Simulation, Scenario Analysis and Risk Modeling

Introduction

Managerial decisions are often made under uncertainty. Demand may fluctuate, project durations may vary, exchange rates may move unpredictably, and operational failures may occur at uncertain times. A single forecast or “most likely” estimate can conceal the range of possible outcomes. Simulation provides a systematic way to represent uncertainty, reproduce the behavior of a system, and examine the consequences of alternative decisions.

Simulation is not a substitute for managerial judgment. It is a structured laboratory for decision-making. A well-designed model makes assumptions explicit, generates a distribution of possible outcomes, identifies the drivers of risk, and supports comparison of policies under normal and stressed conditions.

This chapter focuses on Monte Carlo simulation, discrete-event simulation, scenario analysis, dependence modeling, risk measures, validation, and managerial interpretation.

Learning Objectives

After studying this chapter, you should be able to:

1. Distinguish Monte Carlo simulation, discrete-event simulation, scenario analysis, and sensitivity analysis.
2. Specify uncertain inputs using appropriate probability distributions.
3. Model correlations, common drivers, and dependencies among risks.
4. Design and interpret Monte Carlo and discrete-event simulations.
5. Estimate expected values, percentiles, probabilities, and downside risk measures.
6. Conduct stress tests and scenario analysis for robust decisions.
7. Validate simulation models and communicate their limitations.
8. Apply simulation to a managerial case involving capacity and demand uncertainty.

1. The Logic of Simulation

A simulation model consists of:

• Inputs: controllable decisions and uncertain quantities.
• Relationships: equations, rules, constraints, and behavioral logic.
• State variables: information describing the system at a point in time.
• Random mechanisms: probability distributions or stochastic event rules.
• Outputs: measures used to evaluate performance or risk.

A deterministic model produces one outcome for a specified set of inputs. A stochastic simulation produces many outcomes because uncertain inputs or events vary across replications.

If output Y is a function of uncertain inputs X₁, X₂, …, X_k and decision variables d, the model may be represented as:

The function f may be an algebraic equation, a sequence of operational rules, or a computer program. Simulation estimates properties of the distribution of Y, such as its mean, variance, percentiles, or probability of exceeding a threshold.

Simulation versus Related Methods

These methods are complementary. Scenario analysis may define strategic environments, while Monte Carlo simulation analyzes uncertainty within each environment.

2. Monte Carlo Simulation

Monte Carlo simulation repeatedly samples uncertain inputs and evaluates the model. Suppose a project’s net present value is:

where CF_t is cash flow at time t, r is the discount rate per period, and T is the final period. If future cash flows, costs, and the discount rate are uncertain, each replication samples values for these variables and calculates one NPV.

After N replications, the simulated outputs are Y₁, Y₂, …, Y_N. The estimated mean is:

The sample variance is:

The standard error of the estimated mean is approximately:

These formulas assume independent replications with finite variance. The standard error describes simulation sampling error, not uncertainty in the real-world assumptions.

Probability Distributions for Inputs

A distribution should reflect the nature and quality of available information. Common choices include:

For a uniform random variable X ~ U(a,b), the expected value and variance are:

The uniform model assumes every value between a and b is equally likely. That assumption should not be used merely because it is convenient.

For a triangular distribution with minimum a, mode m, and maximum b, the mean is:

The mean formula assumes a ≤ m ≤ b. A triangular distribution is often useful for early-stage estimates, but it embeds subjective judgments about the mode and bounds.

Random Sampling and Inverse Transform Logic

If U is uniformly distributed on (0,1), a random variable can often be generated through its inverse cumulative distribution function:

where F is the cumulative distribution function of X. The method requires a valid distribution and a reliable random-number generator. Modern software generally handles generation, but analysts remain responsible for choosing and documenting the distribution.

Worked Example: Inventory Profit

A retailer sells a product for £50 per unit. Unit purchase cost is £30, and unsold units have a salvage value of £10. Weekly demand is modeled as discrete values of 80, 100, and 120 units with probabilities 0.25, 0.50, and 0.25. The retailer orders 100 units.

For demand D, sales equal min(D,100), and leftover inventory equals max(100−D,0). Profit is:

The three possible profits are:

Therefore:

The probability of profit below £1,500 is 0.25. A Monte Carlo model would sample demand repeatedly and produce the same probabilities approximately, subject to simulation error.

3. Dependence and Correlation

A major modeling error occurs when related risks are sampled independently. For example, high demand may coincide with higher input prices, or adverse weather may simultaneously reduce supply and increase delivery time.

Correlation measures linear association. For variables X and Y, correlation is:

where Cov(X,Y) is covariance and σ_X, σ_Y are standard deviations. Correlation lies between −1 and +1. It does not prove causation and does not fully describe nonlinear dependence.

Possible approaches include:

1. Common-driver modeling: represent a shared factor, such as an economic index, that influences several inputs.
2. Conditional distributions: specify one variable’s distribution conditional on another.
3. Correlation matrices: impose a target dependence structure, often through a multivariate normal or Gaussian copula.
4. Empirical resampling: sample historical observations as linked vectors rather than sampling each column independently.

A correlation matrix must be mathematically admissible, usually positive semidefinite. Arbitrarily entering pairwise correlations can create an invalid matrix and unreliable simulations.

Dependence may also change during extreme events. Linear correlation estimated from normal periods can understate joint downside risk. Copulas, regime-switching models, and stress scenarios can help, but they require stronger data and modeling judgment.

4. Discrete-Event Simulation

Discrete-event simulation (DES) models a system as a sequence of events that occur at particular times. An event changes the state of the system. Examples include customer arrival, service completion, machine failure, replenishment, or shipment departure.

A basic DES contains:

• a simulation clock;
• a future-event list;
• entities, such as customers or jobs;
• resources, such as servers or machines;
• queues and priority rules;
• event-scheduling logic;
• performance measures.

The event-scheduling cycle is:

1. Select the earliest event.
2. Advance the clock to that event time.
3. Update system state.
4. Record relevant statistics.
5. Schedule new events generated by the state change.
6. Repeat until the stopping condition is reached.

Figure: The event-scheduling cycle in a discrete-event simulation.

DES differs from time-step simulation. In a time-step model, the clock advances by fixed increments; in DES, it jumps from one event time to the next. DES is efficient when important changes occur irregularly.

For a single-server queue, utilization is often summarized by:

where λ is the average arrival rate and μ is the average service rate, both measured in entities per unit time. A stable long-run queue generally requires ρ < 1. This condition alone does not determine waiting time; variability, queue discipline, and the number of servers also matter.

5. Output Analysis and Risk Measures

Simulation output should be treated as a distribution, not merely an average. Important measures include:

• Mean: average outcome.
• Median: 50th percentile.
• Percentile: value below which a specified proportion of outcomes fall.
• Probability of loss: P(Y < 0) for a profit or NPV measure.
• Value at Risk (VaR): a percentile-based loss threshold.
• Expected Shortfall (ES): average loss conditional on being beyond the VaR threshold.

If L denotes loss and α is a confidence level, VaR is commonly defined as:

where q_α(L) is the α-quantile of loss. Expected shortfall is:

for continuous losses or with an appropriate tail convention for discrete losses. VaR does not describe the average severity beyond the threshold; ES does.

For a finite simulation sample, the 95th percentile is estimated by sorting outcomes and selecting an appropriate order statistic. Different software uses different interpolation conventions, so the method should be documented.

Confidence Intervals and Replications

A 95% approximate confidence interval for the simulated mean is:

This approximation assumes sufficiently independent replications and a reasonably stable sampling distribution of the mean. It does not mean there is a 95% probability that the fixed model mean lies inside the computed interval; rather, the procedure has approximately 95% coverage under repeated sampling.

For time-dependent DES, observations within one run may be correlated. Analysts should use independent replications, batch means, or other appropriate time-series methods. Initial conditions can bias results, especially when a system begins empty. A warm-up period may be discarded after examining convergence, and the choice must be justified.

6. Scenario Analysis and Stress Testing

A scenario is a coherent set of assumptions about external conditions and management responses. A scenario is not simply a pessimistic number. It should describe a plausible configuration of drivers, such as:

• demand contraction;
• supplier disruption;
• inflation and interest-rate changes;
• regulatory intervention;
• cyber incident;
• competitor entry;
• accelerated technology adoption.

A useful scenario framework combines:

1. Baseline scenario: continuation of expected conditions.
2. Alternative scenarios: materially different but plausible futures.
3. Stress scenarios: severe conditions designed to test resilience.
4. Reverse stress test: identify conditions that would cause a critical failure.

Scenario analysis is especially valuable when historical data are limited or structural change makes past frequencies unreliable. Monte Carlo simulation is stronger for probabilistic aggregation; scenarios are stronger for strategic narratives and discontinuities.

Sensitivity analysis examines how output changes as one input changes. One-way sensitivity changes one input while holding others constant. Two-way sensitivity examines interactions between two inputs. Tornado charts, elasticity measures, and variance-based methods can rank drivers.

A robust decision should not merely maximize expected value. It may consider downside probability, resilience, flexibility, service commitments, and the cost of acquiring additional information.

7. Model Construction and Validation

A defensible simulation follows a disciplined workflow:

7.1 Define the decision

State the decision, alternatives, objective, time horizon, constraints, and performance measures. Avoid building a sophisticated model before clarifying the decision it must support.

7.2 Build a conceptual model

Map causal relationships, events, states, resources, feedback loops, and boundaries. Consult subject-matter experts and document exclusions.

7.3 Collect and assess data

Check definitions, units, missing values, outliers, structural breaks, selection bias, and representativeness. Historical data describe the past, not automatically the future.

7.4 Select distributions

Use empirical evidence where possible. Fit candidate distributions using graphical checks and goodness-of-fit diagnostics, but do not let statistical fit override operational plausibility.

7.5 Implement and test

Use modular code, fixed random seeds for debugging, version control, input validation, and reproducible configuration files. Unit-test formulas and extreme cases.

7.6 Verify and validate

Verification asks whether the model was implemented correctly. Validation asks whether the model is an adequate representation for its intended purpose. Techniques include:

• dimensional and logic checks;
• extreme-value tests;
• comparison with historical performance;
• face validation by experts;
• sensitivity and perturbation tests;
• comparison with simpler benchmark models;
• out-of-sample testing where data permit.

A model can be computationally correct but decisionally invalid if its assumptions, boundary, or outputs are inappropriate.

7.7 Analyze and communicate

Report assumptions, input sources, dependence structures, number of replications, random-number controls, convergence evidence, output distributions, and limitations. Present ranges and drivers rather than false precision.

8. Illustrative Case: Capacity Planning for a Regional Fulfilment Centre

This is a fictional illustrative case.

Background

A regional fulfilment centre must decide whether to add a second packing line before the holiday season. The investment costs £480,000. Management expects seasonal demand to be uncertain and is concerned about late deliveries.

Problem

The centre currently has one packing line. Orders arrive according to a Poisson process with average rate 42 orders per hour. One line can process orders at an average rate of 48 orders per hour, but service times vary. During peak periods, demand may increase by 20%, while a labour shortage may reduce service capacity by 15%.

Management wants to compare one-line and two-line policies using expected operating cost, probability that average waiting time exceeds 20 minutes, and the probability of failing the delivery-service target.

Analysis and Approach

The model represents arrivals, service completion, queue discipline, labour availability, and an eight-hour operating day. Arrival and service processes are sampled independently in the baseline, but peak demand and labour shortage are modeled as scenario-level common drivers. For each replication, the model records queue length, waiting time, throughput, overtime, and late orders.

The baseline scenarios are:

The scenario probabilities are planning assumptions, not objective frequencies. They must be reviewed with operations and finance teams.

For one line under normal conditions, approximate utilization is:

This is close to full utilization, so variability can generate substantial waiting. Under peak demand, utilization is:

A sustained system with utilization above one cannot clear work in the long run without additional capacity, overtime, demand controls, or a change in operating hours.

The two-line alternative has total nominal service capacity of 96 orders per hour. Even under peak demand, approximate aggregate utilization is:

The simulation should not assume that two lines are perfectly substitutable; shared staff, equipment, and downstream bottlenecks must be represented where material.

Illustrative Results

Suppose 10,000 independent replications are run for each policy and scenario. The following results are illustrative model outputs:

The two-line policy costs £1,200 more per operating day but materially reduces service risk. The investment decision cannot be made from expected operating cost alone. If the avoided cost of late deliveries, customer compensation, reputational damage, and lost future demand exceeds the incremental cost, the second line may be economically justified.

A stress test with simultaneous peak demand and labour shortage should be run even if its assigned probability is small. It may reveal that the second line does not solve the problem if common labour constraints affect both lines. Management might then consider cross-training, temporary labour contracts, order cut-off changes, or flexible outsourcing.

Outcome and Learning Points

The simulation supports a conditional recommendation: add capacity if service penalties and strategic customer effects justify the cost, but combine the investment with labour-contingency planning. The principal learning points are:

• high utilization magnifies variability;
• scenario-level common drivers matter;
• a capacity decision should be evaluated against service risk, not only cost;
• a stress test can expose vulnerabilities hidden by probability-weighted averages;
• simulation output informs judgment but does not determine the decision automatically.

9. Technology, AI and Governance

Simulation software ranges from spreadsheets to specialized DES platforms, statistical programming environments, cloud systems, and digital twins. AI can assist with demand forecasting, anomaly detection, distribution fitting, surrogate models, and automated experimentation. However, AI-generated assumptions require human review. A model that reproduces historical patterns may fail under unprecedented conditions or encode historical bias.

Governance should address:

• ownership of assumptions and model approval;
• data lineage and access controls;
• versioning and reproducibility;
• independent review for material decisions;
• scenario coverage and model-risk limits;
• documentation of overrides and judgment;
• privacy, confidentiality, and cybersecurity;
• explainability for affected stakeholders.

In regulated sectors, simulation outputs may support capital, liquidity, safety, or operational decisions. The organization must distinguish a model estimate from a guaranteed forecast and retain evidence sufficient for audit and challenge.

10. Challenges and Limitations

Simulation has important limitations:

1. Garbage in, garbage out: inappropriate distributions produce misleading output.
2. Model risk: omitted feedback, capacity constraints, or behavioral responses can dominate numerical precision.
3. Rare-event estimation: very small probabilities require many replications or specialized variance-reduction methods.
4. Dependence uncertainty: correlations and tail dependence are often poorly known.
5. Non-stationarity: relationships may change because of technology, regulation, or competition.
6. Optimization bias: selecting the best policy from many noisy simulations can overstate performance.
7. False precision: detailed decimal outputs may conceal subjective assumptions.
8. Communication risk: stakeholders may confuse a simulated percentile with a forecast certainty.

Variance-reduction methods such as common random numbers, antithetic variates, control variates, stratified sampling, and importance sampling can improve efficiency. They must be implemented correctly and documented because they alter sampling behavior without correcting a flawed conceptual model.

Best-Practice Checklist

Before relying on a simulation, ask:

• Is the decision and objective clearly stated?
• Are all inputs defined with units and time bases?
• Are distributions justified by data or expert evidence?
• Are dependencies and common drivers modeled?
• Has the model been verified and validated?
• Are warm-up, termination, and replication rules appropriate?
• Are means, percentiles, tail risks, and sensitivities reported?
• Have severe but plausible scenarios been tested?
• Is simulation error distinguished from real-world uncertainty?
• Can another analyst reproduce the results?
• Are ethical, regulatory, privacy, and governance obligations addressed?

Chapter Summary

Simulation converts uncertain assumptions and system rules into a distribution of possible outcomes. Monte Carlo simulation is appropriate when repeated probabilistic sampling can represent uncertainty, whereas discrete-event simulation is appropriate for systems whose state changes at identifiable events. Scenario analysis provides coherent alternative futures, and stress testing examines resilience under severe conditions.

Effective risk modeling requires defensible distributions, explicit dependencies, appropriate output measures, and careful treatment of simulation error. Means alone are insufficient; percentiles, probabilities of failure, Value at Risk, Expected Shortfall, and sensitivity measures often provide more decision-relevant information. Validation, documentation, reproducibility, and governance are as important as computational sophistication. The managerial value of simulation lies not in producing a seemingly precise number, but in clarifying assumptions, revealing trade-offs, and supporting robust action under uncertainty.

Key Terms

Simulation: A representation of a real system used to examine outcomes under specified assumptions.

Monte Carlo simulation: Repeated random sampling used to estimate the distribution of model outputs.

Discrete-event simulation: Simulation in which system state changes at discrete event times.

Scenario analysis: Comparison of outcomes under coherent alternative assumptions about the future.

Stress test: Evaluation of performance under severe but plausible conditions.

Random variable: A variable whose value is determined by a probability mechanism.

Probability distribution: A mathematical description of possible values and their probabilities.

Cumulative distribution function: The probability that a random variable is less than or equal to a specified value.

Correlation: Standardized linear association between two variables.

Dependence: Any statistical relationship between variables, including nonlinear relationships.

Replication: One complete run of a stochastic simulation.

Simulation error: Difference between a simulated estimate and the corresponding model quantity caused by finite sampling.

Warm-up period: Initial simulation interval excluded to reduce initialization bias.

Utilization: Demand or workload rate divided by service or capacity rate.

Percentile: A value below which a specified percentage of observations fall.

Value at Risk: A specified quantile of a loss distribution.

Expected Shortfall: Average loss conditional on exceeding a specified tail threshold.

Verification: Checking whether the implemented model operates according to its design.

Validation: Assessing whether the model is adequate for its intended real-world purpose.

Model risk: Risk of incorrect decisions caused by flawed model structure, data, assumptions, or implementation.

Review Questions

Short Answer Questions

1. What is the principal difference between deterministic and stochastic simulation?
2. Why can independent sampling of correlated risks understate downside risk?
3. What does utilization measure in a queueing system?
4. Distinguish verification from validation.
5. Why should a simulation report percentiles in addition to the mean?

Long Answer Questions

1. Explain the steps involved in designing and validating a Monte Carlo simulation for a capital-investment decision.
2. Compare Monte Carlo simulation, discrete-event simulation, scenario analysis, and sensitivity analysis.
3. Discuss the strengths and limitations of triangular, normal, lognormal, and empirical input distributions.
4. Explain how dependence and common drivers should be modeled in enterprise risk analysis.
5. Discuss why stress testing is valuable even when a severe scenario has a low assigned probability.

Analytical / Application Questions

1. A service centre receives 36 requests per hour and processes 40 requests per hour. Calculate approximate utilization and explain what additional information is required before concluding that waiting times will be acceptable.
2. Design a simulation experiment to decide whether an organization should outsource a variable-volume logistics operation. Specify inputs, dependencies, outputs, scenarios, and validation tests.
3. A project simulation has a mean NPV of £2.4 million, a 10th percentile of −£0.8 million, and a probability of negative NPV of 0.18. Interpret these results for a board and identify at least three further analyses required before approval.

Multiple Choice Questions with Answers

Question 1

Which method is most appropriate for representing a queue in which arrivals, service completions, and machine failures occur at specific times?

A. Discrete-event simulation

B. Simple point forecasting

C. One-way sensitivity analysis

D. Historical averaging

Correct Answer: A. Discrete-event simulation

Question 2

For a uniform random variable X ~ U(10, 20), what is E[X]?

A. 5

B. 10

C. 15

D. 20

Correct Answer: C. 15

Question 3

If arrival rate is 42 orders per hour and service rate is 48 orders per hour, what is utilization?

A. 0.125

B. 0.875

C. 1.143

D. 90

Correct Answer: B. 0.875

Question 4

What is the principal danger of sampling two strongly related risk inputs independently?

A. It always increases the simulated mean

B. It can understate joint extreme outcomes

C. It eliminates simulation sampling error

D. It guarantees a non-positive correlation

Correct Answer: B. It can understate joint extreme outcomes

Question 5

Which measure describes the average loss conditional on loss exceeding a specified tail threshold?

A. Median

B. Variance

C. Expected Shortfall

D. Mean absolute deviation

Correct Answer: C. Expected Shortfall

Question 6

A 95% simulation confidence interval for an estimated mean primarily reflects which type of uncertainty?

A. Sampling error from a finite number of replications

B. All uncertainty about future structural change

C. The probability that the model is conceptually valid

D. The certainty of the input distribution

Correct Answer: A. Sampling error from a finite number of replications

Question 7

Which statement best distinguishes verification from validation?

A. Verification checks implementation; validation checks adequacy for purpose

B. Verification concerns finance; validation concerns operations

C. Verification uses scenarios; validation uses only averages

D. Verification is optional; validation is automatic

Correct Answer: A. Verification checks implementation; validation checks adequacy for purpose

Question 8

If a queue has an arrival rate of 50.4 orders per hour and service rate of 48 orders per hour, what is the approximate utilization?

A. 0.48

B. 0.95

C. 1.05

D. 2.40

Correct Answer: C. 1.05

Question 9

Which is the strongest reason to use a scenario-level common driver in a simulation?

A. To force every input to have the same distribution

B. To represent several outcomes responding to a shared condition

C. To remove the need for validation

D. To guarantee normally distributed outputs

Correct Answer: B. To represent several outcomes responding to a shared condition

Question 10

What is the main purpose of a reverse stress test?

A. To estimate the median outcome only

B. To identify conditions that would cause a critical failure

C. To replace all baseline forecasts

D. To ensure that every scenario is equally probable

Correct Answer: B. To identify conditions that would cause a critical failure