Quantitative Risk Measurement and Stress Testing

Introduction

Treasury risk management converts uncertain market movements into measures that support limits, hedging, capital allocation and contingency planning. No single measure is sufficient. Volatility and correlation describe ordinary variability; sensitivities approximate exposure to specified risk-factor changes; Value at Risk (VaR) estimates a loss quantile; expected shortfall (ES) measures average tail loss; back-testing evaluates forecast performance; and scenario analysis investigates conditions that may lie outside recent data.

These tools are conditional on assumptions about data, distributions, portfolio composition, liquidity and market behaviour. A numerically precise estimate can therefore be misleading if its assumptions do not fit the decision. Sound practice combines quantitative results with model validation, stress testing, governance and professional judgement.

Learning Objectives

After studying this chapter, students should be able to:

1. calculate and interpret total returns, volatility, covariance and correlation;
2. evaluate portfolio diversification using the variance–covariance framework;
3. apply duration, DV01 and other sensitivity measures;
4. calculate and interpret parametric, historical and simulation-based VaR;
5. explain and estimate expected shortfall;
6. assess VaR models through back-testing and exception analysis;
7. design sensitivity tests, scenarios, stress tests and reverse stress tests;
8. identify model limitations and translate results into treasury decisions.

1. Returns and the Measurement Horizon

Risk measurement begins with a consistent definition of return. Let P_t-1 be the initial asset price, P_t the ending price and C_t cash income received over the period, such as interest or a dividend. The simple total return is:

All price and income variables must use the same currency and valuation basis. If income is treated as part of end-of-period wealth, the corresponding total log return is:

This formulation assumes that income is received at the end of the measurement period. A log price return, ln(P_t ÷ P_t-1), excludes income and should not be compared directly with an income-inclusive simple return.

Worked example: consistent total returns

A security rises from $100 to $103 and pays $2 at the end of the period.

By contrast, its log price return is ln(103 ÷ 100) ≈ 2.956%, because that measure deliberately excludes the $2 income.

Log returns are additive through time when they refer to the same consistently defined wealth process. Simple returns aggregate naturally across portfolio holdings at a single date. For small returns, simple and log returns are close, but they are not identical.

The chosen horizon should match the decision. A trading desk may use one-day risk, while liquidity and capital planning require longer horizons. Square-root-of-time scaling,

relates one-period volatility σ₁ to h periods only when returns are independent, identically distributed and have finite, constant variance. Volatility clustering, autocorrelation, changing positions and market illiquidity can invalidate this rule.

2. Volatility, Covariance and Correlation

2.1 Volatility

Volatility measures dispersion rather than direction. Given n observed returns and sample mean \bar{R}, sample variance and sample standard deviation are:

The denominator n-1 produces the usual unbiased estimator of population variance when observations are independent draws with a common finite variance. Standard deviation has the same periodicity as the returns—for example, daily or monthly—and is commonly annualised under the restrictive square-root-of-time assumptions.

Volatility treats positive and negative deviations symmetrically. It therefore measures uncertainty, not loss alone. Downside deviation, quantile measures and expected shortfall may be more informative when the concern is specifically adverse outcomes.

2.2 Covariance and correlation

Sample covariance between returns R_1,t and R_2,t is:

Covariance depends on the units and volatility of both variables. Correlation standardises covariance:

Correlation indicates linear association, not causation. A value near zero does not rule out nonlinear dependence. Moreover, correlations estimated in calm markets may rise during crises, reducing expected diversification precisely when it is most needed.

2.3 Portfolio volatility

For portfolio weights w_i, covariance matrix Σ and weight vector w, portfolio variance is:

This expression applies to linear portfolio returns when the covariance matrix corresponds to the same horizon and currency basis as the returns. For two assets:

The weights normally sum to one for a fully invested unleveraged portfolio, although the formula also accommodates short and leveraged positions.

Worked example: correlation and diversification

Suppose 60% is invested in Asset 1 with volatility 10%, and 40% in Asset 2 with volatility 15%. Thus:

w₁σ₁ = 0.60 × 0.10 = 0.06 and w₂σ₂ = 0.40 × 0.15 = 0.06.

Therefore:

At ρ=-1, the equally sized weighted volatility exposures offset exactly, producing zero volatility in this special two-asset example. This does not mean every negatively correlated portfolio is riskless; exact cancellation requires w₁σ₁=w₂σ₂ and perfect negative correlation.

Figure: Portfolio volatility equals √[0.0072(1+ρ); it is zero at correlation −1 and rises at a decreasing rate to 12%.

The curve is concave because it is the square root of an increasing affine function of correlation. The benefit from an additional reduction in correlation is therefore nonlinear.

3. Sensitivity Analysis

Sensitivity measures estimate how portfolio value changes when one risk factor moves while other assumptions are held fixed. They are local approximations rather than complete probability models.

3.1 Interest-rate sensitivity

Modified duration approximates the percentage price change of a conventional fixed-income instrument for a small parallel yield change:

Here, V is market value in currency units, D_mod is modified duration measured in years, and Δy is the yield change in decimal form. The approximation assumes a small shift and unchanged cash flows. It becomes less accurate for large changes, nonparallel yield-curve movements or instruments with embedded options.

Dollar value of one basis point, commonly called DV01 or PV01, is:

For a $10 million bond position with modified duration 4.2:

A 25-basis-point yield increase produces an approximate loss of $4,200 × 25 = $105,000. Convexity can be added for a better second-order approximation:

Here C is price convexity under the convention used by the institution. Because convexity conventions differ, treasury systems must document their scaling and units.

3.2 Other sensitivities

Sensitivity reports should state the shock size, direction, currency, valuation date and whether cross-effects are included. Adding sensitivities is reliable only when positions share consistent risk-factor definitions and valuation conventions.

4. Value at Risk

4.1 Definition and interpretation

Let L denote portfolio loss over a specified horizon, where positive values represent losses. At confidence level α, VaR is the loss quantile:

A one-day 99% VaR of $3 million means that the model estimates a 99% probability that the one-day loss will not exceed $3 million. Equivalently, losses may exceed $3 million on approximately 1% of days if the model is correct and conditions remain stable. VaR is not the maximum possible loss and says nothing by itself about the size of losses beyond the threshold.

4.2 Parametric variance–covariance VaR

If loss is normally distributed with mean μ_L and standard deviation σ_L, then:

Here z_α is the standard normal quantile. For example, z_0.99≈2.326. If daily mean loss is assumed to be zero and portfolio daily standard deviation is $1.2 million:

Under independent, identically distributed daily changes and a constant portfolio, ten-day VaR may be approximated as:

This scaling is unreliable when returns are autocorrelated, volatility changes, positions are nonlinear or liquidation affects market prices.

Delta-normal VaR first approximates nonlinear instruments by their deltas. It is efficient but can understate option risk when gamma, volatility shifts or discontinuous payoffs matter.

4.3 Historical simulation

Historical simulation applies observed past risk-factor changes to the current portfolio, revalues it and takes an empirical loss quantile. It avoids a normality assumption and can preserve observed cross-factor dependence. Its weaknesses are equally important:

• the selected window may omit relevant crises;
• all observations are often given equal weight;
• current volatility may differ from historical volatility;
• unprecedented events cannot appear without additional scenarios;
• results depend on quantile and interpolation conventions.

Filtered historical simulation rescales historical innovations using a time-varying volatility model. This responds better to changing volatility but introduces further model risk.

4.4 Monte Carlo simulation

Monte Carlo VaR simulates risk-factor paths from an estimated stochastic model and fully or approximately revalues the portfolio. It can represent nonlinear derivatives and complex dependence structures. Its accuracy depends on distributional assumptions, calibration, the number of simulations and valuation quality. Computational precision does not compensate for an incorrectly specified model.

5. Expected Shortfall

Expected shortfall measures average loss in the tail beyond the VaR threshold. For a continuous loss distribution:

For distributions with probability masses, a quantile-integral definition is more robust:

If loss is normally distributed:

where φ is the standard normal density. For the earlier portfolio, φ(2.326)≈0.02665, so:

Thus, the average loss in the worst 1% of normally modelled days is approximately $3.198 million, compared with 99% VaR of $2.791 million.

Expected shortfall is a coherent risk measure under standard definitions: it satisfies monotonicity, translation invariance, positive homogeneity and subadditivity. VaR can fail subadditivity for some non-elliptical or discontinuous loss distributions. Nevertheless, ES is harder to back-test because it concerns both the frequency and magnitude of relatively rare tail observations.

6. Back-Testing and Model Validation

Back-testing compares ex ante risk forecasts with realised outcomes. A VaR exception occurs when realised loss exceeds the forecast VaR for that date. If the model is correct, the expected exception probability is p=1-α. For T independent observations, the number of exceptions X follows:

For 250 trading days and 99% VaR, the expected number is:

A few exceptions do not automatically invalidate the model. Too many indicate underestimation of risk, while too few may indicate excessive conservatism or an inappropriate testing period.

The Kupiec unconditional coverage statistic tests whether the observed exception rate equals the target rate:

Here x is the observed number of exceptions. Under the null hypothesis and usual large-sample conditions, LR_uc is approximately chi-squared with one degree of freedom. The test examines frequency but not whether exceptions cluster. An independence test is needed to detect clustering; conditional coverage combines frequency and independence.

For regulatory market-risk back-testing, a commonly used 250-day 99% VaR traffic-light framework classifies 0–4 exceptions as green, 5–9 as yellow and 10 or more as red. Institutions must apply the current rules of their own jurisdiction rather than treating these historical thresholds as universal.

Effective validation also includes:

• data-quality and reconciliation checks;
• benchmark-model comparisons;
• sensitivity to windows and parameters;
• profit-and-loss attribution;
• review of valuation reserves and missing risk factors;
• independent validation and documented remediation.

Back-testing can fail because of model weakness, but also because the forecast and realised profit-and-loss series use different positions, fees, valuation adjustments or intraday trading assumptions.

7. Scenario Analysis and Stress Testing

Sensitivity analysis changes one factor, whereas scenario analysis changes several factors coherently. Stress testing applies severe but plausible conditions to identify vulnerabilities that ordinary distribution-based measures may miss.

7.1 Main forms

Historical scenarios replay episodes such as a currency crisis, liquidity shock or rapid rate increase. They are intuitive but may not reflect the current portfolio or financial structure.

Hypothetical scenarios combine forward-looking shocks based on economic narratives. For example, inflation surprise may produce higher yields, wider credit spreads, weaker equities and a stronger safe-haven currency.

Reverse stress testing begins with a failure outcome—such as a liquidity breach or capital ratio falling below a threshold—and identifies scenarios capable of causing it. It is especially useful for revealing hidden concentrations.

Liquidity stress testing estimates cash outflows, collateral calls, deposit withdrawals, market haircuts and asset-sale capacity over defined horizons. Market and funding liquidity should be integrated because forced sales can deepen losses.

7.2 Designing a defensible stress test

A sound process includes:

1. defining the decision, portfolio and horizon;
2. identifying material risk factors and concentrations;
3. constructing a coherent economic narrative;
4. translating the narrative into quantified shocks;
5. revaluing positions, preferably with full valuation for nonlinear instruments;
6. incorporating liquidity, collateral and management constraints;
7. comparing results with capital, liquidity and risk appetite;
8. assigning actions, owners and escalation thresholds.

Stress tests should avoid assuming that all markets remain liquid or that historical correlations remain stable. They should also distinguish first-round valuation losses from second-round effects such as margin calls, fire sales, counterparty deterioration and customer withdrawals.

Figure: Stress testing links an economic narrative to quantified shocks, valuation effects, financial capacity and management action.

8. Illustrative Case: Meridian Bank Treasury

Background and problem

Meridian Bank is a fictional mid-sized bank with $30 million of available risk-bearing capital allocated to treasury activities. Its portfolio contains:

• $80 million of government bonds with modified duration 5.5;
• $40 million of corporate bonds with modified duration 3.8;
• $20 million of equities;
• a net $15 million USD asset position measured in its home currency.

Management specifies a severe hypothetical scenario: government yields rise by 150 basis points, corporate yields rise by 350 basis points in total because of combined benchmark-rate and spread movements, equities fall 25%, and the home currency appreciates 10% against the US dollar. The internal stress-loss limit is 40% of allocated capital.

Analysis

Using first-order duration approximations:

The loss equals 18.42 ÷ 30 = 61.4% of allocated capital, exceeding the 40% limit. The calculation excludes convexity, defaults, basis risk, taxes and market-impact costs; these omissions must be considered before a final decision.

Approach and solution

Treasury proposes interest-rate swaps or futures to reduce government-bond duration to 3.0 and corporate-bond effective duration to 2.0, protective equity options that cap the scenario loss at 15%, and FX forwards that reduce the net USD position to $5 million. Ignoring hedge premiums, basis effects and counterparty adjustments, the revised stress loss is:

Outcome

The revised loss is 33.0% of allocated capital, which is within the 40% limit. Approval should nevertheless depend on hedge cost, liquidity, accounting effects, basis risk and counterparty credit limits.

Learning points

The case demonstrates that VaR alone would not reveal the consequences of the specified joint shock. Stress testing connects exposures to capital capacity and concrete actions. It also shows why first-order sensitivity calculations are useful for screening but should be supplemented by full revaluation when positions are nonlinear or shocks are large.

9. Model Risk, Governance and Contemporary Practice

Every quantitative result is a model output rather than an objective fact. Important sources of model risk include poor data, survivorship bias, unstable correlations, omitted risk factors, inappropriate distributions, valuation errors and behavioural responses not represented in historical data.

Governance should establish clear ownership among model developers, users, independent validators, senior management and the board. Material assumptions, overrides and limitations should be documented. Limits should specify escalation and remediation rather than merely produce reports.

Ethical practice requires transparent communication. Managers should not select confidence levels, data windows or scenarios merely to obtain favourable results. Uncertainty and known exclusions must be disclosed to decision-makers. Compensation arrangements should not reward risk-taking based solely on short-horizon VaR.

Machine learning can improve volatility forecasts, detect nonlinear relationships and generate scenario candidates. However, unstable training data, limited explainability and hidden overfitting can make automated models fragile. Human review, out-of-sample testing, explainability controls and data governance remain essential.

Contemporary practice increasingly combines expected shortfall, stressed calibration, liquidity horizons, climate-related scenarios and network or counterparty analysis. These developments reinforce a central principle: robust risk management uses multiple complementary measures rather than searching for one definitive number.

Chapter Summary

Quantitative treasury risk management begins with consistently defined returns and horizons. Volatility measures dispersion, while covariance and correlation determine how individual exposures combine at portfolio level. Diversification depends on dependence structures that may change during crises.

Sensitivities such as duration and DV01 estimate local effects of risk-factor changes. VaR identifies a loss quantile but does not measure the severity of losses beyond that threshold. Expected shortfall addresses this limitation by averaging tail losses. Historical, parametric and Monte Carlo methods involve different assumptions and implementation risks.

Back-testing evaluates whether forecast exceptions are consistent with the selected confidence level and whether they cluster. Scenario analysis, stress testing and reverse stress testing examine severe joint movements, liquidity pressures and threats to capital or business viability. Effective use requires independent validation, ethical reporting, documented assumptions and translation of quantitative results into limits, hedging and contingency actions.

Key Terms

Back-testing: Comparison of forecast risk measures with realised outcomes.

Correlation: Standardised measure of linear association between two variables.

Covariance: Measure of how two variables vary together.

Delta: First-order sensitivity of an instrument’s value to its underlying variable.

Diversification: Risk reduction achieved by combining imperfectly dependent exposures.

DV01: Approximate currency-value change for a one-basis-point yield movement.

Expected shortfall: Average loss in the tail beyond a selected confidence quantile.

Exception: Realised loss exceeding the VaR forecast for the relevant date.

Historical simulation: Risk estimation by applying historical factor changes to the current portfolio.

Liquidity stress test: Assessment of cash-flow, collateral and asset-sale capacity under stress.

Model risk: Risk of adverse decisions arising from incorrect or misused models.

Monte Carlo simulation: Estimation using repeated draws from a specified stochastic model.

Reverse stress testing: Identification of scenarios capable of causing a predefined failure.

Scenario analysis: Evaluation of portfolio effects from a coherent set of factor changes.

Sensitivity analysis: Measurement of value changes caused by specified risk-factor shocks.

Stress testing: Evaluation under severe but plausible adverse conditions.

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

Volatility: Standard deviation of returns or value changes over a defined horizon.

Review Questions

Short Answer Questions

1. Distinguish a simple total return from a log price return.
2. Why can correlation estimates become unreliable during financial crises?
3. What does a one-day 99% VaR of $5 million mean?
4. Explain the difference between DV01 and modified duration.
5. What is a VaR exception?

Long Answer Questions

1. Compare parametric, historical simulation and Monte Carlo approaches to VaR.
2. Explain why expected shortfall may provide more information about tail risk than VaR.
3. Discuss the assumptions and limitations of square-root-of-time scaling.
4. Describe how an institution should design and govern a comprehensive stress-testing programme.
5. Evaluate the roles of back-testing, independent validation and profit-and-loss attribution in controlling model risk.

Analytical / Application Questions

1. A portfolio invests 40% in an asset with 12% volatility and 60% in an asset with 8% volatility. Calculate portfolio volatility when correlation is 0.25 and interpret the diversification effect.
2. A $25 million bond portfolio has modified duration 6.0. Estimate the loss from a 40-basis-point yield increase, then explain why the estimate may differ from full revaluation.
3. Design a reverse stress test for a bank that breaches its liquidity requirement if cumulative 30-day net cash outflows exceed $500 million. Identify relevant drivers, interactions and potential management responses.

Multiple Choice Questions with Answers

Question 1

A security rises from $100 to $104 and pays $1 at the end of the period. What is its total log return?

A. Approximately 3.922%

B. Approximately 4.879%

C. Exactly 5.000%

D. Approximately 5.127%

Correct Answer: B. Approximately 4.879%

Question 2

In the chapter's 60/40 portfolio example, w₁σ₁=w₂σ₂=0.06. What is portfolio volatility when ρ=-1?

A. 0%

B. 6%

C. 8.485%

D. 12%

Correct Answer: A. 0%

Question 3

A $20 million bond position has modified duration 3.5. What is its approximate DV01?

A. $700

B. $3,500

C. $7,000

D. $70,000

Correct Answer: C. $7,000

Question 4

Which interpretation of a one-day 99% VaR of $4 million is correct?

A. The maximum possible one-day loss is $4 million.

B. The expected one-day loss is $4 million.

C. The model estimates that losses will not exceed $4 million on approximately 99% of days.

D. Losses will equal $4 million on exactly 1% of days.

Correct Answer: C. The model estimates that losses will not exceed $4 million on approximately 99% of days.

Question 5

Daily loss is normally distributed with zero mean and standard deviation $2 million. Using z_0.99=2.326, what is 99% VaR?

A. $2.000 million

B. $4.000 million

C. $4.652 million

D. $6.978 million

Correct Answer: C. $4.652 million

Question 6

For a continuous loss distribution, what does 99% expected shortfall measure?

A. The median of all losses

B. The average loss conditional on being in the worst 1%

C. The largest loss in the sample

D. The probability of any positive loss

Correct Answer: B. The average loss conditional on being in the worst 1%

Question 7

Under a correctly calibrated 99% VaR model, how many exceptions are expected over 250 trading days?

A. 0.25

B. 1.0

C. 2.5

D. 25

Correct Answer: C. 2.5

Question 8

What is the main limitation of the Kupiec unconditional coverage test?

A. It cannot compare the observed exception rate with the target rate.

B. It examines exception frequency but not exception clustering.

C. It can be used only with normally distributed returns.

D. It requires every loss to exceed VaR.

Correct Answer: B. It examines exception frequency but not exception clustering.

Question 9

Which method begins with a predefined failure outcome and identifies conditions capable of causing it?

A. Historical simulation

B. Delta-normal VaR

C. Reverse stress testing

D. Volatility annualisation

Correct Answer: C. Reverse stress testing

Question 10

In the Meridian Bank case, what is the revised stress loss after the proposed exposure reductions, before hedge costs and basis effects?

A. $9.90 million

B. $12.00 million

C. $18.42 million

D. $30.00 million

Correct Answer: A. $9.90 million