. At each chance node, calculate the probability-weighted value.
3. At each decision node, retain the branch with the greatest expected value or utility.

Figure: A decision tree showing an immediate launch and a market-research alternative.

Worked rollback calculation

Suppose an immediate launch earns €4 million under strong demand and loses €2 million under weak demand. The probabilities are 0.6 and 0.4:

Research costs €0.3 million. A favourable report occurs with probability 0.5 and gives posterior demand probabilities of 0.8 strong and 0.2 weak. An unfavourable report also occurs with probability 0.5 and gives probabilities of 0.4 strong and 0.6 weak.

Because both conditional values exceed the zero payoff from not launching, the firm launches after either report. The research strategy is therefore worth:

Research should not be purchased because €1.3 million is below the immediate-launch value of €1.6 million. Information has value only when it may improve the subsequent action sufficiently to cover its cost.

3.1 Value of perfect information

The expected value with perfect information is:

The expected value of perfect information is:

EVPI is an upper bound on what a risk-neutral decision maker should pay for any information about the state. The expected value of sample information is calculated similarly but reflects an imperfect signal and its possible decision consequences.

4. Risk Preferences and Expected Utility

EMV assumes linear preferences for money. In practice, a €1 million loss may matter more to a liquidity-constrained firm than an equal-probability €1 million gain. Expected utility represents such risk preferences.

For wealth or payoff x, expected utility is:

The function u(x) represents preferences. A concave utility function indicates risk aversion, a linear function risk neutrality, and a convex function risk seeking over the relevant range.

The certainty equivalent is the guaranteed payoff CE satisfying:

The risk premium for a risky monetary payoff X is:

For a risk-averse decision maker, the risk premium is normally nonnegative.

Worked expected-utility example

A project pays €2 million with probability 0.7 and loses €1 million with probability 0.3. Its EMV is:

Assume the firm's preferences over payoffs in € millions are represented by the exponential utility function:

Here, R > 0 is risk tolerance in € millions. This function has constant absolute risk aversion and is appropriate only if that behavioral assumption is reasonable. Let R = 2.

Solving 1 − e^−CE/2 = 0.2479 gives:

Thus:

The firm is indifferent between the risky project and a guaranteed payoff of approximately €0.570 million under the assumed utility function. Utility parameters should be elicited from consistent choices or organizational risk limits rather than selected merely to justify a preferred decision.

5. Sensitivity and Scenario Analysis

5.1 Sensitivity analysis

Sensitivity analysis changes one or more inputs to determine how strongly the recommendation depends on assumptions. Common inputs include demand probabilities, exchange rates, discount rates, unit margins and fixed costs.

If the value of alternative A is V_A(z) and that of B is V_B(z), a break-even value of input z solves:

One-way sensitivity is easy to communicate but can be misleading when variables are correlated. Multiway sensitivity and simulation are preferable when exchange rates, inflation, demand and financing costs move jointly.

5.2 Scenario analysis

A scenario is a coherent combination of assumptions, not simply an optimistic or pessimistic change to every variable. For example, weak global growth could reduce demand, lower commodity prices and cause a currency depreciation simultaneously.

Scenario probabilities or weights must be described accurately. Giving the greatest weight to high demand is high-demand emphasis, not a conservative weighting scheme. Labels such as “conservative” should be reserved for clearly defined assumptions rather than used rhetorically.

Stress tests differ from forecasts. Their purpose is to identify vulnerability under severe but plausible conditions, even if precise probabilities cannot be assigned.

6. Monte Carlo Simulation

Analytical calculations become difficult when many uncertain inputs interact nonlinearly. Monte Carlo simulation approximates an output distribution by repeatedly sampling from specified input distributions.

For iteration k, a model can be written generally as:

Here, g(·) is the business model, X_l^(k) is the sampled value of uncertain input l, and Y^(k) is the resulting output. Units are determined by the model; if g calculates profit, Y may be measured in euros.

6.1 Simulation procedure

1. Define the decision and output measure.
2. Separate uncertain inputs from controllable choices.
3. Select distributions using data and expert knowledge.
4. Model dependencies, including correlations.
5. generate random or quasi-random samples.
6. Recalculate the model for many iterations.
7. summarize the output distribution and decision-relevant risks.
8. validate, document and stress-test the model.

A common normal-distribution representation is:

where Z is standard normal, μ is the mean and σ > 0 is the standard deviation. This is suitable only when a normal model is substantively defensible. Normal distributions may be inappropriate for bounded percentages, nonnegative prices or strongly skewed losses.

The standard error of a simulated sample mean declines approximately as:

Here, s_Y is the sample standard deviation and N is the number of independent iterations. Quadrupling N therefore approximately halves this standard error. Tail measures usually require more iterations than mean estimates.

6.2 Completed simulation and tail-risk calculation

Consider a project with the following one-year profit distribution:

The completed simulation produces:

The simulated second moment is €² 13.673m². Using the iteration distribution rather than the unbiased sample correction:

There are 200 loss-making runs, so the estimated probability of a negative profit is 20%. Define loss as L = −X, where X is profit. The simulated loss outcomes are −€6m, −€2m, €1m and €4m.

The empirical 95% value at risk is the 95th percentile of loss. The nearest-rank position is 950. Sorted losses up to positions 801–953 equal €1 million, so:

For empirical expected shortfall, average the worst 5% of the 1,000 runs—the 50 largest losses. These comprise 47 losses of €4 million and 3 losses of €1 million:

This illustrates why VaR alone can hide severity beyond its threshold. In the exact theoretical distribution, the worst 5% all lose €4 million, so theoretical 95% expected shortfall equals €4 million. The difference is simulation sampling error.

7. Risk Measures and Their Interpretation

7.1 Variance and standard deviation

For random payoff X with mean μ, variance is:

Standard deviation is σ = √Var(X). It has the same unit as the payoff. Standard deviation treats upside and downside deviations symmetrically, so it should not automatically be interpreted as loss risk.

7.2 Semideviation and probability of loss

Downside semivariance relative to target T is:

Its square root is downside semideviation. The target could be zero profit, a required return or a debt-service threshold. The probability P(X < T) is also managerially useful but does not show the magnitude of a shortfall.

7.3 Value at risk and expected shortfall

For loss L, value at risk at confidence level α is the lower loss threshold whose cumulative probability reaches α:

Expected shortfall is the average loss in the worst 1 − α fraction of cases. For continuous distributions it may be written:

For discrete distributions, this conditional formula may include more than the intended tail because of probability mass at VaR. The more general definition averages the upper quantiles, using fractional probability mass where necessary. Expected shortfall is therefore generally more informative about tail severity.

8. Illustrative International Business Case: Aurora Foods

Background

Aurora Foods, a fictional European producer, is considering how to enter an overseas market. Management can build a small facility, build a large facility or export from its existing plant. Net present values are stated in € millions.

Problem

The estimated probabilities are 0.55 for low demand, 0.25 for medium demand and 0.20 for high demand.

Analysis

The expected values are:

Thus, export has the highest EMV. The regret table is:

Export also minimizes maximum regret.

Suppose medium demand remains fixed at 0.25, while high-demand probability is h and low-demand probability is 0.75 − h, where 0 ≤ h ≤ 0.75. Then:

Export exceeds small-facility value when:

The estimated value h = 0.20 is above this break-even probability.

An alternative weighting of 0.25 low, 0.25 medium and 0.50 high produces:

• Small facility: €2.00 million.
• Large facility: €2.25 million.
• Export: €4.00 million.

These weights emphasize high demand and should not be described as conservative.

Approach and solution

Management selects export as the provisional strategy because it has the highest EMV, lowest maximum regret and remains preferable to the small facility across a meaningful probability range. Before implementation, managers should simulate exchange rates, shipping costs, tariffs and demand jointly because the simple scenario model treats each state as a complete bundle.

Outcome

Aurora adopts a staged export entry with contractual capacity options. This preserves flexibility while managers collect market information. Expansion triggers are linked to observed demand and contribution margin rather than to calendar dates alone.

Learning points

• Correct probability weighting is essential; small arithmetic errors can reverse recommendations.
• Scenario labels must reflect their actual assumptions.
• A robust decision should be examined using more than one criterion.
• Flexible, staged commitments can have strategic value not captured by static payoffs.

9. Model Risk, Governance and Contemporary Practice

Decision models can create false confidence when their structure is treated as objective fact. Important limitations include omitted states, unreliable probabilities, unstable correlations, poor-quality data and unmodeled strategic responses.

International models require particular attention to exchange-rate convertibility, political intervention, transfer restrictions, tax rules and supply-chain concentration. Correlations often increase during crises, reducing the diversification benefit predicted from normal-period data.

Good governance requires:

• documented assumptions, data sources and model ownership;
• independent review of formulas and code;
• version control and reproducible simulation seeds;
• comparison with historical outcomes and benchmark models;
• explicit reporting of uncertainty and tail exposure;
• approval limits for high-impact model-based decisions;
• periodic recalibration and monitoring.

Artificial intelligence can help estimate demand, identify scenarios and automate simulation. However, opaque forecasts should not be converted mechanically into precise probabilities. Human review remains necessary to detect leakage, bias, unstable relationships and implausible causal conclusions.

Ethical analysis also extends beyond shareholder profit. A sourcing decision may transfer environmental, labor or safety risk to suppliers and communities. Multicriteria decision analysis, constraints or shadow prices can incorporate material nonfinancial consequences. Sensitivity analysis should disclose when a recommendation depends on controversial valuations of social or environmental harm.

Chapter Summary

Decision analysis structures alternatives, uncertain states, consequences and preferences. Under credible probabilities, expected monetary value supports risk-neutral choices, while expected utility incorporates risk attitudes through certainty equivalents and risk premiums. Under ambiguity, maximin, Hurwicz and minimax-regret criteria provide alternative perspectives.

Decision trees represent sequential choices and information, while rollback determines the preferred strategy. The value of information depends on whether new evidence changes subsequent action. Sensitivity and scenario analysis reveal break-even assumptions and vulnerabilities. Monte Carlo simulation extends analysis to complex, interacting uncertainties, producing full outcome distributions rather than single-point estimates. Variance, probability of loss, VaR and expected shortfall describe different dimensions of risk. Reliable application requires transparent assumptions, correct arithmetic, realistic dependency modeling, validation and strong governance.

Key Terms

Ambiguity: A condition in which probabilities are unknown, unreliable or disputed.

Certainty equivalent: A guaranteed amount considered equally desirable as a risky prospect.

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

Downside semivariance: Expected squared shortfall below a specified target.

Expected monetary value: The probability-weighted average monetary payoff.

Expected shortfall: The average loss in the worst specified fraction of outcomes.

Expected utility: Probability-weighted utility of uncertain outcomes.

Hurwicz criterion: A weighted combination of an alternative's best and worst payoffs.

Maximin criterion: Selection of the alternative with the best worst-case payoff.

Minimax regret: Selection of the alternative with the smallest maximum opportunity loss.

Monte Carlo simulation: Repeated random sampling used to approximate an output distribution.

Payoff table: A matrix showing consequences for alternatives and states of nature.

Regret: The payoff sacrificed by not choosing the best action for the realized state.

Risk: Uncertainty for which probabilities can reasonably be assigned.

Risk premium: Expected payoff minus certainty equivalent.

Scenario analysis: Evaluation of coherent combinations of uncertain assumptions.

Sensitivity analysis: Examination of how results change when model inputs change.

State of nature: An uncertain external condition affecting consequences.

Value at risk: A loss quantile at a specified confidence level and horizon.

Value of perfect information: The maximum expected gain from learning the state before deciding.

Review Questions

Short Answer Questions

1. Distinguish risk from ambiguity.
2. What conditions should states of nature satisfy in a payoff table?
3. Why is EMV not a forecast of the outcome of one decision?
4. What does a certainty equivalent represent?
5. Why can expected shortfall be more informative than VaR?

Long Answer Questions

1. Compare expected monetary value, maximin and minimax regret as decision criteria.
2. Explain how a decision tree is constructed and evaluated by rollback.
3. Discuss the assumptions and limitations of expected-utility analysis.
4. Explain how Monte Carlo simulation should be designed, validated and governed.
5. Evaluate the ethical and managerial risks of relying on AI-generated probabilities in international business decisions.

Analytical / Application Questions

1. A market entry pays €8 million under high demand and loses €3 million under low demand. If high demand has probability 0.4, calculate EMV and identify the break-even high-demand probability.
2. Using the Aurora Foods payoff table, calculate the Hurwicz value of every alternative when α = 0.6. Which alternative is selected?
3. Design a Monte Carlo model for an exporting decision involving uncertain foreign demand, exchange rates, tariffs and shipping costs. Specify distributions, dependencies, outputs and validation procedures.

Multiple Choice Questions with Answers

Question 1

Using Aurora Foods' probabilities of 0.55 low demand, 0.25 medium demand and 0.20 high demand, what is the EMV of the large facility?

A. −€1.05 million

B. €0.15 million

C. €1.90 million

D. €2.25 million

Correct Answer: B. €0.15 million

Question 2

For the Aurora Foods case, which alternative minimizes maximum regret?

A. Small facility

B. Large facility

C. Export

D. All three have equal maximum regret

Correct Answer: C. Export

Question 3

A project has expected payoff €1.100 million and certainty equivalent €0.570 million. What is its risk premium?

A. €0.530 million

B. €0.570 million

C. €1.100 million

D. €1.670 million

Correct Answer: A. €0.530 million

Question 4

In the market-research decision-tree example, what is the expected value of the research strategy after its €0.3 million cost?

A. €0.4 million

B. €1.3 million

C. €1.6 million

D. €2.8 million

Correct Answer: B. €1.3 million

Question 5

In the completed 1,000-run simulation, what is the empirical 95% VaR of loss?

A. −€2 million

B. €1 million

C. €3.82 million

D. €4 million

Correct Answer: B. €1 million

Question 6

In the same simulation, what is the empirical 95% expected shortfall based on the worst 50 runs?

A. €1.00 million

B. €2.46 million

C. €3.82 million

D. €4.70 million

Correct Answer: C. €3.82 million

Question 7

Which expression correctly defines the expected value of perfect information?

A. Best EMV minus expected value with perfect information

B. Expected value with perfect information minus the best current EMV

C. Expected utility minus certainty equivalent

D. Maximum payoff minus minimum payoff

Correct Answer: B. Expected value with perfect information minus the best current EMV

Question 8

Why should correlations be modeled in a Monte Carlo analysis?

A. They guarantee that simulated profits are normally distributed.

B. They eliminate sampling error from tail estimates.

C. They represent how uncertain inputs may move together.

D. They ensure every scenario has equal probability.

Correct Answer: C. They represent how uncertain inputs may move together.

Question 9

How should scenario weights of 0.25 low demand, 0.25 medium demand and 0.50 high demand be described?

A. Conservative because low demand has positive weight

B. Risk-neutral because the weights sum to one

C. High-demand emphasis because high demand receives the greatest weight

D. Probability-free because the weights are subjective

Correct Answer: C. High-demand emphasis because high demand receives the greatest weight

Question 10

Which decision environment exists when relevant probabilities are unknown or seriously disputed?

A. Certainty

B. Risk

C. Ambiguity

D. Deterministic optimization

Correct Answer: C. Ambiguity