Probability, Statistics and Financial Risk Analytics

Introduction

Financial risk management converts uncertain future outcomes into measurable quantities that support pricing, hedging, capital allocation and governance. Probability provides a language for uncertainty; statistics allows analysts to estimate unknown relationships from data; and financial risk analytics applies these tools to market, credit, liquidity and operational exposures.

The resulting measures are not objective facts. Volatility, correlation, value at risk and regression coefficients depend on data selection, assumptions and estimation methods. A technically correct calculation may therefore remain unsuitable for decision-making if the underlying data are poor, the model is misapplied or structural market changes are ignored.

Learning Objectives

After studying this chapter, students should be able to:

1. Represent financial gains and losses using random variables and probability distributions.
2. Calculate and interpret expected value, variance, standard deviation, covariance and correlation.
3. Explain the roles and limitations of commonly used probability distributions.
4. Estimate relationships using correlation and linear regression.
5. analyze time-series returns and estimate volatility using historical, EWMA and GARCH approaches.
6. Calculate and interpret value at risk and expected shortfall.
7. Apply historical and Monte Carlo simulation to financial risk problems.
8. Evaluate sampling error, data-quality weaknesses and model risk.
9. Apply quantitative evidence to portfolio and hedging decisions.

1. From Financial Data to Risk Measures

A financial risk variable must be defined before it is measured. Examples include the daily loss on a bond portfolio, the annual default rate of a loan book and the change in the domestic-currency value of a foreign receivable.

Let P_t be an asset price at time t. Its simple return is:

Here, D_t is any cash distribution received during the period. Prices and distributions must use the same currency, while R_t is dimensionless and is usually reported as a percentage.

The continuously compounded, or log, return excluding distributions is:

Log returns are additive through time because ln(P_T/P₀) = ∑_t=1^T r_t. Simple returns aggregate through compounding rather than addition. Log returns are undefined when the relevant price is non-positive and may be inappropriate for positions whose values can cross zero.

Risk systems frequently define loss as the negative of profit and loss:

Thus, a positive L_t represents a loss. The sign convention must be documented because reversing it changes the interpretation of tail quantiles.

Figure: Financial risk analytics is an iterative process rather than a one-time calculation.

2. Probability Foundations

2.1 Random variables and distributions

A random variable assigns a numerical value to each possible outcome. A discrete variable has countable outcomes and is described by a probability mass function p(x) = P(X=x). A continuous variable is described by a probability density function f(x), for which probabilities are areas under the density.

The cumulative distribution function is applicable to both types:

A valid distribution has non-negative probabilities and total probability equal to one. For a continuous distribution, P(X=x)=0 at any single point even though intervals can have positive probability.

2.2 Expected value and dispersion

For a discrete random variable, expected value is:

For a continuous variable with density f(x), it is:

Expected value is a probability-weighted long-run average, not necessarily the most likely outcome. Variance and standard deviation measure dispersion around the mean μ = E(X):

Variance has squared units; standard deviation has the same units as X. Both require a finite second moment.

Worked Example: Discrete Operational Loss

Suppose an incident produces losses of $0, $10,000 or $30,000 with probabilities 0.80, 0.15 and 0.05.

Therefore:

The standard deviation exceeds the expected loss because infrequent large losses create substantial dispersion.

2.3 Conditional probability and dependence

Conditional probability is:

Bayes’ theorem reverses conditioning:

Two events are independent only if P(A∩B)=P(A)P(B). Zero correlation does not generally imply independence, although independence with finite variances implies zero covariance.

3. Probability Distributions in Finance

If X ∼ N(μ,σ²), its standardized value is:

The familiar 68–95–99.7 rule is specific to a normal distribution. It must not be applied mechanically to heavy-tailed financial data. Student’s t distributions can represent more extreme observations, although very low degrees of freedom may imply that some moments do not exist.

Under the geometric Brownian motion model, an asset price follows:

where S_t is price, μ is instantaneous drift, σ is volatility per square root of time and W_t is a Wiener process. Under constant parameters and continuous paths, future prices are lognormally distributed. Real markets may exhibit jumps, stochastic volatility and changing parameters, so this is a model rather than a universal law.

4. Covariance, Correlation and Portfolio Risk

Covariance measures joint variation:

Its units are the product of the units of X and Y. Correlation standardizes covariance:

Provided both standard deviations are positive, -1 ≤ ρ_XY ≤ 1. Correlation measures linear dependence, not causality, and can conceal nonlinear or tail dependence.

For portfolio return R_p = ∑_i=1ⁿw_iR_i, where weights are fractions of portfolio value, portfolio variance is:

For two assets:

This assumes finite variances and covariances. If w_A=0.60, w_B=0.40, σ_A=2%, σ_B=1% and ρ_AB=0.25, then:

Thus σ_p = √0.000184 ≈ 0.01356, or 1.356% per period. Diversification reduces risk because correlation is below one; it does not eliminate risk.

5. Sampling and Statistical Estimation

Risk analysts usually observe a sample rather than the full population. For observations x₁,…,x_n, the sample mean and unbiased sample variance are:

The denominator n-1 corrects bias when estimating population variance from an independent, identically distributed sample. It does not guarantee that the estimate is accurate or stable.

Sample covariance is:

Sampling error is the difference between a sample estimate and the population parameter caused by observing only a sample. Under independent observations with finite variance, the standard error of the sample mean is s/√n. Serial dependence requires methods such as heteroskedasticity-and-autocorrelation-consistent standard errors or time-series models.

Confidence intervals express estimation uncertainty. Under normality, or approximately for large independent samples, a mean may be reported as:

The relevant Student’s t critical value should be used when population variance is unknown and the classical assumptions apply.

6. Regression for Risk-Factor Analysis

A simple linear regression relates a dependent variable to a risk factor:

Here, Y_t may be a portfolio return, X_t a market return, α the intercept, β the exposure or sensitivity and ε_t the unexplained component.

Ordinary least squares chooses coefficients that minimize ∑ε̂_t². In a simple regression with an intercept:

If weekly covariance between a fund and market return is 0.00030 and market-return variance is 0.00020, then β̂=1.50. On average, a one-percentage-point increase in market return is associated with an estimated 1.5-percentage-point increase in fund return, holding the linear model interpretation constant.

The coefficient of determination, R², is the proportion of sample variation in the dependent variable explained by the fitted regression. A high R² does not establish causation or guarantee accurate forecasts.

Classical inference typically assumes correct linear specification, zero conditional mean E(ε_t|X_t)=0, no perfect multicollinearity and suitable error behavior. Heteroskedasticity and serial correlation do not necessarily bias OLS coefficients under exogeneity, but they can invalidate conventional standard errors. Endogeneity, omitted variables and measurement error can bias coefficients themselves.

7. Financial Time Series and Volatility

Financial observations are ordered through time. Returns often display weak autocorrelation in levels but strong autocorrelation in squared or absolute returns. This pattern, called volatility clustering, means tranquil and turbulent periods tend to persist.

7.1 Rolling historical volatility

Using a window of m returns, rolling volatility is their sample standard deviation. A short window responds quickly but is noisy; a long window is more stable but slower to recognize regime changes. Old observations receive equal weight until they abruptly leave the window.

7.2 Exponentially weighted moving average

The EWMA model updates conditional variance as:

Here, 0<λ<1, returns are usually demeaned, and all terms use consistent periodic units. Larger λ produces slower decay.

If yesterday’s volatility was 1.2%, yesterday’s demeaned return was −2.0%, and λ=0.94, then:

7.3 GARCH volatility

A GARCH(1,1) model is:

For non-negative conditional variances, standard sufficient restrictions are ω>0, α≥0 and β≥0. Under the conventional specification, covariance stationarity requires α+β<1. Its unconditional variance is then:

GARCH captures persistent conditional variance but may need extensions for asymmetric responses, structural breaks or fat-tailed innovations.

7.4 Time aggregation

The square-root-of-time rule is:

It is valid when one-period returns have constant variance and are serially uncorrelated. It is not a general law. Serial correlation, changing volatility and overlapping observations can invalidate the scaling.

8. Tail-Risk Measurement

8.1 Value at risk

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

A one-day 99% VaR of $1 million means that, under the model and current conditions, the probability of a loss exceeding $1 million over one day is at most approximately 1%. It does not mean the maximum possible loss is $1 million.

If losses are normally distributed with mean μ_L and standard deviation σ_L, then:

For μ_L=$20,000, σ_L=$150,000 and z_0.99=2.326, the 99% VaR is approximately:

Normal VaR can seriously understate risk when losses are skewed or heavy-tailed.

8.2 Expected shortfall

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

For normally distributed losses:

where φ is the standard normal density. Expected shortfall captures the severity of tail losses and, under standard definitions, is a coherent risk measure. Its estimation is difficult because relatively few observations occur in the tail.

9. Simulation Methods

Historical simulation

Historical simulation applies observed past changes in risk factors to today’s positions and ranks the resulting hypothetical losses. It makes no explicit normality assumption and preserves historical cross-factor relationships. However, it assumes the selected past scenarios are informative about the future. Rare events absent from the sample cannot appear without supplementary stress scenarios.

Monte Carlo simulation

Monte Carlo simulation specifies a stochastic model, draws many random scenarios, revalues the portfolio and estimates the loss distribution. A standard process is:

1. Define risk factors and their joint distribution.
2. Estimate parameters and dependence.
3. Generate correlated random draws.
4. Revalue positions in each scenario.
5. Rank losses and calculate risk measures.
6. assess convergence and conduct sensitivity tests.

Simulation error decreases slowly: for many estimators, precision improves approximately with the square root of the number of independent simulations. Doubling precision may therefore require roughly four times as many trials. Monte Carlo flexibility does not remove model risk; it transfers assumptions into distributions, correlations and valuation functions.

10. Illustrative Case: Hedging a Foreign-Currency Receivable

Background

This is an illustrative case. Meridian Components expects a foreign-currency customer payment in three months. A depreciation of the foreign currency would create a dollar loss. Management is considering standardized futures contracts whose gains tend to offset this loss.

Problem

Let:

• X = unhedged dollar loss on the receivable;
• Y = dollar gain on one futures contract;
• n = number of contracts;
• L_H=X-nY = hedged dollar loss.

Both X and Y are measured in dollars, so their variances and covariance are measured in dollar-squared units. Based on validated scenario data, the treasury team estimates:

Analysis and approach

The variance of the hedged loss is:

Differentiating with respect to n and setting the derivative equal to zero gives:

Therefore, when Var(Y)>0, the minimum-variance hedge is:

Substitution gives:

The dollar-squared units cancel, leaving a number of contracts. The calculation divides covariance by variance; neither quantity is squared again.

At the continuous optimum:

Thus the hedged standard deviation is approximately √$19,000,000=$4,358.90, compared with $10,000 before hedging.

Because contracts are indivisible, management must choose four or five contracts. Either produces:

and a standard deviation of approximately $4,472.14. Transaction costs, margin liquidity, accounting treatment and basis risk may determine the final choice.

Outcome

The analysis indicates that either four or five contracts substantially reduces estimated variance, but the hedge is imperfect. Management chooses four contracts to avoid over-hedging and supplements the statistical analysis with exchange-rate stress tests.

Learning points

• Hedge ratios must use consistent definitions, signs and units.
• A minimum-variance hedge does not necessarily minimize expected loss or extreme tail loss.
• Estimated covariance may change during market stress.
• Contract indivisibility, liquidity and basis risk affect implementation.
• Model output should inform, not replace, treasury judgment.

11. Data Quality, Model Risk and Governance

11.1 Data-quality risks

Common weaknesses include missing observations, stale prices, survivorship bias, inconsistent timestamps, corporate-action errors, currency mismatches and duplicate records. Illiquid assets may appear artificially stable because prices are infrequently updated. Outliers should be investigated rather than automatically deleted: they may be errors, but they may also be genuine stress events.

Analysts should maintain data lineage, reconcile inputs to authoritative sources, document transformations and preserve an auditable raw-data layer. Training, validation and test periods should be separated to prevent look-ahead bias.

11.2 Model risk

Model risk is the possibility of adverse consequences from incorrect models or inappropriate use. It arises from:

• specification error, such as assuming normality for heavy-tailed losses;
• estimation error from limited samples;
• implementation error in code or data mapping;
• parameter instability and structural breaks;
• use outside the model’s intended scope;
• human misinterpretation of outputs.

Sound governance includes independent validation, benchmark models, back-testing, stress testing, sensitivity analysis, version control, change approval and clear limitations. VaR back-testing compares predicted exceptions with realized outcomes, but passing a back-test does not prove that the model is correct.

11.3 Ethics, regulation and AI

Risk estimates can influence customer limits, capital allocation and executive remuneration. Analysts have an ethical responsibility not to conceal uncertainty, manipulate model choices or suppress unfavorable stress results. Assumptions and overrides should be transparent and reproducible.

Regulatory frameworks increasingly emphasize expected shortfall, stress testing, model governance and prudent valuation. Exact requirements depend on jurisdiction and institution type.

Machine learning can detect nonlinear relationships and process large alternative datasets, but it can amplify bias, overfit changing markets and reduce explainability. Responsible use requires lawful data, out-of-sample testing, explainability proportionate to impact, human oversight and continuous drift monitoring.

12. Professional Best Practices

A defensible financial risk analysis should:

1. Define the exposure, horizon, currency and loss sign convention.
2. Match the model to the economic purpose rather than selecting it solely for convenience.
3. Validate data before estimation.
4. Report point estimates together with uncertainty and sensitivity.
5. Compare multiple methods, such as historical, parametric and simulation estimates.
6. Use stress scenarios to supplement probability-based measures.
7. Back-test forecasts and investigate exceptions.
8. Document assumptions, limitations, ownership and approval.
9. Recalibrate only under controlled procedures.
10. Communicate results in language decision-makers can understand.

Chapter Summary

Probability distributions describe uncertain financial outcomes, while expected value and variance summarize location and dispersion. Covariance and correlation quantify linear co-movement and provide the basis for portfolio diversification and minimum-variance hedging. Statistical estimates are sample-dependent and must be interpreted with awareness of sampling uncertainty.

Regression estimates exposure to risk factors, but valid interpretation depends on specification, exogeneity and suitable treatment of heteroskedasticity and serial correlation. Financial time series commonly exhibit volatility clustering, motivating rolling, EWMA and GARCH estimates. VaR describes a loss quantile, whereas expected shortfall measures average tail severity. Historical and Monte Carlo simulation offer flexible alternatives, but their outputs remain conditional on data and model assumptions.

Reliable risk analytics combines quantitative methods with validated data, stress testing, professional judgment and independent governance. Models simplify reality; they do not eliminate uncertainty.

Key Terms

Random variable: A numerical representation of an uncertain outcome.

Probability distribution: A rule assigning probabilities to possible values of a random variable.

Expected value: The probability-weighted mean of a random variable.

Variance: Expected squared deviation from the mean.

Standard deviation: The positive square root of variance.

Covariance: A measure of joint variation between two variables.

Correlation: Standardized covariance bounded between −1 and 1.

Sampling error: Difference between a sample estimate and its population parameter caused by sampling.

Regression: A method for estimating relationships between a dependent variable and explanatory variables.

Beta: Estimated sensitivity of an asset or portfolio return to a specified risk factor.

Time series: Observations indexed in chronological order.

Volatility clustering: Persistence of high- and low-volatility periods.

EWMA: A volatility model assigning exponentially declining weights to past squared returns.

GARCH: A model in which conditional variance depends on past shocks and past variance.

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

Expected shortfall: Expected loss in the tail beyond a specified quantile.

Historical simulation: Risk estimation using observed historical risk-factor changes.

Monte Carlo simulation: Risk estimation using repeated random draws from a specified model.

Model risk: Risk of adverse outcomes from incorrect models or their misuse.

Back-testing: Comparison of model forecasts with realized outcomes.

Review Questions

Short Answer Questions

1. Distinguish between a simple return and a log return.
2. Why does standard deviation have more interpretable units than variance?
3. Explain why zero correlation does not generally imply independence.
4. What does a one-day 99% VaR measure?
5. Identify two common sources of financial data-quality risk.

Long Answer Questions

1. Explain how expected value, variance, covariance and correlation contribute to financial risk analysis.
2. Compare the normal and Student’s t distributions for modeling asset returns.
3. Discuss the assumptions, strengths and limitations of ordinary least squares in risk-factor analysis.
4. Compare rolling historical volatility, EWMA and GARCH models.
5. Evaluate historical simulation and Monte Carlo simulation as methods of estimating tail risk.

Analytical / Application Questions

1. A portfolio invests 40% in an asset with volatility 3% and 60% in an asset with volatility 2%. Their correlation is 0.30. Calculate portfolio volatility and explain the diversification effect.
2. A risk analyst reports a low VaR after deleting all observations exceeding four standard deviations. Evaluate the statistical, ethical and governance implications.
3. A hedge ratio estimated from calm-market data performs poorly during a currency crisis. Analyze possible statistical and economic causes and recommend improvements.

Multiple Choice Questions with Answers

Question 1

A loss is $0 with probability 0.80, $10,000 with probability 0.15 and $30,000 with probability 0.05. What is the expected loss?

A. $1,500

B. $3,000

C. $6,000

D. $10,000

Correct Answer: B. $3,000

Question 2

For the same loss distribution, what is the variance of loss?

A. $9,000,000 dollar squared

B. $51,000,000 dollar squared

C. $60,000,000 dollar squared

D. $7,141 dollar squared

Correct Answer: B. $51,000,000 dollar squared

Question 3

Two returns have covariance 0.00012 and standard deviations 0.02 and 0.03. What is their correlation?

A. 0.02

B. 0.12

C. 0.20

D. 0.60

Correct Answer: C. 0.20

Question 4

A portfolio has weights 0.60 and 0.40, asset volatilities 2% and 1%, and correlation 0.25. What is its approximate volatility per period?

A. 0.92%

B. 1.36%

C. 1.60%

D. 3.00%

Correct Answer: B. 1.36%

Question 5

Under EWMA, yesterday's volatility was 1.2%, yesterday's demeaned return was −2.0%, and λ=0.94. What is the new volatility estimate?

A. 1.20%

B. 1.26%

C. 1.88%

D. 2.00%

Correct Answer: B. 1.26%

Question 6

In a simple regression with an intercept, sample covariance between fund and market returns is 0.00030 and market-return variance is 0.00020. What is the estimated beta?

A. 0.67

B. 1.00

C. 1.50

D. 2.00

Correct Answer: C. 1.50

Question 7

Which statement correctly interprets a one-day 99% VaR of $1 million?

A. The portfolio cannot lose more than $1 million in one day.

B. The expected one-day loss is exactly $1 million.

C. Under the model, losses exceed $1 million on approximately 1% of days.

D. There is a 99% probability that the portfolio earns $1 million.

Correct Answer: C. Under the model, losses exceed $1 million on approximately 1% of days.

Question 8

Which risk measure describes average loss severity in the tail beyond a specified loss quantile?

A. Expected shortfall

B. Sample variance

C. Regression beta

D. Tracking correlation

Correct Answer: A. Expected shortfall

Question 9

If Cov(X,Y)=$18,000,000 dollar squared and Var(Y)=$4,000,000 dollar squared, what is the continuous minimum-variance hedge quantity in L_H=X-nY?

A. 0.22 contracts

B. 4.00 contracts

C. 4.50 contracts

D. 20.25 contracts

Correct Answer: C. 4.50 contracts

Question 10

Which practice most directly prevents look-ahead bias when developing a forecasting model?

A. Increasing the confidence level of VaR

B. Separating training, validation and test periods chronologically

C. Deleting all extreme observations

D. Assuming returns are normally distributed

Correct Answer: B. Separating training, validation and test periods chronologically