Risk, Uncertainty and Scenario Modelling

Introduction and context

Financial analysis is often presented as if future cash flows, prices, default rates and operating outcomes were known. In practice, forecasts are conditional beliefs. A model may be internally consistent and still be wrong because its assumptions are wrong, incomplete or exposed to unexpected events. Risk modelling therefore complements valuation and planning by asking not only “What is the expected result?” but also “How sensitive is it?”, “What could go wrong?”, “How likely is it?”, and “What action would be appropriate?”

Risk is commonly defined as uncertainty with measurable or estimable probabilities. Uncertainty, in a narrower sense, refers to situations in which probabilities are unknown, unreliable or disputed. This distinction is useful but not absolute. A probability estimate is itself a model judgement and can be affected by limited data, structural change and behavioural bias.

For a bank, insurer or corporate finance department, risk analysis supports credit decisions, capital planning, pricing, liquidity management, investment appraisal, budgeting and regulatory reporting. Its purpose is not to eliminate uncertainty. Rather, it is to make assumptions visible, quantify material exposures, identify vulnerabilities and improve decisions.

Learning objectives

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

• distinguish risk, uncertainty, variability and model risk;
• perform one-way and two-way sensitivity analysis and calculate break-even values;
• construct and interpret scenario models and stress tests;
• analyse decision trees using expected monetary value and expected utility;
• select probability distributions and model correlations responsibly;
• estimate and interpret value at risk, expected shortfall and downside measures;
• explain Monte Carlo simulation and assess simulation quality;
• communicate ranges and probabilities without implying false precision; and
• recognise governance, ethical and regulatory implications of risk models.

1. A framework for thinking about uncertainty

A financial model contains three distinct layers. The first is the structure: equations linking revenues, costs, cash flows, balance-sheet items and financing. The second is the inputs: rates, volumes, prices, probabilities and timing assumptions. The third is the decision rule: for example, accept an investment if its net present value is positive, or hold capital above a risk threshold.

Risk analysis can fail at any layer. A spreadsheet may contain an incorrect debt-pricing formula (structural risk), an unrealistic default rate (input risk), or an inappropriate decision threshold (decision risk). Model risk is the possibility that a model is unsuitable, incorrectly implemented or misinterpreted. It includes data, coding, estimation, specification and use errors.

A useful decomposition is:

• Market risk: losses from movements in interest rates, exchange rates, equity prices, commodity prices or credit spreads.
• Credit risk: loss caused by a counterparty’s failure to meet obligations.
• Liquidity risk: inability to meet obligations when due without unacceptable loss.
• Operational risk: loss from failed processes, people, systems or external events.
• Strategic and business risk: adverse effects of decisions, competition or changes in the operating environment.
• Model risk: loss arising from a deficient model or its inappropriate use.

Expected outcomes do not fully describe risk. Two investments may have the same expected return but very different downside exposure, liquidity and tail loss. A responsible analysis therefore considers central estimates, dispersion, adverse scenarios, dependencies and the reversibility of decisions.

2. Sensitivity analysis and break-even analysis

One-way sensitivity analysis

One-way sensitivity analysis changes one input while holding other inputs constant. If project value is V and input x changes, a simple local sensitivity is:

Here ∆V and ∆x are changes in value and the input. S_x is dimensionless and approximates the percentage change in value caused by a 1% change in the input. The measure is most informative for small changes and a reasonably smooth model. Holding all other variables constant is a useful diagnostic, not a prediction of what will happen in reality.

For an investment with cash flow CF_t in period t, discount rate r, and life n, net present value is:

I₀ is the initial investment. The formula assumes cash flows are measured consistently, occur at specified dates, and r is appropriate for their risk and timing. If cash flows are periodic, the rate must match the period; a nominal annual rate should not be applied directly to monthly cash flows without conversion.

A sensitivity table may vary sales volume, selling price, variable cost, fixed cost, discount rate and project life. A tornado chart ranks variables by their effect on NPV, but ranking does not establish probability. A highly sensitive variable may be very predictable, while a less sensitive variable may be extremely uncertain.

Two-way and multidimensional sensitivity

Two-way analysis changes two inputs simultaneously, such as price and volume or default rate and recovery rate. It helps reveal interactions. For example, a bank’s credit loss depends jointly on exposure at default, probability of default and loss given default:

Expected loss EL is in currency units; EAD is exposure at default, PD is probability of default, and LGD is loss given default. The expression is a simplified expected-loss model. It assumes the three quantities are appropriately defined for the same exposure and period. In a portfolio, dependence, default timing, provisions, recoveries and discounting may require a more complete model.

Break-even analysis

Break-even analysis identifies the operating level at which profit is zero. With fixed costs F, price per unit P, and variable cost per unit V, the accounting break-even volume is:

This requires P > V and assumes fixed cost and unit contribution are constant over the relevant range. If target operating profit is Π, required volume is:

For capital investment, an NPV break-even may instead solve for the price, volume, discount rate or terminal value that makes NPV = 0. Such a break-even is not necessarily a forecast; it is a threshold for decision discussion.

Worked example. A service has price 80, variable cost 50 and annual fixed costs 240,000. Contribution per unit is 30, so:

At 10,000 units, operating profit is (10,000 × 30) − 240,000 = 60,000. If price falls to 74 while other assumptions remain unchanged, contribution becomes 24 and break-even rises to 10,000 units. The analysis exposes operating leverage: a small price change can create a large change in profit when fixed costs are substantial.

3. Scenarios and stress testing

A scenario is a coherent set of assumptions describing a possible environment. A baseline scenario might combine moderate growth, stable funding costs and normal credit losses. An adverse scenario should change related variables consistently—for example, weaker economic activity, higher unemployment, lower collateral values, wider credit spreads and higher funding costs.

Scenarios should not be confused with forecasts. A baseline is often a planning assumption, while an adverse scenario is a conditional “if” statement. Probability-weighted scenarios can estimate an expected value:

where X_i is the outcome in scenario i, p_i is its probability, p_i ≥ 0, and ∑p_i = 1. This calculation is meaningful only when the scenarios are sufficiently exhaustive, probabilities are defensible and outcomes are measured consistently.

Stress testing deliberately examines severe but plausible conditions. It may be historical, hypothetical or reverse-engineered. A historical stress repeats a past episode; a hypothetical stress constructs a new combination; a reverse stress asks what circumstances would cause failure of a specified limit or capital position. Stress tests are valuable precisely because they explore the tails, but they should not be assigned precise probabilities merely because a numerical result has been produced.

A good stress test specifies:

1. the risk purpose and decision threshold;
2. the scenario narrative and transmission channels;
3. the variables shocked and their time horizon;
4. management actions, such as repricing or asset sales;
5. balance-sheet and behavioural assumptions;
6. effects on profit, liquidity, capital and covenants; and
7. escalation, remediation and monitoring arrangements.

Avoid double counting. If a recession scenario already raises unemployment and defaults, an additional unexplained default shock may exaggerate losses. Conversely, assuming management can immediately eliminate all exposure may understate risk.

4. Decision trees and decision quality

A decision tree represents sequential choices and uncertain events. Decision nodes are controlled by the decision-maker; chance nodes have alternative outcomes with probabilities; terminal nodes contain payoffs. The tree is solved from right to left by rollback.

For mutually exclusive outcomes with probabilities p_i and monetary outcomes X_i, expected monetary value is:

Choose the action with the highest EMV when outcomes are comparable, probabilities are credible and the decision-maker is risk neutral. Risk-averse decision-makers may instead maximise expected utility:

where u(·) is a utility function. A concave utility function represents diminishing marginal utility of wealth and therefore aversion to risk. EMV should not be used mechanically where liquidity constraints, risk limits, capital requirements or non-financial consequences matter.

The expected value of perfect information is the expected value with perfect information minus the best EMV without it. It is an upper bound on what information is worth; no organisation should pay more than this amount for perfect information, and real information is imperfect.

5. Probability distributions and correlations

A probability distribution describes possible values and their likelihoods. A discrete distribution assigns probabilities to countable outcomes. A continuous distribution uses a density, and the probability of an exact single value is zero; probabilities are assigned to intervals.

Common choices include:

For a random variable X, expected value and variance are:

where f(x) is the density, μ = E(X), and σ is standard deviation. For a discrete variable, the integral is replaced by ∑ x_i p_i. Standard deviation has the same units as the variable; variance has squared units.

Correlation between X and Y is:

It lies between −1 and +1 and measures linear association. Correlation is not causation and zero correlation does not imply independence except in special distributional settings, notably jointly normal variables. Dependence can strengthen in crises, so historical correlation may be unsuitable for stress analysis.

When simulating correlated normal risk factors, one common method is to generate independent standard normal vector Z and use:

where L is a matrix satisfying LL^T = Σ, and Σ is the target covariance matrix. Cholesky decomposition requires a valid positive-definite covariance matrix. This technique creates linear-Gaussian dependence; it does not automatically reproduce asymmetric or tail dependence.

6. Monte Carlo simulation

Monte Carlo simulation repeatedly samples uncertain inputs, calculates the model outcome and records the distribution of results. A basic procedure is:

1. define the output and model equations;
2. identify uncertain inputs and assign distributions;
3. specify dependencies and correlations;
4. generate random draws;
5. calculate outputs for each trial;
6. summarise means, percentiles, losses and sensitivities; and
7. validate, document and communicate the results.

If N independent trials produce outputs Y₁, …, Y_N, the simulated mean is:

The standard error of the mean is approximately s_Y ÷ √N, where s_Y is the sample standard deviation. Increasing trials reduces simulation noise at a rate proportional to 1/√N; it does not correct a wrong model or bad assumptions.

Simulation should use reproducible random seeds for testing, sensitivity checks for distributional assumptions and convergence checks for important percentiles. Results should be reported with suitable precision. A 95th percentile estimated from a small sample should not be printed as an exact monetary truth.

7. Value at risk and downside measures

For a loss variable L, value at risk at confidence level α is a quantile:

Thus, 99% one-day VaR of 10 million means that, under the model and horizon, losses exceed 10 million on approximately 1% of days. It does not mean the maximum possible loss is 10 million, nor does it describe the average loss beyond the threshold.

Expected shortfall at level α is the conditional tail expectation:

For continuous distributions this expression is direct; with discrete losses, a careful tail-mean definition is required because the quantile may have positive probability mass. Expected shortfall captures severity beyond VaR and is generally more informative for tail risk, although it remains model-dependent.

Other downside measures include maximum drawdown, lower partial moments and the Sortino ratio. If returns are R_t and a target return is T, a downside semideviation can be estimated from negative deviations:

The denominator convention may vary; analysts must state it. No single risk measure is sufficient. VaR and expected shortfall should be accompanied by stress results, liquidity analysis and exposure information.

8. Illustrative case: regional bank lending expansion

Background. This fictional case concerns Meridian Regional Bank, which is considering a 100 million expansion of small-business lending. The proposed portfolio has an estimated one-year exposure of 100 million, baseline probability of default of 3%, and loss given default of 45%.

Problem. Management must decide whether the expected return compensates for credit loss, capital usage and adverse economic conditions. It is concerned that average assumptions hide concentration in one industry.

Analysis. Baseline expected credit loss is:

An adverse scenario raises probability of default to 8%, loss given default to 55% and reduces expected fee income by 2 million. Adverse expected credit loss is 2.2 million. Relative to baseline, the credit-loss increase is 850,000, and the additional fee-income loss produces a total deterioration of 2.85 million before tax and capital effects.

A two-way analysis varies default probability from 2% to 10% and loss given default from 35% to 65%. It shows that the product of the two drivers, rather than either driver alone, determines expected loss. A concentration stress then assumes that 40% of borrowers in the exposed industry default together. This is not equivalent to independently applying a portfolio-average PD; it represents dependence in a downturn.

The bank simulates correlated defaults, collateral values and funding spreads. Outputs include net income, capital ratio, liquidity buffer and expected shortfall of annual loss. It also performs a reverse stress: the level of default and recovery deterioration that would breach the board’s capital limit.

Approach and solution. The credit committee approves only a smaller initial allocation, imposes sector limits, requires stronger covenants and prices the facility using an adverse—not merely average—loss estimate. The model is independently validated, assumptions are documented, and quarterly monitoring compares observed arrears and recoveries with forecasts.

Outcome. The decision is not “approved because expected loss is low” or “rejected because the stress is severe.” It is a conditional approval with risk controls, escalation triggers and a review point. This illustrates how scenario analysis converts a numerical forecast into a governance process.

Learning points. Expected loss is not a tail-loss measure; dependence matters; management actions must be realistic; and model outputs should be linked to limits, pricing and monitoring.

9. Governance, ethics and communication

Risk models influence credit access, insurance premiums, investment allocation and capital requirements. Governance should include clear ownership, independent validation, version control, change logs, data-quality controls, outcome monitoring and periodic back-testing. Material models should have documented limitations and an approved use case.

Ethically, analysts should avoid selective scenario presentation, hidden assumptions and unjustified claims of objectivity. Automated credit or insurance models may reproduce historical discrimination through biased data or proxy variables. Fairness testing, explainability, human review and lawful data use are therefore important. Privacy rules, consumer-protection obligations, accounting standards and prudential requirements may apply depending on the institution and jurisdiction.

Technology enables spreadsheet add-ins, cloud simulation, machine learning and real-time dashboards. These tools improve speed but can increase opacity, cyber exposure and false confidence. Artificial intelligence should support—not replace—expert challenge, especially where data are sparse or structural breaks are likely. A complex model is not automatically a better model.

Communicate uncertainty using ranges, scenario labels, confidence or credibility qualifications, time horizons and assumptions. Distinguish “the model estimates” from “the organisation knows.” Use rounded values when estimation uncertainty is material, show adverse outcomes alongside central estimates, and explain what would cause the result to change. Precision in a displayed number should not exceed precision in the evidence.

Best-practice checklist

• Define the decision, horizon, output and material risks before modelling.
• Separate known data, assumptions, expert judgements and model-generated values.
• Use sensitivity analysis to identify drivers and scenarios to examine coherent combinations.
• Validate distributions, correlations, tails and units.
• Test implementation independently and retain reproducible simulation settings.
• Include stress and reverse-stress analysis where failure is consequential.
• Compare model predictions with realised outcomes and investigate exceptions.
• Report limitations, not only attractive central estimates.
• Link results to decisions, limits, capital, pricing and contingency actions.

Chapter Summary

Risk analysis extends financial modelling beyond a single forecast. Sensitivity analysis identifies influential inputs, while break-even analysis identifies thresholds at which profit or value changes sign. Scenarios combine assumptions into coherent narratives, and stress tests examine severe but plausible conditions. Decision trees support sequential choices through expected monetary value or expected utility, but probabilities and risk preferences must be explicit.

Probability distributions describe uncertain inputs; correlation describes dependence but not causation or complete tail behaviour. Monte Carlo simulation produces a distribution of outcomes and is useful only when its structure, inputs and dependencies are credible. VaR describes a loss quantile, whereas expected shortfall describes the average severity beyond a quantile. Neither replaces stress testing, liquidity analysis or judgement.

Professional risk modelling requires validation, governance, ethical awareness and careful communication. The objective is not false precision but better-informed, conditional decisions.

Key Terms

Risk: Uncertainty with outcomes whose probabilities can be estimated.

Uncertainty: Limited or disputed knowledge about future outcomes or probabilities.

Model risk: Loss arising from an unsuitable, incorrectly implemented or misused model.

Sensitivity analysis: Study of output changes caused by changes in inputs.

Break-even point: Level at which revenue or value equals relevant costs, producing zero profit or NPV.

Scenario: Coherent set of assumptions describing a possible future environment.

Stress test: Analysis of severe but plausible adverse conditions.

Reverse stress test: Analysis that identifies conditions capable of breaching a limit or causing failure.

Decision tree: Diagrammatic representation of sequential decisions and uncertain events.

Expected monetary value: Probability-weighted average of monetary outcomes.

Expected utility: Probability-weighted average of utility values.

Probability distribution: Description of possible values and their likelihoods.

Correlation: Standardised measure of linear association between two variables.

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

Value at risk: Loss quantile at a stated confidence level and horizon.

Expected shortfall: Average loss in the tail beyond a specified VaR threshold.

Downside risk: Risk associated with outcomes below a target or reference level.

Back-testing: Comparison of model predictions with realised outcomes.

False precision: Presentation of estimates with more numerical detail than the evidence supports.

Review Questions

Short Answer Questions

1. Distinguish risk from uncertainty and explain why the distinction can be difficult in practice.
2. What is the purpose of one-way sensitivity analysis?
3. State two differences between a scenario and a stress test.
4. What does a 99% one-day VaR measure?
5. Why does increasing the number of Monte Carlo trials not repair a misspecified model?

Long Answer Questions

1. Explain how sensitivity analysis, scenario analysis and stress testing complement one another in a bank’s planning process.
2. Discuss the assumptions and limitations of using expected monetary value in a decision tree.
3. Compare VaR and expected shortfall as measures of market or credit-loss risk.
4. Explain how probability distributions and correlations should be selected and validated in a financial model.
5. Discuss governance and ethical controls required for a model used in lending or insurance pricing.

Analytical / Application Questions

1. A project has fixed costs of 300,000, a selling price of 75 and variable cost of 45. Calculate accounting break-even volume. Then explain how a 10% fall in price affects the result, holding other assumptions constant.
2. A loan portfolio has exposure of 60 million, PD of 4% and LGD of 50%. Calculate expected loss. Construct an adverse scenario with PD of 9% and LGD of 60%, and interpret the change.
3. Design a stress-testing framework for an insurer exposed to rising interest rates, falling equity prices and increased claims. Identify variables, dependencies, management actions and communication requirements.

Multiple Choice Questions with Answers

Question 1

A project has fixed costs of 240,000, a price of 80 per unit and variable cost of 50 per unit. What is its accounting break-even volume?

A. 6,000 units

B. 8,000 units

C. 10,000 units

D. 12,000 units

Correct Answer: B. 8,000 units

Question 2

Which statement best describes a stress test?

A. It is always the most statistically probable forecast.

B. It examines severe but plausible conditions.

C. It changes exactly one input while holding all others constant.

D. It eliminates model risk through simulation.

Correct Answer: B. It examines severe but plausible conditions.

Question 3

Exposure at default is 20 million, probability of default is 5%, and loss given default is 40%. What is expected loss?

A. 200,000

B. 400,000

C. 800,000

D. 1,000,000

Correct Answer: B. 400,000

Question 4

A decision-maker who chooses the action with the highest expected monetary value is implicitly using which assumption?

A. Risk neutrality

B. Perfect information

C. Zero uncertainty

D. Guaranteed outcomes

Correct Answer: A. Risk neutrality

Question 5

What does correlation measure?

A. Causation between two variables

B. The maximum possible loss

C. Linear association between two variables

D. The probability that both variables equal zero

Correct Answer: C. Linear association between two variables

Question 6

If the number of independent Monte Carlo trials increases by a factor of four, the approximate standard error of the simulated mean becomes:

A. Twice as large

B. Unchanged

C. Half as large

D. Four times as large

Correct Answer: C. Half as large

Question 7

A 99% one-day VaR of 5 million indicates that, under the model:

A. The maximum one-day loss is 5 million.

B. The average one-day loss is 5 million.

C. Losses exceed 5 million on approximately 99% of days.

D. Losses exceed 5 million on approximately 1% of days.

Correct Answer: D. Losses exceed 5 million on approximately 1% of days.

Question 8

Which measure focuses on the average loss beyond a VaR threshold?

A. Expected shortfall

B. Break-even volume

C. Correlation

D. Operating leverage

Correct Answer: A. Expected shortfall

Question 9

For a loss variable with discrete outcomes, why must expected shortfall be defined carefully?

A. A discrete variable has no probabilities.

B. The quantile may contain positive probability mass.

C. VaR cannot be calculated for losses.

D. Discrete outcomes always have normal distributions.

Correct Answer: B. The quantile may contain positive probability mass.

Question 10

Which practice most directly helps prevent false precision?

A. Reporting every output to many decimal places

B. Using only the baseline forecast

C. Matching reported precision to the quality of evidence

D. Ignoring model limitations

Correct Answer: C. Matching reported precision to the quality of evidence