Statistical Foundations for Finance

Introduction

Financial decisions are made under uncertainty. Banks estimate default probabilities, insurers forecast claim frequencies, investment analysts assess expected returns, and regulators evaluate whether observed risks are genuine or products of sampling variation. Statistical analysis provides a disciplined way to quantify this uncertainty.

A statistical result is not automatically a useful financial result. Analysts must distinguish between population characteristics and sample estimates, association and causation, statistical significance and economic significance, and model precision and real-world uncertainty. These distinctions are essential because financial data commonly exhibit volatility clustering, outliers, dependence, structural change and selection bias.

Learning Objectives

After studying this chapter, students should be able to:

• apply major probability distributions to financial problems;
• explain random sampling and sampling distributions;
• calculate point estimates, standard errors and confidence intervals;
• formulate and conduct hypothesis tests;
• distinguish statistical significance from economic significance;
• calculate and interpret covariance and correlation;
• estimate and interpret a simple linear regression;
• calculate a regression slope standard error and conduct a regression t-test;
• identify common errors in financial analytics; and
• communicate statistical uncertainty responsibly.

1. Probability Foundations

1.1 Random variables and probabilities

A random variable assigns a numerical value to each possible outcome of an uncertain process. A discrete random variable has countable values, such as the number of defaults in a loan portfolio. A continuous random variable can take values over an interval, such as a daily portfolio return.

For mutually exclusive outcomes, probabilities satisfy 0 ≤ P(A) ≤ 1 and the probabilities of all possible outcomes sum to one. Conditional probability is

where P(A|B) is the probability of event A given that B has occurred. Independence requires P(A ∩ B)=P(A)P(B). In finance, assuming independence without evidence can materially understate portfolio risk.

1.2 Expected value and variance

For a discrete random variable X with possible values x_i and probabilities p_i, its expected value is

The variance and standard deviation are

Expected return has the same unit as return, while variance is measured in squared return units. Standard deviation restores the original return unit.

Worked example: expected investment return

Suppose an annual investment return is −10%, 8% or 20%, with probabilities 0.20, 0.50 and 0.30 respectively. Expressing returns as decimals,

The expected return is 8%. Its variance is

Thus σ=√0.0108≈0.1039, or 10.39% per year. The expected return is not a guaranteed outcome.

2. Probability Distributions in Finance

A probability distribution specifies the possible values of a random variable and their probabilities or probability density. Distribution choice should reflect the economic process, measurement scale and empirical evidence.

2.1 Bernoulli and binomial distributions

A Bernoulli variable has two outcomes. For example, let D=1 if a borrower defaults within one year and D=0 otherwise. If the default probability is p,

If X counts successes across n independent Bernoulli trials with a common probability p, then X∼Binomial(n,p) and

Here C(n,k)=n!÷(k!(n−k)!). The binomial model requires a fixed number of trials, two outcomes per trial, constant p and independence. Defaults may become correlated during recessions, so these conditions must be checked rather than assumed.

2.2 Poisson distribution

The Poisson distribution models event counts over a specified interval when events occur independently at a constant average rate λ:

Its mean and variance both equal λ. Applications include claim arrivals, transaction failures and operational-loss events. If the observed variance greatly exceeds the mean, the data are overdispersed and a Poisson model may be unsuitable.

2.3 Normal distribution

A normal random variable is symmetric around its mean and is characterized by μ and σ². Standardization gives

where Z∼N(0,1). The normal distribution is useful for measurement errors, sampling distributions and some short-horizon return approximations. Actual asset returns often exhibit skewness and heavier tails, so a normal model may underestimate extreme losses.

Normal value-at-risk application. Suppose a portfolio is worth $10 million and its one-day return is modeled as normal with μ=0.0004 and σ=0.012. The 1% standard-normal quantile is approximately −2.326. The 1% return quantile is

The corresponding 99% one-day value at risk, stated as a positive loss amount, is approximately

This estimate assumes the stated normal model, unchanged portfolio positions and an appropriate one-day horizon. It is not the maximum possible loss.

2.4 Lognormal distribution

A positive variable X is lognormally distributed if ln(X) is normally distributed. A lognormal model prevents negative prices. If continuously compounded returns are normal under a specified model, the associated future price is lognormal. However, this is a model assumption rather than a universal law of market prices.

2.5 Student’s t, chi-square and F distributions

The Student’s t distribution resembles the normal distribution but has heavier tails. It arises in inference about a mean when population variance is unknown and the underlying observations are normal. It is also sometimes used to model heavy-tailed financial innovations, although that modeling use is conceptually distinct from its role in hypothesis testing.

The chi-square distribution supports inference about variance under normality. The F distribution compares variance estimates and underlies overall regression significance tests. These exact small-sample results depend on their assumptions; robust or resampling methods may be preferable when assumptions fail.

2.6 Empirical and simulated distributions

Historical simulation uses observed outcomes without imposing a specific parametric distribution. Monte Carlo simulation generates outcomes from an assumed data-generating model. Both methods can represent nonlinear portfolios, but their reliability still depends on data quality and model design.

3. Populations, Samples and Sampling Distributions

A population is the full set of units or outcomes of interest. A sample is the subset actually observed. A parameter, such as μ or p, describes a population; a statistic, such as x̄ or p̂, is calculated from a sample.

A random sample should represent the target population. Convenience samples, rejected-loan records, surviving funds or only currently active customers may create selection bias. Larger samples reduce random sampling error but do not correct systematic bias.

3.1 Sample estimators

For observations x₁,…,x_n, the sample mean is

The unbiased sample variance is

The divisor n−1 corrects the downward bias that would arise from using n when estimating population variance from the sample mean.

An estimator is unbiased if its expected value equals the parameter. Consistency means that it converges to the true parameter as sample size grows. Efficiency concerns the estimator’s variance relative to alternatives.

3.2 Standard error and the central limit theorem

The standard deviation of an estimator’s sampling distribution is its standard error. For an independent and identically distributed sample with finite variance,

When σ is unknown, it is estimated by s÷√n. The central limit theorem states that, under appropriate independence and finite-variance conditions, the standardized sample mean approaches a normal distribution as n increases. There is no universal sample size that guarantees a good approximation; strong skewness, heavy tails or dependence may require much larger samples.

For time-series returns, autocorrelation changes the variance of the sample mean. Therefore, the simple standard-error formula must not be used mechanically when observations are dependent.

4. Estimation and Confidence Intervals

4.1 Point and interval estimation

A point estimate gives one value for a parameter. A confidence interval combines an estimate with a margin of error. If observations are independent and normally distributed, population variance is unknown, and n observations are used, a confidence interval for μ is

Here t_α÷2,n−1 denotes the positive t critical value leaving probability α÷2 in the upper tail, and 1−α is the confidence level. For nonnormal populations, this interval may be approximately valid in sufficiently large samples under central-limit conditions.

Worked example: confidence interval for average monthly return

A sample of 64 monthly returns has x̄=0.006 and s=0.020. Assuming independent observations and that t-based inference is appropriate,

Using t_0.025,63≈1.998,

The 95% confidence interval is approximately 0.0010 to 0.0110, or 0.10% to 1.10% per month.

The correct repeated-sampling interpretation is that 95% of intervals constructed by this method would contain μ if the assumptions held. It is not correct, in frequentist inference, to state that the fixed parameter has a 95% probability of lying in this already-calculated interval.

For a population proportion, the estimated standard error is

The conventional Wald interval p̂±z_α÷2SE(p̂) can perform poorly for small samples or proportions near zero or one. Wilson or exact intervals are often preferable in those settings.

5. Hypothesis Testing

A hypothesis test evaluates how compatible sample evidence is with a specified null hypothesis.

1. Define the parameter and hypotheses before examining results.
2. Choose a significance level α.
3. Select a test statistic whose assumptions fit the data.
4. Calculate the statistic and p-value.
5. Reject or fail to reject the null hypothesis.
6. Report effect size, confidence interval, assumptions and practical implications.

For a mean with unknown population variance,

where μ₀ is the null value. Under independent normal observations and H₀:μ=μ₀, the statistic follows a t distribution with n−1 degrees of freedom.

A p-value is the probability, assuming the null hypothesis and test assumptions are true, of observing a test statistic at least as incompatible with the null as the statistic obtained. It is not the probability that the null hypothesis is true.

5.1 Decision errors and power

A Type I error rejects a true null hypothesis; its designed probability is α. A Type II error fails to reject a false null hypothesis. Test power is 1−β, where β is the Type II error probability for a specified alternative.

Power generally increases with sample size, effect magnitude and measurement precision. Lowering α while holding everything else fixed reduces Type I error but also tends to reduce power.

5.2 Statistical and economic significance

Statistical significance indicates that data provide evidence against a stated null under the model and testing procedure. It does not prove that the null is false, establish causality or demonstrate commercial importance.

A tiny reduction in funding cost can be statistically significant in a very large sample yet economically negligible. Conversely, a financially important effect may fail to reach significance in a small, noisy sample. Analysts should report effect sizes in decision-relevant units—for example, percentage points of default probability, basis points of yield or currency amounts of expected loss.

Testing many strategies inflates the probability of false discoveries. Bonferroni adjustments, false-discovery-rate controls, preregistration and out-of-sample validation can reduce this risk.

6. Covariance and Correlation

Covariance measures how two variables move together:

Its units are the product of the variables’ units. The sample covariance is

Pearson correlation standardizes covariance:

Correlation is unit-free. Values near +1 indicate strong positive linear association, values near −1 indicate strong negative linear association, and values near zero indicate weak linear association. A zero correlation does not rule out a nonlinear relationship.

Correlation does not establish causation. A relationship may reflect reverse causality, omitted variables, common trends or data selection. In portfolio analysis, correlation also need not remain stable during market stress.

For a two-asset portfolio with weights w₁ and w₂, return variances σ₁² and σ₂², and correlation ρ₁₂,

Weights are dimensionless and variance is in squared return units. Diversification benefits arise when correlation is below +1, although they may weaken when correlations rise in crises.

7. Simple Linear Regression

Simple linear regression models the conditional mean of Y as a linear function of X:

Here β₀ is the intercept, β₁ is the slope and ε_i is the unobserved error. In a market model, Y may be an asset return and X a market return. The slope is dimensionless if both are measured in the same return units; the intercept is return per observation period.

Ordinary least squares chooses b₀ and b₁ to minimize

Setting the two first derivatives to zero gives

The fitted value is Ŷ_i=b₀+b₁X_i, and the residual is e_i=Y_i−Ŷ_i. The coefficient of determination is

With an intercept, 0≤R²≤1. A high R² describes in-sample fit; it does not prove causality, correct specification or predictive performance.

7.1 Regression inference and the slope standard error

Under the classical simple-regression assumptions—including linear conditional mean, random sampling or suitable independence, variation in X, exogeneity E(ε_i|X_i)=0, and constant conditional error variance Var(ε_i|X_i)=σ²—the error variance is estimated by

The loss of two degrees of freedom reflects estimation of the intercept and slope. Let

Then the conventional slope standard error is

To test H₀:β₁=β_1,0, use

With normally distributed errors, the statistic has an exact t distribution with n−2 degrees of freedom under the null. Without normality, large-sample inference may be approximate. Heteroskedasticity-robust or autocorrelation-consistent standard errors should be considered when error variance or dependence assumptions fail.

A confidence interval for the slope is

Worked regression example

Suppose market returns X and bank-share returns Y, both in percentage points, are observed for five periods:

Here X̄=0, Ȳ=1.0, S_xx=10 and ∑(X_i−X̄)(Y_i−Ȳ)=10.1. Therefore,

The fitted equation is Ŷ=1.0+1.01X. The residuals are −0.08, 0.21, −0.20, 0.09 and −0.02, giving

Testing H₀:β₁=0 gives

with three degrees of freedom. Using t_0.025,3≈3.182, the 95% confidence interval is

Also, R²=1−0.099÷10.3≈0.990. The sample supplies strong evidence against a zero slope under the model assumptions. Nevertheless, five observations are too few for confident diagnosis of normality, structural stability or out-of-sample performance.

8. Illustrative Case: Evaluating a Bank Credit Scorecard

Background

The following case is fictional and illustrative. Meridian Bank introduces a credit scorecard intended to reduce one-year default rates among approved personal loans. Its historical benchmark default rate is 9%.

Problem

In a validation sample of 600 newly approved loans, 42 default. Management asks whether the observed rate supports the claim that default risk is below 9% and whether the difference is economically important.

Analysis

The estimated default rate is

The hypotheses are H₀:p=0.09 against H₁:p<0.09. Under the null, the one-proportion z statistic is

The one-sided p-value is approximately 0.0435. At the 5% level, the sample provides evidence against the 9% null in favor of a lower rate. It does not prove that the scorecard caused the reduction.

An approximate 95% Wald confidence interval is

A Wilson interval would generally be preferable for formal reporting.

Economic assessment

A two-percentage-point reduction applied to 6,000 comparable loans would correspond to 120 fewer defaults. If average exposure at default is $20,000 and loss given default is 45%, expected loss per default is $9,000. The potential annual reduction in expected credit loss is therefore

This calculation assumes unchanged loan volume, exposure and loss severity.

Approach and outcome

The bank should compare contemporaneous treatment and control groups, examine borrower mix, approval rates and macroeconomic conditions, and validate performance out of time. It should also test whether protected groups experience unjustified adverse outcomes. The initial evidence is statistically suggestive and potentially economically material, but causal and governance conclusions require stronger validation.

Learning points

• A p-value must be interpreted conditionally on assumptions.
• Confidence intervals communicate estimation uncertainty.
• Economic significance requires decision-relevant units.
• Model validation must consider selection, fairness and changing conditions.

9. Common Analytical Errors and Best Practice

Frequent errors in financial statistics include:

• Look-ahead bias: using information unavailable at the decision date.
• Survivorship bias: excluding failed institutions, funds or borrowers.
• Data snooping: repeatedly testing models until a favorable result appears.
• Omitted-variable bias: excluding a factor related to both X and Y.
• Spurious regression: interpreting common trends as meaningful relationships.
• Outlier sensitivity: allowing a few observations to dominate means, correlations or regressions.
• Nonstationarity: assuming historical parameters remain constant through structural change.
• Dependence: treating clustered defaults or serially correlated returns as independent.
• Heteroskedasticity: using conventional standard errors when error variance changes.
• Significance misuse: treating failure to reject as proof of no effect or treating significance as proof of causation.

Best practice includes defining the target population, preserving time order, documenting data transformations, checking assumptions, reporting confidence intervals and effect sizes, using robust standard errors where justified, validating out of sample and conducting sensitivity analysis.

10. Governance, Ethics and Contemporary Practice

Financial statistical models affect credit access, pricing, capital allocation and insurance coverage. Institutions should maintain data lineage, access controls, reproducible code, independent validation and documented model limitations. Privacy and fair-lending requirements apply when personal or proxy variables are used.

Machine-learning systems can capture nonlinear relationships, but prediction accuracy does not eliminate sampling bias, instability or discrimination. Explainability tools should supplement—not replace—sound experimental design and statistical inference. Contemporary practice increasingly uses bootstrap intervals, robust covariance estimators, Bayesian updating, stress testing and model monitoring. These tools improve analysis only when their assumptions and limitations are clearly disclosed.

Chapter Summary

Statistical foundations enable finance professionals to quantify uncertainty rather than conceal it. Probability distributions model defaults, claims, prices and returns, but their assumptions must match the financial process. Sampling distributions and standard errors explain why estimates vary across samples, while confidence intervals communicate plausible parameter ranges under a specified procedure.

Hypothesis tests evaluate evidence against null hypotheses but neither prove hypotheses nor establish economic importance. Covariance and correlation measure co-movement, while regression estimates conditional relationships. Reliable regression inference requires an appropriate slope standard error, t-test and careful assessment of exogeneity, variance and dependence assumptions. Professional analysis combines statistical evidence with effect sizes, financial consequences, validation, governance and ethical scrutiny.

Key Terms

Bernoulli distribution: A two-outcome probability distribution.

Binomial distribution: A distribution for the number of successes in independent Bernoulli trials with constant probability.

Central limit theorem: A result describing the approximate normality of suitably standardized sample means under stated conditions.

Confidence interval: A range generated by a procedure designed to cover a parameter at a specified long-run rate.

Correlation: A unit-free measure of linear association between two variables.

Covariance: A measure of joint variation between two variables.

Economic significance: The practical or financial importance of an estimated effect.

Estimator: A sample-based rule used to estimate a population parameter.

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

Heteroskedasticity: Nonconstant conditional error variance.

Hypothesis test: A procedure for evaluating evidence against a null hypothesis.

Lognormal distribution: A positive distribution whose logarithm is normally distributed.

Ordinary least squares: A method that minimizes the sum of squared regression residuals.

P-value: The probability, under the null and model assumptions, of a result at least as incompatible with the null as the observed result.

Regression slope: The estimated change in conditional mean Y associated with a one-unit increase in X.

Sampling distribution: The probability distribution of a statistic over repeated samples.

Standard error: The standard deviation of an estimator’s sampling distribution.

Statistical significance: Evidence against a null hypothesis at a specified significance level under the testing assumptions.

Type I error: Rejection of a true null hypothesis.

Type II error: Failure to reject a false null hypothesis.

Review Questions

Short Answer Questions

1. Distinguish a population parameter from a sample statistic.
2. State the assumptions of a binomial distribution.
3. Explain the meaning of a standard error.
4. Why does correlation not establish causation?
5. Distinguish statistical significance from economic significance.

Long Answer Questions

1. Explain how normal, lognormal, binomial and Poisson distributions are applied in finance, including their limitations.
2. Discuss the central limit theorem and the consequences of dependence for financial sampling distributions.
3. Explain the complete hypothesis-testing process, including Type I error, Type II error and power.
4. Derive and interpret the ordinary least-squares slope and its conventional standard error.
5. Evaluate the governance and ethical responsibilities associated with statistical credit models.

Analytical / Application Questions

1. A portfolio has 100 independent loans, each with a 3% default probability. Calculate the expected number and standard deviation of defaults, and discuss why actual loan defaults may violate binomial assumptions.
2. A sample of 36 monthly returns has a mean of 0.9% and a standard deviation of 3.0%. Construct an appropriate 95% confidence interval using a supplied critical value of 2.030, and interpret it.
3. Design a validation plan for a new insurance-pricing model that addresses sampling bias, statistical significance, economic significance, fairness and out-of-sample performance.

Multiple Choice Questions with Answers

Question 1

An investment returns 5% with probability 0.6 and −2% with probability 0.4. What is its expected return?

A. 1.4%

B. 2.2%

C. 3.0%

D. 4.2%

Correct Answer: B. 2.2%

Question 2

Which condition is required for a binomial model?

A. The event probability changes after every trial.

B. Each trial has at least three possible outcomes.

C. Trials are independent with a constant event probability.

D. The number of trials is itself normally distributed.

Correct Answer: C. Trials are independent with a constant event probability.

Question 3

For an independent sample with n=100 and sample standard deviation s=0.20, what is the estimated standard error of the sample mean?

A. 0.002

B. 0.020

C. 0.100

D. 2.000

Correct Answer: B. 0.020

Question 4

A sample has x̄=8, s=10 and n=25. Using a t critical value of 2.064, what is the 95% confidence interval for the population mean?

A. [3.872, 12.128]

B. [5.936, 10.064]

C. [−12.640, 28.640]

D. [7.587, 8.413]

Correct Answer: A. [3.872, 12.128]

Question 5

What is a Type I error?

A. Failing to reject a false null hypothesis

B. Rejecting a true null hypothesis

C. Accepting every alternative hypothesis

D. Using a sample that is larger than necessary

Correct Answer: B. Rejecting a true null hypothesis

Question 6

Which statement correctly defines a p-value?

A. It is the probability that the null hypothesis is true.

B. It is the probability that the research hypothesis is false.

C. It is the probability, under the null and test assumptions, of a result at least as incompatible with the null as the observed result.

D. It is the proportion of observations explained by a regression.

Correct Answer: C. It is the probability, under the null and test assumptions, of a result at least as incompatible with the null as the observed result.

Question 7

Which conclusion is valid when Pearson correlation is zero?

A. The variables are statistically independent.

B. The variables have no possible relationship.

C. There is no linear association in the measured population or sample, although a nonlinear relationship may exist.

D. One variable cannot cause the other under any circumstances.

Correct Answer: C. There is no linear association in the measured population or sample, although a nonlinear relationship may exist.

Question 8

In simple linear regression, which expression is the conventional slope standard error under homoskedastic errors?

A. SSE÷S_xx

B. s÷√S_xx

C. s√S_xx

D. √SSE÷n

Correct Answer: B. s÷√S_xx

Question 9

A regression slope is statistically significant at the 5% level in a correctly specified test. Which interpretation is most appropriate?

A. The data provide evidence against the tested null value under the model assumptions.

B. The slope is proven to be the true causal effect.

C. The slope must be economically important.

D. The fitted model must predict accurately out of sample.

Correct Answer: A. The data provide evidence against the tested null value under the model assumptions.

Question 10

A bank expects a credit policy to reduce defaults by 2 percentage points across 6,000 loans. How many fewer defaults does this represent, assuming the rates apply as expected?

A. 12

B. 30

C. 120

D. 300

Correct Answer: C. 120