Probability, Statistics and Risk Variables

Introduction

Insurance is an institution for managing uncertain financial consequences. An insurer cannot know in advance which policyholders will claim, when claims will occur, or how large they will be. Actuarial mathematics converts this uncertainty into quantitative models used for pricing, reserving, underwriting, solvency assessment and risk transfer.

Probability provides the language for uncertain events. Statistics supplies methods for learning about uncertainty from data. Random variables describe uncertain quantities such as claim frequency, claim severity, mortality, lapse rates and aggregate losses. A sound actuarial analysis must also distinguish three sources of uncertainty:

• Process uncertainty: randomness inherent in the future outcome, even if the model and parameters are known.
• Parameter uncertainty: uncertainty about the numerical values governing the model.
• Model uncertainty: uncertainty about whether the chosen probability model adequately represents reality.

These distinctions are central to professional judgment. A large insurer may have extensive data and sophisticated software, yet still face material risk if its assumptions about dependence, tail behaviour or inflation are inappropriate.

Learning objectives

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

1. Apply probability axioms, union rules, conditional probability and Bayes’ theorem to insurance events.
2. Define random variables, probability distributions, expectation, variance and covariance.
3. Use common discrete and continuous distributions in insurance applications.
4. Explain dependence, conditional modelling and aggregate risk.
5. Estimate unknown parameters and assess sampling uncertainty.
6. Distinguish process, parameter and model uncertainty in actuarial decisions.
7. Interpret statistical outputs critically rather than treating a fitted model as reality.

1. Probability foundations

Let Ω denote the sample space, the set of all possible outcomes, and let A, B ⊆ Ω denote events. A probability measure P assigns a number between 0 and 1 to each event and satisfies the following axioms:

The last equality applies when the events Aᵢ are mutually exclusive. From these axioms follow several important laws:

where Aᶜ is the complement of A, and

The intersection term must be subtracted because outcomes in both A and B would otherwise be counted twice. If A and B are mutually exclusive, P(A ∩ B) = 0.

Conditional probability and independence

The conditional probability of A given B, provided P(B) > 0, is

Rearranging gives the multiplication rule:

Events A and B are independent when

Equivalently, P(A | B) = P(A), when the conditional probability is defined. Independence is a strong assumption. For example, two claims may be dependent because they arise from the same flood, pandemic, economic recession or geographic concentration. A portfolio can therefore contain many policies without containing fully independent risks.

Bayes’ theorem

Suppose B₁, B₂, …, Bₖ form a mutually exclusive and exhaustive partition of Ω. Bayes’ theorem is

It updates prior probabilities P(Bⱼ) using observed evidence A. In insurance, the theorem can update the probability that a risk belongs to a high-risk class after observing claims, medical test results, telematics data or suspicious behaviour.

Worked example: underwriting evidence

Suppose 20% of applicants are high risk. A screening indicator is positive for 80% of high-risk applicants and for 10% of low-risk applicants. Let H denote high risk and + denote a positive indicator. Then

P(H | +) = (0.80 × 0.20) ÷ [(0.80 × 0.20) + (0.10 × 0.80)]

= 0.16 ÷ 0.24 = 0.667.

Thus, a positive indicator implies a 66.7% probability of high risk, not certainty. The result demonstrates the importance of the base rate: evidence must be interpreted together with the prior composition of the portfolio.

2. Random variables and distributions

A random variable X assigns a numerical value to each outcome. In insurance, X may represent the number of claims during a year, the amount of one claim, the time to death, or the total annual loss.

A discrete random variable takes countable values. Its probability mass function (pmf) is p(x) = P(X = x), with

A continuous random variable is described by a probability density function f(x). Probabilities are areas under the density:

For a continuous variable, P(X = x) = 0 for every particular value x. Its cumulative distribution function (cdf) is

For a continuous distribution, F(x) = ∫₋∞ˣ f(t) dt. The survival function is S(x) = P(X > x) = 1 − F(x), subject to the convention used at points of discontinuity.

Expectation and variance

The expected value is the probability-weighted average. For a discrete X,

and for a continuous X,

These expressions require the expectation to exist; in rigorous terms, E[|X|] must be finite for an ordinary finite mean. For a function g(X), the law of the unconscious statistician gives E[g(X)] by weighting g(x), without first finding the distribution of g(X).

Variance measures dispersion around the mean μ = E[X]:

The standard deviation is σ = √Var(X), in the same units as X. Variance is measured in squared units and should not be confused with a probability.

For constants a and b,

The second relationship shows why a negative scale factor still produces a positive variance.

Insurance interpretation of expectation

If X is an insurer’s claim amount, E[X] is the pure expected claim cost before expenses, profit, capital costs and risk margins. It is not necessarily a fair premium in practice. A premium may be represented conceptually as

The precise premium structure depends on the product, regulatory rules, competition and the insurer’s capital objectives.

3. Important probability distributions

The choice of distribution should follow the risk mechanism and evidence, not convenience alone.

Bernoulli and binomial models

A Bernoulli variable X takes value 1 with probability p and 0 with probability 1 − p. It is useful for whether a policy claims during a period. If n trials are independent and have the same claim probability p, the total N follows a binomial distribution:

The assumptions of fixed n, identical probabilities and independence may be unrealistic for insurance portfolios. Exposure may differ across policies, and claims may be correlated.

Poisson claim frequency

The Poisson distribution is commonly used for claim counts over a fixed exposure period:

Here λ > 0 is the expected number of claims. Its mean and variance are both λ, a property called equidispersion. If observed variance materially exceeds the mean, a negative binomial model, a mixture model or a dependence structure may be more appropriate.

The Poisson process additionally assumes independent increments and a constant event rate. Those assumptions should not be silently extended to catastrophe claims, seasonal events or changing exposure.

Continuous severity models

Claim severities are non-negative and often right-skewed. The exponential model has density

and survival function S(x) = e⁻λˣ. Its memoryless property is often mathematically convenient but may be a poor representation of large-loss tails.

If ln X is normally distributed with mean μ and variance σ², then X is lognormal. This permits positive values and right skewness. The lognormal may describe ordinary claims but can understate extreme catastrophe losses.

The Pareto model is useful for heavy-tailed losses. With scale xₘ > 0 and shape α > 0,

The mean exists only when α > 1, and the variance exists only when α > 2. This condition is professionally important: sample averages can be unstable when the tail is sufficiently heavy.

4. Dependence and aggregate loss

Let N be claim frequency and X₁, X₂, …, Xₙ be severities. Aggregate loss is

This is a random sum. If N is independent of identically distributed severities Xᵢ, with E[N] = m and E[X] = μₓ, then

and

This is the random-sum variance formula. It separates severity variation from frequency variation. For N ~ Poisson(λ), Var(N) = E[N] = λ, so

because Var(X) + [E(X)]² = E[X²].

The independence assumption is not automatically valid. Weather may simultaneously increase the number and size of claims. Inflation may affect all severities. A common latent risk factor can generate dependence even when individual conditional risks are independent.

Covariance and correlation

For random variables X and Y,

Correlation is

when both standard deviations are positive. Correlation lies between −1 and 1 and measures linear association. Zero correlation does not generally imply independence. Dependence in the tails may be especially important for solvency even when ordinary correlation appears modest.

A practical tool is conditional modelling: estimate risk conditional on age, territory, vehicle type, occupation, climate zone or economic conditions. This can improve discrimination, but it also creates governance obligations concerning fairness, privacy and explainability.

5. Estimation from data

A population is the conceptual set of all relevant risks; a sample is the observed subset. A statistic is a function of sample data, while a parameter describes the population or model. For observations X₁, …, Xₙ, the sample mean is

The unbiased sample variance is

The n − 1 denominator corrects, under standard random-sampling assumptions, the downward bias arising from estimating the mean from the same observations.

Maximum likelihood estimation

If observations have density or mass function f(x; θ), the likelihood is

The maximum likelihood estimator θ̂ maximizes L(θ), or equivalently the log-likelihood ℓ(θ) = ln L(θ). For independent Poisson observations N₁, …, Nₙ with common mean λ, the log-likelihood is

Differentiating and setting the derivative to zero gives

Thus the fitted Poisson rate is the observed average count per equal exposure period. With unequal exposure eᵢ, a common rate estimate is total claims divided by total exposure, under the Poisson rate assumption:

Confidence intervals and prediction

A confidence interval quantifies sampling uncertainty about a parameter. For a sufficiently large sample mean with standard error SE(X̄) = s/√n, an approximate 95% interval is

This approximation relies on suitable sampling conditions and, commonly, the central limit theorem. It does not mean there is a 95% probability that a fixed parameter lies in this particular interval under the frequentist interpretation. It means that the procedure has approximately 95% long-run coverage under its assumptions.

Prediction intervals address future observations and are generally wider because they include both parameter estimation uncertainty and future process variability.

6. Three layers of uncertainty

Process uncertainty

Even if θ is known exactly and the model is correct, a future claim count or loss remains random. For example, if annual claims follow Poisson(100), the expected count is 100 but the actual count varies. Process uncertainty is addressed through diversification, capital, reinsurance and risk margins.

Parameter uncertainty

The true rate, severity mean or tail index is unknown and estimated from finite data. Parameter uncertainty can be quantified through confidence intervals, bootstrap methods, Bayesian posterior distributions or scenario analysis. It is especially material for new products, sparse catastrophe data and rapidly changing environments.

Model uncertainty

Model uncertainty concerns structural choices: Poisson versus negative binomial frequency, lognormal versus Pareto severity, independence versus copula dependence, or alternative inflation and mortality assumptions. More data do not automatically eliminate model uncertainty if competing models are all plausible.

A useful governance decomposition is:

Treating model uncertainty as if it were merely parameter uncertainty can create false precision. Model selection should include out-of-sample validation, residual analysis, sensitivity testing and expert review.

7. Worked aggregate-loss example

Suppose an insurer models annual claim count N as Poisson with λ = 4. Each claim has independent severity X with mean £2,000 and variance £1,500² = £2,250,000. Assume frequency and severity are independent.

The expected annual aggregate loss is

E[S] = λE[X] = 4 × £2,000 = £8,000.

The aggregate variance is

Var(S) = λVar(X) + λ[E(X)]²

= 4(£2,250,000) + 4(£2,000)²

= £9,000,000 + £16,000,000

= £25,000,000.

Therefore the aggregate standard deviation is

√£25,000,000 = £5,000.

The result does not say that losses will usually equal £8,000 or that £5,000 is a sufficient capital amount. It describes the mean and dispersion under a particular model. A solvency decision would also consider skewness, extreme quantiles, reinsurance, expenses, dependence, parameter uncertainty and regulatory requirements.

8. Illustrative case: catastrophe exposure and dependence

Background. A regional insurer writes 50,000 household policies in a coastal area. Historical ordinary claims appear approximately independent, but a severe storm can generate thousands of claims in a short period.

Problem. The existing pricing model treats household claims as independent Bernoulli events. Its aggregate simulation produces a modest 99.5% loss quantile. Management questions whether the result understates catastrophe risk.

Analysis. The independence assumption fails because a shared storm affects many policyholders simultaneously. The insurer separates ordinary attritional claims from catastrophe claims and introduces a common event variable. Conditional on no severe storm, claims are relatively dispersed; conditional on a storm, both frequency and severity increase. Historical data are sparse, so the catastrophe frequency and event-size parameters have substantial parameter uncertainty. Climate trends and exposure growth create model uncertainty about whether the historical event catalogue remains representative.

Approach. The insurer:

1. maps policies by location and construction characteristics;
2. uses event scenarios for storm intensity and affected areas;
3. calibrates ordinary claims with exposure-adjusted frequency models;
4. validates severity tails using external engineering and industry data;
5. performs sensitivity tests for event frequency, inflation and geographic correlation;
6. evaluates reinsurance attachment points and aggregate limits;
7. documents assumptions and obtains independent model validation.

Outcome. The revised model produces a higher and more credible extreme-loss estimate. Management increases catastrophe reinsurance, revises capital allocation and improves disclosure of uncertainty. The outcome is not a single “true” number; it is a decision supported by transparent alternative assumptions.

Learning points. Diversification is weaker when risks share a common cause. A model can fit ordinary historical claims and still fail in the tail. Governance must address process, parameter and model uncertainty separately.

9. Professional practice, ethics and contemporary methods

Actuarial models influence premiums, access to insurance, claims investigation and capital allocation. Statistical accuracy does not by itself establish ethical acceptability. Analysts should assess whether variables are lawful, relevant, proportionate and non-discriminatory. Proxy variables may reproduce protected characteristics even when those characteristics are excluded directly. Data protection, consent, retention, cybersecurity and explainability are therefore part of model governance.

Machine-learning methods can improve prediction through flexible nonlinear relationships and interaction detection. However, predictive performance must be evaluated out of sample, with attention to calibration, stability, interpretability, fairness and drift. A highly accurate model trained on biased historical claims data may institutionalise past inequities. Artificial intelligence should support—not replace—actuarial judgment, documentation and accountability.

Best practice includes:

• define the exposure, observation period and units before modelling;
• distinguish claim frequency from claim severity;
• inspect missingness, censoring, truncation and reporting delays;
• use exposure offsets where appropriate;
• check calibration and residuals, not only overall fit;
• compare plausible alternative distributions;
• test dependence and tail sensitivity;
• separate development, validation and test data where feasible;
• communicate uncertainty intervals and scenario ranges;
• maintain version control, change logs and independent review.

Chapter Summary

Probability laws provide the foundation for combining insurance events and updating risk information. Random variables and their distributions represent claim counts, severities and aggregate losses. Expectation describes average financial cost, while variance and covariance describe dispersion and dependence. Bernoulli, binomial, Poisson, exponential, lognormal and Pareto models are useful but depend on explicit assumptions.

Aggregate loss is a random sum whose mean and variance depend on both frequency and severity. Dependence, especially common-cause catastrophe dependence, can materially increase tail risk. Statistical estimation uses sample information to infer unknown parameters, while confidence intervals describe sampling uncertainty. Actuarial analysis must distinguish process uncertainty from parameter and model uncertainty. Effective professional practice combines mathematical modelling, diagnostics, scenario analysis, ethical governance and clear communication.

Key Terms

Probability: A numerical measure of the likelihood of an event.

Sample space: The set of all possible outcomes.

Conditional probability: Probability of an event given that another event has occurred.

Independence: A relationship in which joint probability equals the product of marginal probabilities.

Bayes’ theorem: A rule for updating probabilities using observed evidence.

Random variable: A numerical mapping from outcomes to values.

Probability mass function: Probabilities assigned to values of a discrete random variable.

Probability density function: A function whose integral gives probabilities for a continuous variable.

Cumulative distribution function: P(X ≤ x), the probability that a variable does not exceed x.

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

Variance: The expected squared deviation from the mean.

Covariance: A measure of joint directional variation between two variables.

Correlation: Standardised covariance measuring linear association.

Frequency: The number of claims or events in a specified period.

Severity: The monetary amount of an individual claim.

Aggregate loss: The total of all claim amounts in a period.

Maximum likelihood: A method estimating parameters that make observed data most plausible.

Process uncertainty: Random variation in outcomes under a known model and parameters.

Parameter uncertainty: Uncertainty about numerical model parameters.

Model uncertainty: Uncertainty about the structure or assumptions of the model.

Review Questions

5 Short Answer Questions

1. State the addition rule for two events and explain why their intersection is subtracted.
2. Define conditional probability and independence.
3. Distinguish a probability mass function from a probability density function.
4. What do expected value and variance measure in an insurance context?
5. Define process, parameter and model uncertainty.

5 Long Answer Questions

1. Explain Bayes’ theorem and illustrate its use in underwriting classification.
2. Compare the Poisson, binomial and negative binomial approaches to claim frequency.
3. Explain why dependence is important in catastrophe insurance portfolios.
4. Derive and interpret the mean and variance of a compound Poisson aggregate loss.
5. Discuss the limitations of maximum likelihood estimation in a sparse, heavy-tailed claims dataset.

3 Analytical / Application Questions

1. An insurer observes 120 claims over 10,000 policy-years. Under a Poisson rate model, estimate the annual claim rate per policy-year and explain what additional uncertainty should be considered.
2. A portfolio has expected claim frequency 3 and expected severity £4,000. Explain why the expected aggregate loss alone is insufficient for setting capital or reinsurance limits.
3. Design a model-governance plan for an AI underwriting model. Identify how you would test data quality, calibration, dependence, fairness, parameter uncertainty and model uncertainty.

Multiple Choice Questions with Answers

Question 1

If P(A) = 0.30, P(B) = 0.40, and A and B are mutually exclusive, what is P(A ∪ B)?

A. 0.10

B. 0.30

C. 0.70

D. 1.20

Correct Answer: C. 0.70

Question 2

Using the illustrative screening example, what is P(H | +) when P(H)=0.20, P(+ | H)=0.80 and P(+ | Hᶜ)=0.10?

A. 0.167

B. 0.400

C. 0.667

D. 0.800

Correct Answer: C. 0.667

Question 3

For a Poisson random variable with λ = 6, which statement is correct?

A. Its mean is 6 and its variance is 36.

B. Its mean and variance are both 6.

C. Its mean is 36 and its variance is 6.

D. Its variance is always zero.

Correct Answer: B. Its mean and variance are both 6.

Question 4

If Var(X) = 9, what is Var(5 − 2X)?

A. −18

B. 9

C. 18

D. 36

Correct Answer: D. 36

Question 5

Under an independent compound Poisson model with λ = 4 and E[X] = £2,000, what is E[S]?

A. £500

B. £2,004

C. £8,000

D. £16,000

Correct Answer: C. £8,000

Question 6

Which condition is required for the Pareto mean to be finite?

A. α ≤ 0

B. α > 1

C. α = 0

D. α < 1

Correct Answer: B. α > 1

Question 7

What does zero correlation generally imply?

A. Independence in every probability model

B. Perfect positive dependence

C. No linear association, but not necessarily independence

D. Equal means and variances

Correct Answer: C. No linear association, but not necessarily independence

Question 8

For equal-exposure independent Poisson observations, the maximum likelihood estimate of the common rate is the:

A. Sample median

B. Sample maximum

C. Sample variance

D. Sample mean count

Correct Answer: D. Sample mean count

Question 9

Which example best represents model uncertainty?

A. Whether a particular policyholder claims this year

B. The estimated value of a claim rate from finite data

C. Choosing between a Poisson and a negative binomial frequency model

D. Random variation in a known severity distribution

Correct Answer: C. Choosing between a Poisson and a negative binomial frequency model

Question 10

In an aggregate-loss model, a common severe storm affecting many policies primarily challenges the assumption of:

A. Positive claim amounts

B. Dependence between risks

C. A finite sample size

D. The definition of expectation

Correct Answer: B. Dependence between risks