7. General Insurance Loss Models

Introduction and Context

General insurance contracts protect policyholders against uncertain financial losses arising from events such as motor accidents, property damage, liability claims, medical expenses, and natural catastrophes. Unlike life insurance, where the timing and amount of benefit may be relatively structured, general insurance losses are usually irregular in both frequency and severity. An insurer may receive no claims from a policyholder during a period, or several claims of widely different sizes. Actuarial loss models provide a disciplined way to represent this uncertainty and to transform historical data into pricing, reserving, underwriting, capital, and reinsurance decisions.

A useful aggregate-loss model separates the number of claims from the amount of each claim. If N denotes claim frequency and X1, X2, ..., XN denote individual claim severities, aggregate loss is represented by

S = X1 + X2 + ... + XN.

This frequency–severity framework is central to general insurance actuarial work. Frequency models describe how often claims occur; severity models describe the size of a claim conditional on its occurrence. The observed data, however, may be affected by deductibles, policy limits, inflation, reporting delays, censoring, truncation, and changes in exposure. Consequently, selecting a distribution is only one part of the modelling task.

Learning Objectives

After studying this chapter, the reader should be able to:

1. Distinguish claim frequency, claim severity, and aggregate loss models.

2. Explain and apply Poisson and negative binomial frequency models.

3. Use exponential, gamma, lognormal, Pareto, and related severity distributions.

4. Calculate the effects of deductibles, limits, inflation, censoring, and truncation.

5. Estimate parameters using moments, maximum likelihood, and related methods.

6. Evaluate model suitability using diagnostics and professional judgement.

7. Apply loss models to pricing, reserving, reinsurance, and risk management.

8. Recognise governance, data, ethical, and technological issues in actuarial modelling.

The Frequency–Severity Architecture

Let N be the number of claims in a defined exposure period and let Xi be the amount of claim i. A collective risk model writes aggregate loss as S = Σ Xi, with the convention that S = 0 when N = 0. A frequency distribution has probability function pN(n), while a severity distribution has distribution function FX(x), density fX(x), and, where applicable, survival function SX(x) = 1 − FX(x).

The expected aggregate loss is

E[S] = E[N]E[X],

provided frequency and severity are independent and the expectations exist. If N and X are independent, then

Var(S) = E[N]Var(X) + Var(N)[E(X)]².

The first component represents variation in claim amounts when the number of claims is known; the second represents variation in the number of claims. These formulas are useful for first-order pricing and capital analysis, although realistic portfolios may require dependence between frequency and severity. For example, a severe weather event can increase both the number and size of claims.

Important terminology includes exposure, claim count, incurred loss, paid loss, reported claim, ultimate claim, attritional loss, large loss, and catastrophe loss. Exposure may be measured in vehicle-years, policy-years, insured property values, or payroll units. Rates and expected losses should generally be expressed per consistent exposure unit.

Modelling Claim Frequency

The Poisson Model

The Poisson distribution is the standard starting point for claim counts. If N follows a Poisson distribution with parameter λ, then

P(N = n) = e^(−λ) λ^n/n!, n = 0,1,2,...

Both the mean and variance equal λ. The model is appropriate when claims occur independently, the occurrence rate is approximately constant over the observation period, and one claim does not materially alter the probability of another. Its simplicity makes it valuable for tariff analysis, credibility models, and aggregate-loss calculations.

A practical weakness is the equidispersion assumption. Insurance portfolios frequently exhibit overdispersion, where Var(N) exceeds E(N), because policyholders differ in risk, observations are clustered, or unobserved factors influence claim propensity.

The Negative Binomial Model

The negative binomial distribution allows greater variance than the mean. One parameterisation uses mean μ and dispersion parameter α:

E[N] = μ, Var(N) = μ + αμ², α > 0.

As α approaches zero, the distribution approaches a Poisson model. A common interpretation is that Poisson claim counts have heterogeneous rates: conditional on a risk-specific rate Λ, N is Poisson, while Λ follows a gamma distribution. Mixing over Λ produces a negative binomial distribution. This provides a statistical explanation for overdispersion caused by unobserved heterogeneity.

The negative binomial is often preferable for portfolios containing different driver types, geographic areas, or policyholder classes. It should not, however, be adopted automatically: overdispersion may also arise from incorrect exposure measurement, dependence, time trends, or claims inflation.

Other Frequency Considerations

A binomial model may be suitable when there are a fixed number of opportunities for claims, such as a limited number of insured items, although it is less common for annual policy counts. A zero-inflated model may be considered when there are more zero-claim policies than a standard count distribution can explain. Hurdle models separately describe whether any claim occurs and, conditional on a positive outcome, how many claims occur.

Frequency models may be estimated using grouped count data, individual policy records, or generalized linear models. Under a Poisson generalized linear model, the logarithm of expected frequency is commonly linked to rating variables and exposure through a log link:

log(μi) = log(exposurei) + β0 + β1zi1 + ... + βkzik.

The exposure term is an offset rather than an ordinary explanatory variable. This ensures that a policy with twice the exposure is expected to generate twice the claims, all else equal.

Modelling Claim Severity

Severity distributions must represent non-negative amounts and, in many lines, substantial right-skewness. The mean alone is insufficient: insurers must understand dispersion, quantiles, tail probabilities, and the cost of extreme claims.

Exponential Distribution

The exponential distribution with rate parameter θ has density

f(x) = θe^(−θx), x ≥ 0,

mean 1/θ and variance 1/θ². Its survival function is e^(−θx). The exponential distribution has the memoryless property:

P(X > x + y | X > x) = P(X > y).

This property is mathematically convenient but often unrealistic for insurance losses, because the probability of a very large loss generally depends on the amount already exceeded. The model can nevertheless be useful for simple risks and as a benchmark.

Gamma Distribution

The gamma distribution has shape parameter a and scale parameter b, with density proportional to x^(a−1)e^(−x/b). Its mean is ab and variance is ab². It can represent a range of right-skewed shapes. When the shape is one, it reduces to the exponential distribution. Gamma models are useful for medical, property, and aggregate operational losses where the tail is heavier than exponential but not necessarily extremely heavy.

Lognormal Distribution

If log X is normally distributed with mean μ and variance σ², then X is lognormal. Its mean and variance are

E[X] = exp(μ + σ²/2),

Var(X) = [exp(σ²) − 1] exp(2μ + σ²).

The lognormal distribution is positively skewed and often provides a good empirical description of claim sizes over a central range. It has a lighter asymptotic tail than Pareto, but high variance can still produce large losses. Lognormal parameters are conveniently estimated by applying normal methods to logarithmic claim amounts, subject to appropriate treatment of zeros and policy modifications.

Pareto and Generalised Pareto Distributions

The Pareto distribution is especially important for large-loss modelling. A simple Pareto survival function is

P(X > x) = (x0/x)^α, x ≥ x0,

where α is the tail index. Its mean is finite only when α > 1, and its variance is finite only when α > 2. Thus, tail-index estimation has direct implications for pricing, capital, and reinsurance.

A Pareto model may be fitted above a threshold u using excess losses Y = X − u. The generalised Pareto distribution is widely used in peaks-over-threshold extreme-value modelling. Its shape parameter governs tail heaviness. Threshold selection involves a trade-off: a low threshold provides more observations but may violate the asymptotic approximation; a high threshold improves theoretical relevance but increases estimation uncertainty.

Weibull and Mixture Models

The Weibull distribution can model different hazard-rate patterns and includes the exponential as a special case. Mixture models combine distributions—for example, a body distribution for ordinary claims and a Pareto tail for large claims. Such models often reflect the operational reality that small attritional claims and major losses arise from different mechanisms.

Policy Modifications: Deductibles, Limits, and Coinsurance

Insurance data are frequently observed only after contractual modification. Let X be the ground-up loss.

For an ordinary deductible d, the insurer payment per loss is

Y = (X − d)+ = max(X − d, 0).

The expected payment per loss is the limited expected value

E[(X − d)+] = ∫d∞ S_X(x) dx.

If the analysis is conditional on a claim payment being positive, the payment distribution is a shifted and truncated distribution. The probability that a loss exceeds the deductible is S_X(d), and the expected payment conditional on exceeding d is E[(X − d)+]/S_X(d).

Under a policy limit u and deductible d, payment is

Y = min((X − d)+, u).

Equivalently, if the attachment point is d and the limit is L, payment is min((X − d)+, L). The expected payment can be written as

E[Y] = ∫d^(d+L) S_X(x) dx.

A per-payment deductible differs from an ordinary deductible. Under a franchise deductible, the insurer may pay the entire loss when the deductible is exceeded, creating a discontinuity in the payment function. These distinctions are material in pricing and must be reflected in the data transformation.

Coinsurance, policy limits, sublimits, and recoveries from salvage or subrogation further modify the observed amount. A model fitted to net paid claims cannot be applied directly to ground-up losses without reconstructing the contractual transformation.

Inflation, Censoring, and Truncation

Historical claims should be restated to a common valuation date. If a claim amount x occurred when the price index was It and the target index is I0, an inflation-adjusted amount may be approximated by

x0 = x(I0/It).

The selected index should reflect the loss cost, not merely a general consumer price index. Wage inflation, medical inflation, construction costs, legal costs, and social inflation may behave differently. Future inflation assumptions should be separated from historical development where possible.

Censoring occurs when the exact claim amount is not observed, although information about it is available. For example, a claim may be known to exceed a policy limit. Truncation occurs when observations outside a range are absent from the dataset. Left truncation may result when only claims above a deductible are recorded; right truncation may result when large claims are capped.

Ignoring censoring or truncation usually biases parameter estimates. If claims are observed only when X > d, the likelihood must be conditional on X > d. If an observation is known only to exceed c, its likelihood contribution is S_X(c), not the density evaluated at c. Proper treatment is particularly important when estimating tail parameters.

Parameter Estimation and Model Selection

Method of Moments

The method of moments equates sample moments with theoretical moments. For a Poisson model, the sample mean estimates λ. For a gamma model, sample mean and variance can estimate shape and scale. This method is simple and transparent, but it may be inefficient and sensitive to extreme observations—especially for heavy-tailed severity data.

Maximum Likelihood

Maximum likelihood estimation selects parameters that maximise the probability of the observed data. For independent uncensored observations x1,...,xn, the likelihood is the product of densities. With deductibles, limits, or censoring, the likelihood must use the appropriate payment density, conditional density, or survival probability. Maximum likelihood is flexible and supports standard errors and likelihood-ratio comparisons, but numerical optimisation may be unstable for highly heavy-tailed data.

Regression and Credibility

Generalised linear models relate frequency or severity to rating variables. Frequency often uses Poisson or negative binomial errors with a log link. Severity may use gamma errors with a log link or model log claim amounts using a normal regression. Credibility methods combine individual experience with collective portfolio information, reducing volatility for small portfolios.

Diagnostics

No distribution should be accepted solely because it has a familiar name. Useful diagnostics include empirical and fitted distribution plots, probability–probability plots, quantile–quantile plots, residual analysis, likelihood-based criteria such as AIC, and out-of-sample validation. Tail models require special attention to exceedance plots, threshold stability, extreme quantiles, and sensitivity analysis. A model that fits ordinary claims well may materially understate catastrophe risk.

Aggregate Loss and Simulation

When N is Poisson and severities are independent, the aggregate-loss distribution is compound Poisson. Its mean and variance follow the formulas given earlier. Exact recursive methods may be available for discretised severity distributions, while simulation is often more practical for complex policy conditions, dependence, reinsurance, and multiple claim types.

A Monte Carlo procedure generates a frequency observation, generates that many severity observations, applies deductibles and limits, sums payments, and repeats the process. The resulting simulated distribution supports estimates of expected loss, variance, value at risk, and expected shortfall. Simulation error should be reported, especially for extreme quantiles. Variance-reduction methods, importance sampling, and scenario analysis may improve efficiency for rare events.

Applied Case: Motor Insurance Portfolio

Background

An insurer has 100,000 private motor policies observed for one year. Historical analysis indicates an average of 0.08 reported claims per policy-year. Claim counts show a variance materially greater than their mean. Ground-up claim severities are strongly right-skewed, and the policy has a deductible of 500 monetary units and a payment limit of 50,000.

Situation and Problem

The insurer must estimate the expected annual claim cost per policy and assess whether its existing pricing model understates variability. Claims above 50,000 are recorded only at the limit, and historical amounts span several years with differing repair and medical cost levels.

Analysis

A Poisson model would impose mean–variance equality and is therefore a questionable starting point. A negative binomial model can represent overdispersion, while explanatory variables such as driver age, vehicle type, territory, and mileage can be incorporated through a regression structure.

For severity, ordinary claims may be represented by a lognormal or gamma distribution, with a Pareto or generalised Pareto tail above a selected threshold. All historical claims are first adjusted to a common cost level. Claims subject to the deductible are modelled using the payment distribution rather than the ground-up distribution. Claims recorded at 50,000 are treated as right-censored or capped observations, depending on the available information.

The insurer estimates frequency and severity separately, then simulates annual aggregate payments. Sensitivity tests vary the tail index, inflation assumption, deductible interpretation, and dependence between frequency and severity.

Approach and Outcome

The resulting model produces an expected loss cost based on the mean frequency multiplied by expected policy payment, but the aggregate simulation reveals a much larger upper-tail risk than the previous model. Management therefore separates attritional pricing from large-loss and catastrophe provisions, reviews its excess-of-loss reinsurance, and introduces monitoring thresholds for changes in claim frequency and severity.

Learning Points

The case illustrates that a technically sophisticated distribution cannot correct poor data treatment. Exposure definition, inflation adjustment, policy limits, censoring, and dependence are as important as the choice between lognormal and Pareto models. It also demonstrates why expected cost and capital requirement are different outputs: the latter depends strongly on the full aggregate distribution.

Professional, Governance, and Ethical Considerations

Actuarial models influence premiums, underwriting decisions, claims strategy, and access to insurance. Variables used in pricing should have a defensible relationship with risk and should comply with applicable insurance regulation. Proxy discrimination may arise when apparently neutral variables reproduce protected characteristics. Governance should include documented assumptions, data lineage, validation, independent review, change control, and clear communication of uncertainty.

Common mistakes include treating capped claims as exact observations, ignoring zero inflation, fitting a single distribution to fundamentally different claim types, using nominal historical amounts without adjustment, extrapolating a tail from too few observations, and confusing statistical fit with business validity. Model users should distinguish parameter uncertainty, process uncertainty, and model error.

Digital systems now permit automated data collection, real-time monitoring, telematics, geospatial analysis, and machine-learning-based prediction. Machine learning may improve predictive accuracy, but it does not remove the need for interpretable exposure definitions, fairness testing, calibration, tail validation, and actuarial oversight. Black-box predictions may be unsuitable where explanations are required by regulation or where extreme outcomes are poorly represented in training data.

Emerging Directions and Best Practices

Modern loss modelling increasingly combines classical actuarial distributions with flexible statistical learning, hierarchical models, Bayesian estimation, and extreme-value methods. Climate change, cyber risk, litigation trends, supply-chain disruption, and changing mobility patterns create non-stationarity that historical data alone may not capture. Scenario analysis and forward-looking expert judgement are therefore important complements to statistical estimation.

Sound practice includes:

• defining the unit of exposure and the valuation basis before modelling;
• separating frequency, severity, and aggregate-loss questions;
• adjusting for inflation, development, deductibles, limits, and changes in underwriting;
• comparing several plausible distributions rather than selecting one mechanically;
• validating central and tail behaviour separately;
• documenting assumptions, uncertainty, and limitations;
• using out-of-sample tests and sensitivity analysis; and
• reviewing the model regularly as claims, products, regulation, and risk conditions change.

Chapter Summary

General insurance loss models represent uncertainty through claim frequency, claim severity, and aggregate loss. Poisson models are useful for independent counts with equal mean and variance, while negative binomial models accommodate overdispersion. Exponential, gamma, lognormal, Pareto, Weibull, and mixture distributions provide alternative representations of severity, with Pareto-type models being especially relevant to heavy tails.

Deductibles, limits, coinsurance, inflation, censoring, and truncation transform the observed data and must be incorporated into estimation and pricing. Parameters may be estimated by moments, maximum likelihood, regression, or credibility methods. Model selection requires diagnostic testing, validation, professional judgement, and sensitivity analysis. Aggregate simulations support pricing, reinsurance, capital, and risk management decisions. Ethical governance requires attention to fairness, transparency, data quality, uncertainty, and regulatory standards.

Key Terms

Frequency: Number of claims occurring in a specified exposure period.

Severity: Monetary size of an individual claim.

Aggregate loss: Total loss from all claims in a specified period.

Exposure: Unit measuring the amount of insurance risk observed.

Poisson distribution: Count distribution whose mean and variance are both equal to its rate parameter.

Negative binomial distribution: Count distribution commonly used to model overdispersed claim frequencies.

Overdispersion: Variance exceeding the mean in a count distribution.

Exponential distribution: Continuous severity model with a constant hazard rate and memoryless property.

Gamma distribution: Flexible positive distribution described by shape and scale parameters.

Lognormal distribution: Distribution whose logarithm is normally distributed.

Pareto distribution: Heavy-tailed distribution commonly used for large losses.

Tail index: Parameter describing the heaviness of a Pareto-type tail.

Deductible: Amount of loss retained by the policyholder before insurer payment begins.

Policy limit: Maximum amount payable by the insurer under a coverage.

Censoring: Partial observation of a value, such as knowing only that it exceeds a limit.

Truncation: Exclusion of observations outside a specified range.

Limited expected value: Expected payment after applying a deductible or limit.

Credibility: Method combining individual and collective experience in estimation.

Aggregate simulation: Repeated generation of frequency and severity outcomes to approximate total-loss distributions.

Expected shortfall: Average loss conditional on exceeding a specified risk percentile.

Review Questions

Short Answer Questions

1. What is the difference between claim frequency and claim severity?

2. Why may a negative binomial model be preferred to a Poisson model?

3. State the effect of an ordinary deductible on an individual loss.

4. What is meant by censoring in insurance claims data?

5. Why are Pareto models useful for large claims?

Long Answer Questions

1. Explain the frequency–severity framework and derive the mean and variance of aggregate loss.

2. Compare the Poisson and negative binomial distributions in general insurance applications.

3. Discuss the characteristics and uses of exponential, gamma, lognormal, and Pareto severity models.

4. Explain how deductibles, limits, inflation, censoring, and truncation affect loss modelling.

5. Describe maximum likelihood estimation and diagnostic procedures for selecting a severity model.

Analytical / Application Questions

1. A portfolio has expected claim frequency 0.12 per policy-year and expected severity of 4,000. Calculate expected aggregate loss per policy-year and explain what additional information is needed to assess capital risk.

2. An insurer observes many claims exactly equal to its policy limit. Explain how treating these observations as uncensored exact values could distort tail estimation and propose an appropriate modelling approach.

3. A motor portfolio has claim-count variance three times its mean. Recommend a frequency-modelling strategy, identify possible causes of the overdispersion, and describe validation tests.

22. Multiple Choice Questions with Answers

Question 1

Which frequency model directly allows the variance of claim counts to exceed the mean?

A. Poisson

B. Negative binomial

C. Bernoulli

D. Uniform

Correct Answer: B. Negative binomial

Question 2

If claim frequency and severity are independent, expected aggregate loss equals:

A. E[N] + E[X]

B. E[N]E[X]

C. Var(N)Var(X)

D. E[N]/E[X]

Correct Answer: B. E[N]E[X]

Question 3

For an ordinary deductible d, insurer payment per loss is:

A. X + d

B. min(X,d)

C. (X − d)+

D. X/d

Correct Answer: C. (X − d)+

Question 4

Which distribution is particularly associated with heavy-tailed large losses?

A. Pareto

B. Bernoulli

C. Uniform

D. Binomial

Correct Answer: A. Pareto

Question 5

A claim recorded only as exceeding a policy limit is an example of:

A. Right censoring

B. Left truncation

C. Deflation

D. Zero inflation

Correct Answer: A. Right censoring

Question 6

Under a Poisson model, the mean and variance of claim count are:

A. Always zero

B. Equal

C. Opposite

D. Unrelated

Correct Answer: B. Equal

Question 7

Historical claims are restated to a common valuation date primarily to account for:

A. Deductible selection

B. Inflation

C. Claim frequency

D. Exposure cancellation

Correct Answer: B. Inflation

Question 8

Which method selects parameters by maximising the probability of observed data?

A. Method of moments

B. Maximum likelihood

C. Random sampling

D. Simple averaging

Correct Answer: B. Maximum likelihood

Question 9

If a Pareto tail index is less than or equal to one, the ordinary mean is:

A. Guaranteed to be zero

B. Finite and stable

C. Not finite

D. Equal to the variance

Correct Answer: C. Not finite

Question 10

In aggregate-loss simulation, the usual sequence is to generate:

A. Only the largest claim

B. Frequency, severities, policy payments, and their sum

C. Inflation first and no claims

D. Only expected values

Correct Answer: B. Frequency, severities, policy payments, and their sum