Simulation, Risk and Decision Analysis

Introduction and Context

Business decisions are rarely made with complete information. A retailer must choose inventory before demand is known; a bank must approve loans before repayment is observed; and a manufacturer may invest in capacity before prices, regulation, or technology become clear. Advanced business analytics helps decision makers represent this uncertainty explicitly rather than hiding it behind a single forecast.

This chapter develops an integrated toolkit for decisions under uncertainty. Decision trees represent sequential choices, utility concepts incorporate preferences toward risk, sensitivity analysis identifies influential assumptions, scenario planning explores coherent alternative futures, and Monte Carlo simulation estimates the distribution of possible outcomes. Together, these methods support risk exposure assessment, trade-off analysis, robustness evaluation, and the value-of-information perspective.

The objective is not to produce a mechanically precise answer. It is to make assumptions visible, compare alternatives consistently, and identify what should be learned, monitored, or controlled before committing resources.

Learning Objectives

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

1. Distinguish decisions under certainty, risk, and uncertainty.
2. Construct and evaluate decision trees using expected monetary value.
3. Explain expected utility and risk attitudes.
4. Perform one-way, two-way, and threshold sensitivity analysis.
5. Develop useful scenarios without treating them as simple forecasts.
6. Design and interpret a Monte Carlo simulation.
7. Estimate the expected value of perfect and sample information.
8. Evaluate robustness, model risk, ethics, and governance in analytical decisions.

1. Decisions Under Uncertainty

A decision problem consists of alternatives, uncertain states of nature, consequences, and information available at the time of choice. Let an alternative be denoted by a, a state of nature by s, and the resulting payoff by V(a,s). If probabilities are credibly assigned to states, the situation is commonly called decision-making under risk. If probabilities cannot be justified, it is decision-making under uncertainty.

A payoff may be profit, net present value, customer retention, service level, or a broader measure of value. It is essential to define the time horizon, currency, and whether values are nominal or inflation-adjusted. Comparing a one-year profit with a ten-year net present value without adjustment is not analytically valid.

Expected monetary value (EMV) is:

Here, p_i is the probability of state s_i, V(a,s_i) is the payoff under alternative a, and probabilities must satisfy 0 ≤ p_i ≤ 1 and ∑p_i = 1. EMV is appropriate when monetary outcomes can be treated as the relevant value measure and the decision maker is risk neutral, or when the analysis is explicitly being used as a financial benchmark.

EMV should not be interpreted as the outcome that will occur. It is a probability-weighted average across repeated or comparable situations. A project with an EMV of $40,000 may still have a substantial probability of losing money.

Terminology of Decision Trees

A decision tree is a chronological representation of choices and uncertain events.

• A decision node, usually represented by a square, is controlled by the decision maker.
• A chance node, usually represented by a circle, represents an uncertain event.
• A branch represents an action or event.
• A terminal node contains the resulting payoff.
• Rollback analysis evaluates the tree from the end toward the initial decision.

At a chance node, the value is the probability-weighted average of successor values. At a decision node, the decision maker selects the branch with the greatest value, subject to feasibility and the chosen objective.

Figure: A simple decision tree with a launch decision and uncertain demand.

2. Worked Decision-Tree Analysis

Consider a firm deciding whether to launch a service. Launching produces $120,000 if demand is high, with probability 0.6, and a loss of $60,000 if demand is low, with probability 0.4. Not launching produces $0.

The EMV of launching is:

The EMV of not launching is $0. Therefore, under a risk-neutral EMV criterion, launching is preferred. However, the probability of a loss is 0.4. A risk-averse manager may reject the launch if the potential loss threatens liquidity or violates a risk limit.

A decision tree becomes more realistic when actions occur after information arrives. For example, the firm might conduct a market test before launching. The test has a cost, may be imperfect, and may change the probabilities assigned to high and low demand. The correct analysis must then include conditional probabilities such as P(High demand | Positive test). Treating a test result as certain evidence is a common analytical error.

Expected Value of Perfect Information

The expected value of perfect information (EVPI) measures the maximum amount a decision maker should pay to know the future state before choosing an action. It is calculated as:

The first term is the expected payoff if the best action could be selected separately for every state. The second is the best EMV without additional information. EVPI is never negative under the same assumptions because information permits, but does not require, better choices.

In the launch example, suppose that under high demand launching yields $120,000 and not launching yields $0; under low demand launching yields −$60,000 and not launching yields $0. The expected value with perfect information is:

0.6($120,000) + 0.4($0) = $72,000. Since the best current EMV is $48,000, EVPI is $24,000. No perfect forecast should cost more than $24,000 in expected monetary terms.

Expected Value of Sample Information

Information is usually imperfect. The expected value of sample information (EVSI) is the expected value of using a test, survey, pilot, or forecast, less the value without that information. If the information itself costs C, its net value is:

EVSI cannot exceed EVPI when both are calculated consistently. A positive net EVSI does not guarantee a positive cash outcome; it indicates that information improves the expected decision value.

Bayes’ rule is often required to update beliefs:

Here, S_i is a state, E is observed evidence, and P(E) = ∑_j P(E | S_j)P(S_j). The rule assumes that the probabilities describe the same event definitions and that the evidence model is appropriate.

3. Utility and Risk Preferences

Money is not always the correct measure of value. The utility of an outcome reflects its desirability to a particular decision maker. A risk-neutral person has linear utility, so equal expected money produces equal expected utility. A risk-averse person has a concave utility function: an additional dollar provides less utility as wealth increases. A risk-seeking person has a convex utility function over the relevant range.

Expected utility is:

The preferred action maximizes expected utility rather than necessarily maximizing EMV. Utility functions must be defined over the relevant outcome range and should reflect the decision maker’s preferences, not the analyst’s assumptions.

A frequently used risk-averse utility function is exponential utility:

where x is a monetary outcome and R > 0 is the coefficient of absolute risk tolerance, measured in currency units. Larger R indicates greater tolerance for risk. This function is useful because it is increasing and concave, but it is not universally appropriate. It also assumes constant absolute risk tolerance, which may be unrealistic when wealth changes substantially.

The certainty equivalent (CE) is the guaranteed amount that gives the same utility as the risky prospect:

For exponential utility:

The risk premium is:

For a risk-averse decision maker, the risk premium is normally non-negative. Utility analysis is especially relevant when downside losses affect solvency, safety, employment, reputation, or strategic flexibility.

4. Sensitivity and Robustness Analysis

A model is sensitive to an input when modest changes in that input materially change the recommendation or outcome. Sensitivity analysis does not predict which input will change; it reveals where attention is most valuable.

One-Way Sensitivity Analysis

One input is varied while all other inputs are held constant. For the launch example, let p be the probability of high demand. The launch EMV is:

Launch is preferred when:

Thus, the break-even probability of high demand is 33.33%. If management believes the probability could reasonably fall below this threshold, the decision is fragile.

Sensitivity can also vary costs, prices, conversion rates, or discount rates. A tornado chart ranks the changes in outcome caused by plausible low and high values, but the chart must state the ranges and whether inputs are varied independently.

Two-Way and Threshold Analysis

Two-way sensitivity varies two inputs simultaneously, such as price and volume. It may reveal interaction effects that one-way analysis misses. Threshold analysis identifies the value at which a decision changes. Thresholds are useful for operational monitoring: if demand falls below a specified level, management may pause investment or activate a contingency plan.

Robustness

A decision is robust when it remains acceptable across a wide range of plausible assumptions, scenarios, and model specifications. Robustness is not the same as selecting the alternative with the highest expected value. A slightly lower expected return may be preferable if it has substantially lower downside exposure and remains competitive across multiple futures.

Useful robustness questions include:

• Does the recommendation survive plausible changes in probabilities?
• Does it remain acceptable under adverse but credible conditions?
• Are results driven by one poorly measured assumption?
• Can the decision be staged, delayed, reversed, or hedged?
• Are operational and ethical constraints satisfied in every relevant scenario?

5. Scenario Planning

Scenario planning develops several internally coherent descriptions of how the external environment could evolve. Scenarios are not optimistic, expected, and pessimistic versions of a single forecast unless their causal assumptions are explicitly different. They may combine trends and uncertainties involving regulation, technology, customer behavior, competitors, supply, and macroeconomic conditions.

A disciplined scenario process is:

1. Define the decision and time horizon.
2. Identify major drivers and predetermined trends.
3. Assess uncertainty and impact.
4. Select critical uncertainties that are not redundant.
5. Construct a small set of contrasting, plausible scenarios.
6. Describe causal logic, not merely labels.
7. Test strategies against each scenario.
8. Identify indicators and trigger points for updating action.

Scenario planning differs from probability forecasting. A scenario need not receive a probability; its purpose may be to expose strategic vulnerabilities and stimulate contingency planning. Probabilities can nevertheless be assigned when evidence supports them, particularly for quantitative evaluation.

A useful strategy table might compare alternatives by outcome, capability requirements, downside, and reversibility. Strategies that perform adequately across all scenarios are often more robust than strategies optimized for one assumed future. Real-options thinking strengthens this analysis: a small pilot may preserve the option to expand after uncertainty is reduced.

6. Monte Carlo Simulation

Monte Carlo simulation propagates uncertainty through a model by repeatedly sampling uncertain inputs and calculating the resulting output. Instead of reporting only one expected profit, it estimates a distribution of profit, cost, service level, or net present value.

Suppose annual profit is modeled as:

If price, quantity, and costs are uncertain, each simulation iteration draws one value for each input and computes profit. After N iterations, the simulated outputs approximate the outcome distribution.

Simulation Workflow

1. Define the output and the decision question.
2. Identify uncertain inputs and fixed inputs.
3. Specify probability distributions using evidence or defensible expert judgment.
4. Represent dependencies among inputs.
5. Sample inputs and calculate outputs repeatedly.
6. Summarize mean, median, standard deviation, percentiles, and downside probabilities.
7. Validate the model and test sensitivity to distributional assumptions.
8. Communicate uncertainty and decision implications.

The sample mean of simulated outputs is:

where y_j is the output from iteration j. The sample standard deviation is:

The standard error of the simulated mean is approximately:

These formulas describe simulation sampling error, not uncertainty in the business model. A million iterations cannot correct a biased demand distribution or an omitted risk.

Distributions and Dependence

Common choices include normal, lognormal, triangular, uniform, binomial, Poisson, and empirical distributions. Selection should reflect the variable’s meaning and data. For example, a positive skewed revenue variable may be better represented by a lognormal distribution than a normal distribution, which permits negative values.

Correlation matters. Demand and price may be negatively related; disruption duration and expedited freight cost may be positively related. Independently sampling correlated variables can understate or overstate risk. Dependence may be modeled through historical data, a correlation structure, copulas, or explicit conditional relationships, but each approach requires validation.

Important outputs include:

• Expected value and median outcome.
• Percentiles, such as the 5th and 95th percentiles.
• Probability of loss: P(Profit < 0)
• Value at Risk (VaR), defined as a specified quantile of loss at a confidence level.
• Conditional Value at Risk (CVaR), the expected loss beyond the VaR threshold.

VaR must specify a horizon, confidence level, and loss convention. It does not describe the severity of losses beyond the threshold; CVaR addresses that limitation. Neither metric is a complete measure of business risk.

Simulation Quality and Validation

A simulation is credible only if its conceptual model, inputs, dependencies, and implementation are credible. Analysts should compare simulated ranges with historical observations, conduct extreme-value checks, document random-number seeds when reproducibility matters, and use convergence diagnostics. Results should be tested under alternative distributions and dependence assumptions.

A common implementation pattern is:

for j = 1 to N:
    draw uncertain inputs using the specified dependencies
    calculate the outcome y[j]
report distributional and decision metrics

Simulation should be used to support decisions, not to create false precision. Reporting “the probability is 17.38%” may be inappropriate when the input assumptions are highly uncertain; “approximately 15–20% under the stated assumptions” may be more honest.

7. Integrated Illustrative Case: Regional Distribution-Centre Expansion

The following is a fictional illustrative case.

Background and Problem

A consumer-goods distributor is considering a regional distribution centre. The investment requires $2.4 million immediately. Annual net benefits depend on demand and operating performance. Management may either build immediately, conduct a six-month pilot costing $180,000, or reject the project. The pilot provides imperfect evidence about demand and may allow management to defer construction.

The board is concerned not only with expected net present value (NPV), but also with cash-flow downside, reversibility, and the possibility that a regulatory change increases operating cost.

Analysis Approach

The analytics team first constructs a decision tree with the alternatives: build, pilot, or reject. Demand states are high, medium, and low. The pilot has different probabilities of a positive result conditional on each state, so Bayes’ rule is used to update the state probabilities. The model includes the pilot cost and the value of waiting.

Next, the team performs threshold analysis. It finds that immediate construction is attractive only if the probability of high demand exceeds 0.42, assuming the stated cost and discount-rate assumptions. The decision is therefore sensitive to demand assessment.

A Monte Carlo model then represents annual demand, fuel cost, construction cost, and regulatory expense as uncertain inputs. Demand and fuel cost are negatively related because high demand raises transportation activity and market tightness affects fuel contracts. The model produces the following hypothetical results:

The figures are illustrative outputs, not guaranteed forecasts. The pilot has a lower mean NPV because of its cost and delay, but it reduces downside exposure and creates information. Its value is therefore partly financial and partly strategic.

Outcome and Recommendation

The board adopts a staged pilot with explicit stop/go criteria. Construction proceeds only if the pilot generates sufficient demand evidence and if updated cost estimates remain within a defined limit. A monitoring dashboard tracks order density, service-level improvement, fuel cost, regulatory announcements, and customer retention. The decision is reviewed at the pilot endpoint rather than being treated as irreversible.

Learning Points

First, EMV alone would have favored immediate construction, but utility and downside analysis made the capital risk visible. Second, the value of information depends on whether information changes action. Third, correlation and tail outcomes can materially alter risk conclusions. Fourth, a staged decision can be robust even when it is not the highest-mean alternative. Finally, governance requires documenting assumptions, ownership of thresholds, and the authority to stop a project.

8. Professional, Ethical, and Governance Considerations

Risk analysis can influence employment, credit access, insurance, pricing, and public services. Analysts must therefore distinguish uncertainty from ignorance, disclose material assumptions, and avoid presenting model outputs as objective facts. A probability estimate based on limited historical data should not be described as a measured frequency without qualification.

Governance practices include version control, independent review, sensitivity testing, model documentation, access controls, and clear approval authority. Models used in lending, insurance, employment, or health-related decisions may require fairness assessment, explainability, privacy protection, and regulatory compliance. Sensitive variables and proxy variables should be examined for discriminatory effects.

AI-enabled forecasting and simulation can automate scenario generation, parameter estimation, and code production. However, generated models require human validation. AI may reproduce historical bias, obscure assumptions, or generate apparently plausible but unsupported distributions. An auditable workflow should retain source data, transformations, model versions, assumptions, random seeds where appropriate, and decision records.

9. Best Practices and Limitations

Effective decision analysis follows several principles:

• Define the decision before selecting a technique.
• Separate controllable decisions from uncertain states.
• Use consistent time horizons, units, and valuation rules.
• Make probability assumptions and data limitations explicit.
• Model dependencies rather than assuming independence automatically.
• Combine expected value with downside, utility, and feasibility constraints.
• Use sensitivity and scenario analysis to test robustness.
• Prefer staged or reversible decisions when uncertainty is high.
• Validate models against history and expert knowledge.
• Communicate ranges, thresholds, and actions—not only point estimates.

Limitations remain. Probabilities may be subjective, utility functions may be difficult to elicit, scenarios may omit disruptive events, and simulation outputs can create false confidence. Decision trees can become unwieldy when there are many stages. Correlations may change over time, and historical data may not represent structural breaks. Analytical sophistication cannot compensate for a poorly defined objective or weak governance.

Chapter Summary

Decision analysis makes uncertainty explicit by linking alternatives, states, probabilities, consequences, and preferences. Decision trees represent sequential choices and are evaluated by rollback analysis. Expected monetary value is a probability-weighted benchmark, while expected utility incorporates risk attitude. Perfect and sample information quantify the economic value of learning.

Sensitivity analysis identifies assumptions that influence recommendations, and threshold analysis converts uncertainty into operational trigger points. Scenario planning examines coherent alternative futures and supports strategic robustness. Monte Carlo simulation propagates uncertain inputs through a model and provides distributions, percentiles, and downside probabilities rather than a single forecast. Its quality depends on valid inputs, dependencies, implementation, and validation.

Professional decisions should integrate expected outcomes, risk exposure, utility, reversibility, information value, ethics, and governance. The best decision is not always the one with the highest average payoff; it is the one that fits the organization’s objectives, constraints, risk capacity, and ability to learn and adapt.

Key Terms

Decision under risk: A decision situation in which credible probabilities are assigned to uncertain states.

Decision under uncertainty: A situation in which probabilities are unavailable or not sufficiently justified.

Decision tree: A chronological representation of decisions, chance events, and payoffs.

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

Rollback analysis: Evaluating a decision tree from terminal outcomes backward to the initial decision.

Expected utility: The probability-weighted utility of possible outcomes.

Risk aversion: Preference for a less risky prospect when expected monetary value is comparable.

Certainty equivalent: A guaranteed outcome providing the same utility as a risky prospect.

Risk premium: The difference between expected monetary value and certainty equivalent.

Sensitivity analysis: Examination of how outputs change when inputs change.

Threshold analysis: Identification of the input value at which a decision or conclusion changes.

Scenario planning: Construction of coherent alternative descriptions of future conditions.

Robustness: The extent to which a decision remains satisfactory across plausible assumptions and futures.

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

Value at Risk: A specified quantile of loss over a defined horizon and confidence level.

Conditional Value at Risk: The expected loss beyond a specified Value-at-Risk threshold.

Expected value of perfect information: The maximum expected benefit of knowing the future state before acting.

Expected value of sample information: The expected benefit of imperfect information from a test or study.

Model risk: The risk of error arising from model structure, inputs, implementation, or use.

Review Questions

5 Short Answer Questions

1. What is the difference between decision-making under risk and decision-making under uncertainty?
2. Why can a project with a positive EMV still be rejected by a risk-averse manager?
3. Define a certainty equivalent.
4. What is the purpose of threshold analysis?
5. Why is dependence among simulation inputs important?

5 Long Answer Questions

1. Explain how a decision tree represents sequential decisions and uncertain events. Include the role of rollback analysis.
2. Compare EMV and expected utility as decision criteria, including their assumptions and limitations.
3. Describe a complete Monte Carlo simulation workflow for an uncertain investment decision.
4. Explain how scenario planning differs from probability forecasting and how both can support strategy.
5. Discuss the governance and ethical risks of using simulation or AI-generated forecasts in consequential decisions.

3 Analytical / Application Questions

1. A project returns $200,000 with probability 0.55 and loses $100,000 otherwise. Calculate its EMV and determine the probability threshold at which the project breaks even.
2. Design a decision tree for a firm choosing between immediate investment, a paid pilot, and no investment. Identify the information required to calculate the EVSI of the pilot.
3. Propose a Monte Carlo model for a retailer’s inventory decision. Specify at least four uncertain inputs, their plausible distributions, possible dependencies, and three risk outputs that management should monitor.

Multiple Choice Questions with Answers

Question 1

A project returns $120,000 with probability 0.6 and loses $60,000 with probability 0.4. What is its expected monetary value?

A. $24,000

B. $48,000

C. $60,000

D. $72,000

Correct Answer: B. $48,000

Question 2

In a decision tree, which node represents an event controlled by the decision maker?

A. Terminal node

B. Chance node

C. Decision node

D. Payoff node

Correct Answer: C. Decision node

Question 3

A risk-averse decision maker generally has which type of utility function over wealth?

A. Concave

B. Convex

C. Constant and decreasing

D. Discontinuous

Correct Answer: A. Concave

Question 4

For the launch payoff in the chapter, EMV is zero when the probability of high demand is approximately:

A. 0.167

B. 0.250

C. 0.333

D. 0.600

Correct Answer: C. 0.333

Question 5

Which statement about EVPI is correct?

A. It can never exceed the value of current information

B. It is the cost of the most accurate forecast

C. It is the expected value of choosing the best action after observing the state

D. It is always negative for risk-averse decision makers

Correct Answer: C. It is the expected value of choosing the best action after observing the state

Question 6

What is the main purpose of one-way sensitivity analysis?

A. To vary every input simultaneously

B. To identify the effect of changing one input while holding others constant

C. To eliminate all model uncertainty

D. To assign probabilities to scenarios

Correct Answer: B. To identify the effect of changing one input while holding others constant

Question 7

Which simulation output directly measures the likelihood that profit is negative?

A. Median profit

B. P(Profit < 0)

C. 95th percentile of profit

D. Sample mean of input price

Correct Answer: B. P(Profit < 0)

Question 8

Why can independently sampling correlated inputs produce misleading simulation results?

A. It always increases the number of iterations required

B. It may distort the joint distribution and understate or overstate risk

C. It prevents the calculation of a mean

D. It makes all outputs identical

Correct Answer: B. It may distort the joint distribution and understate or overstate risk

Question 9

Which statement best describes scenario planning?

A. It is necessarily a single most-likely forecast

B. It creates coherent alternative futures to test strategies

C. It ignores external drivers

D. It replaces all quantitative analysis

Correct Answer: B. It creates coherent alternative futures to test strategies

Question 10

A certainty equivalent is best defined as:

A. The highest possible payoff

B. The guaranteed amount with the same utility as a risky prospect

C. The probability of avoiding loss

D. The difference between VaR and CVaR

Correct Answer: B. The guaranteed amount with the same utility as a risky prospect