Non-Life Loss and Aggregate Claims Models

Introduction and context

Non-life insurance protects policyholders against uncertain losses arising from motor accidents, property damage, liability, health events, natural catastrophes and many other causes. Unlike life insurance, where timing and amount of benefits may often be specified in advance, non-life claims are usually uncertain both in frequency and severity. An insurer therefore needs models that answer two related questions:

1. How many claims will occur during a specified period?
2. How large will each claim be after policy terms, inflation and reinsurance are considered?

The aggregate loss model combines these two sources of uncertainty. It is fundamental to pricing, reserving, capital adequacy, solvency assessment, reinsurance design and enterprise risk management.

The central aggregate-loss representation is:

Here, N is the number of claims during the period, Xᵢ is the amount of the ith claim, and S is total aggregate loss. The model is conditional on a defined exposure period, coverage basis, currency and loss definition.

Learning objectives

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

• distinguish claim frequency, severity and aggregate-loss models;
• select and interpret common frequency and severity distributions;
• incorporate deductibles, policy limits, inflation and coinsurance;
• derive the mean and variance of a compound aggregate model;
• explain reinsurance transformations and their effect on risk;
• use exact, recursive, approximation and simulation methods;
• evaluate modelling assumptions, data limitations and governance issues;
• interpret aggregate-loss results for underwriting and capital decisions.

1. The architecture of a non-life loss model

A useful model separates the insurance process into layers:

A basic assumption is that claim severities X₁, X₂, … are independent and identically distributed and independent of N. This assumption is convenient but not universally appropriate. For example, a severe storm can produce both unusually high claim frequency and unusually large claims, creating dependence.

The insurer must also distinguish between:

• ground-up loss: the full economic loss before insurance terms;
• reported loss: loss known to the insurer;
• incurred loss: reported payments plus case reserves;
• ultimate loss: the estimated final cost after all development;
• paid loss: cash payments made during the period.

A frequency–severity model is meaningful only when the claim definition is consistent. Splitting one event into several claims, changing reporting thresholds or mixing paid and incurred data can materially distort both distributions.

2. Claim frequency models

Claim frequency is a non-negative integer-valued random variable. The most widely used starting model is the Poisson distribution:

λ is the expected number of claims in the exposure period. For a Poisson variable,

The equality of mean and variance is called equidispersion. It may be unsuitable where policyholders differ substantially in risk or where claims are clustered.

The binomial distribution applies when there are m fixed, approximately independent opportunities for a claim, each with probability p:

Its mean and variance are mp and mp(1−p). It is appropriate only when the number of opportunities is fixed and at most one claim can arise from each opportunity.

The negative binomial distribution is often used for overdispersed counts. One parameterisation has mean μ and variance:

where r is a dispersion parameter. As r increases, the variance approaches the Poisson variance. Other parameterisations exist, so the chosen convention must always be stated.

Frequency models may also include exposure and rating variables. A Poisson generalized linear model commonly uses:

where eᵢ is exposure for observation i, zᵢⱼ are risk characteristics and βⱼ are estimated coefficients. The exposure term is an offset. It ensures that a policy observed for twice as long is not treated as having the same opportunity for claims as a policy observed for half as long.

3. Claim severity models

Severity X is non-negative and usually right-skewed. Important candidate distributions include:

• Exponential: simple, light-tailed and governed by one rate parameter;
• Gamma: flexible positive distribution with shape and scale parameters;
• Lognormal: suitable where multiplicative factors produce skewness;
• Weibull: flexible hazard shape, often used for property or operational losses;
• Pareto: heavy-tailed and useful for large losses;
• Generalized Pareto: commonly applied to losses above a high threshold in extreme-value analysis.

For a lognormal severity, if ln X ∼ N(μ, σ²), then:

and

The parameters μ and σ² are the mean and variance of the logarithm of loss, not of the original loss. This distinction prevents a frequent interpretation error.

For a Pareto Type I distribution with scale θ and tail parameter α, survival function:

The mean exists only when α > 1 and the variance exists only when α > 2. Thus, heavy-tailed models can imply very high uncertainty in capital estimates even when expected loss is finite.

Severity fitting must address deductibles, policy limits, inflation, claim truncation and reporting thresholds. Fitting an ordinary distribution to payments that are already capped or excess-of-deductible generally produces biased parameters.

4. Policy modifications: deductibles, limits and coinsurance

Let X denote ground-up loss.

Ordinary deductible

With an ordinary deductible d, the insurer payment is:

The expected payment is:

The integral identity applies to non-negative continuous losses and is often called the stop-loss expectation. For a discrete loss distribution, summation or direct probability calculation is required.

Policy limit

With a limit u, payment is:

and

Deductible and limit together

For an ordinary deductible d and limit u, where u is the maximum insurer payment, the payment is:

Equivalently, the insurer pays zero below d, pays x − d between d and d + u, and pays u above that point.

Coinsurance

If the insurer pays proportion c, with 0 ≤ c ≤ 1, then:

when no deductible or limit applies. Coinsurance changes both mean and standard deviation by factor c, and variance by c².

Worked example: deductible and limit

Suppose a claim has severity X equal to 4,000, 10,000 or 30,000 currency units with probabilities 0.50, 0.35 and 0.15. The policy has a deductible of 5,000 and a payment limit of 12,000.

The insurer payments are respectively 0, 5,000 and 12,000. Therefore:

The expected ground-up loss is:

The policy structure therefore reduces expected insurer payment from 9,500 to 3,550 currency units per claim. It also removes small losses and caps exposure to large losses.

5. Inflation and trend

Loss inflation may arise from general price inflation, medical-cost inflation, wage growth, litigation, changes in repair technology or social inflation. If a historical loss X is developed by factor g, the adjusted loss is:

If annual inflation is i for t years, compound development is:

For example, a loss of 20,000 units developed for three years at 6% annual inflation becomes 20,000 × 1.06³ = 23,820.32 units. Inflation should be applied consistently with the policy limit, deductible and valuation date. If a deductible is contractual in nominal currency, the treatment differs from a deductible indexed to inflation.

Trend assumptions are uncertain. A model should usually include sensitivity tests rather than presenting one deterministic inflation estimate as fact.

6. Aggregate claims and the compound distribution

Let N be claim count and Xᵢ independent severities. The aggregate loss is:

This is a compound distribution. Conditional on N = n, the mean and variance are:

where μₓ = E[X] and σₓ² = Var(X). Applying the law of total expectation and total variance gives:

The first component is severity risk conditional on frequency; the second is frequency risk.

For Poisson frequency with mean λ, this becomes:

The equality uses E[X²] = Var(X) + E[X]². It is valid only under independence of frequency and severity and identical independent severities.

Probability generating and moment generating functions

For a non-negative integer-valued N, let its probability generating function be G_N(z) = E[zᴺ]. For severity moment generating function M_X(t) = E[eᵗˣ], the aggregate moment generating function is:

For Poisson frequency:

These expressions are useful theoretically and computationally, although a moment generating function may not exist for some heavy-tailed distributions and positive values of t.

7. Computing aggregate distributions

Expected loss and variance are not enough for solvency decisions because insurers often need P(S > s), a high quantile or a tail conditional expectation.

Exact convolution

If severity is discrete with probabilities f_j, the aggregate distribution can be obtained by convolving severity distributions across possible claim counts and mixing over frequency:

This method is transparent but can be computationally expensive for fine-grained amounts.

Recursive methods

Panjer recursion computes compound distributions efficiently for frequency distributions satisfying a suitable recursion, including Poisson and negative binomial cases under standard parameterisations. It requires discretising severity and selecting a sufficiently high maximum loss. Discretisation error, rounding and truncation must be tested.

Approximations

A normal approximation uses:

It may be acceptable for high-frequency, relatively light-tailed portfolios but can assign probability to negative losses and may underestimate tail risk. A translated gamma approximation often better reflects positive skewness. Heavy-tailed or catastrophe portfolios generally require more appropriate methods.

Monte Carlo simulation

A basic simulation algorithm is:

1. Generate a claim count N^(b) from the fitted frequency model.
2. Generate N^(b) severities from the fitted severity model.
3. Apply deductibles, limits, inflation and reinsurance to every simulated claim.
4. Sum the transformed claims to obtain S^(b)
5. Repeat for b = 1, …, B.
6. Estimate probabilities and quantiles from the simulated aggregate values.

The simulated mean is:

Simulation error decreases approximately at rate 1 ÷ √B, so increasing simulations from 10,000 to 40,000 roughly halves ordinary Monte Carlo standard error. Estimating very remote tail probabilities may require importance sampling, stratification, extreme-value methods or other variance-reduction techniques.

8. Reinsurance structures

Reinsurance transforms the insurer’s gross loss into a retained loss and a ceded loss.

For per-risk excess-of-loss cover with retention r and limit l, the reinsurer payment for a claim is:

The insurer’s retained payment, ignoring other policy terms, is:

For aggregate excess-of-loss reinsurance with annual attachment a and limit l, if gross aggregate loss is S, the ceded amount is:

and retained annual loss is S − C. Per-risk and aggregate covers are not interchangeable: the former acts claim by claim, while the latter acts after annual aggregation.

Reinsurance reduces retained tail risk but introduces counterparty credit risk, reinstatement premiums, basis risk, contract ambiguity and dependence on the reinsurer’s interpretation of loss events. The expected ceded loss is only one part of the decision; capital relief, pricing, collateral, claims control and solvency treatment also matter.

9. An illustrative portfolio case

This is a fictional educational case. Meridian Mutual insures 50,000 household policies for one year. Historical data show an average of 0.08 weather claims per policy-year. Claims are moderately overdispersed, so the actuarial team selects a negative binomial frequency model rather than a Poisson model. After adjusting historical claims for two years of 5% annual inflation, the team models individual ground-up severity using a lognormal distribution. Each policy has a 1,000-unit deductible and a 100,000-unit limit.

Problem

Management wants an estimate of expected annual retained loss and the 99.5th percentile for capital planning. It is also considering a per-risk excess-of-loss treaty with retention 50,000 and limit 250,000.

Analysis and approach

Expected annual claim count is:

The severity model is fitted to inflation-adjusted, policy-level data while explicitly accounting for deductible and limit censoring. The team tests goodness of fit using graphical diagnostics, out-of-sample validation and tail comparisons. It then runs a large Monte Carlo simulation. Each iteration samples the annual count, samples severities, applies policy terms, calculates ceded recoveries claim by claim and sums retained losses.

The team performs sensitivity tests for:

• a higher frequency trend;
• a heavier lognormal tail or Pareto tail above a threshold;
• dependence between claim count and severity during major storms;
• inflation one percentage point above the central assumption;
• delayed reporting and changes in claims settlement practice.

Outcome

The results are reported not as a single “true” capital number but as a distribution with confidence intervals and scenario results. Management compares the treaty’s premium and expected recovery with its reduction in the 99.5th percentile, liquidity needs, counterparty exposure and regulatory capital treatment.

Learning points

A credible aggregate model requires more than selecting familiar distributions. Exposure definition, inflation, censoring, dependence, reinsurance wording, parameter uncertainty and model validation can materially alter the result. Tail decisions should be supported by stress testing and expert review.

10. Governance, ethics and contemporary practice

Actuarial models influence premiums, coverage availability, claims decisions and capital allocation. Governance should therefore include:

• documented data lineage and reproducible transformations;
• clear separation of training, validation and monitoring data;
• diagnostics for calibration, overdispersion and tail fit;
• sensitivity analysis and independent model review;
• version control for assumptions and code;
• transparent communication of uncertainty;
• controls against discriminatory or proxy variables;
• protection of policyholder and claims data.

Machine-learning methods can improve frequency segmentation, image-based damage assessment and fraud detection. However, predictive accuracy does not eliminate the need for interpretability, fairness, stability and regulatory compliance. A highly flexible model can learn reporting practices, socioeconomic proxies or temporary catastrophe patterns rather than underlying risk. Actuarial judgment remains necessary for exposure changes, emerging risks and extrapolation beyond observed data.

Contemporary aggregate modelling increasingly combines classical credibility methods, generalized linear models, gradient-boosting techniques, geospatial catastrophe models and extreme-value theory. Climate change creates particular challenges: historical stationarity may fail, hazard frequency and severity may evolve, and correlations across locations may increase. Scenario analysis and forward-looking models are therefore essential complements to purely historical estimation.

Challenges and limitations

The principal limitations are:

1. Sparse tail data: very large losses are rare, so tail estimates are parameter-sensitive.
2. Dependence: catastrophe events can produce correlated frequency and severity.
3. Changing exposure: policy counts, limits, deductibles and risk mix evolve over time.
4. Censoring and truncation: deductibles, limits and reporting thresholds hide parts of the loss distribution.
5. Parameter uncertainty: estimated parameters are themselves random.
6. Operational complexity: claims systems may record events inconsistently.
7. Model risk: a mathematically elegant model can still be inappropriate for the portfolio.

Best practice is to compare alternative models, validate against hold-out periods, examine residuals and tail diagnostics, stress key assumptions, and communicate ranges rather than false precision.

Chapter Summary

Non-life aggregate-loss modelling separates claim frequency from claim severity and recombines them through the compound model S = ∑ᵢ₌₁ᴺ Xᵢ. Poisson, binomial and negative binomial distributions provide common frequency frameworks, while lognormal, gamma, Weibull, Pareto and generalized Pareto distributions describe severity. Deductibles, limits, coinsurance and inflation transform ground-up losses and must be reflected in the fitted model. Under independent and identically distributed severities independent of frequency, E[S] = E[N]E[X] and Var(S) = E[N]Var(X) + Var(N)E[X]². Aggregate distributions may be obtained through convolution, recursion, approximation or simulation. Reinsurance changes retained loss and tail risk but also creates contractual and counterparty considerations. Reliable professional practice requires validation, sensitivity analysis, governance, ethical data use and explicit treatment of dependence and parameter uncertainty.

Key Terms

Frequency: Number of claims during a specified period.

Severity: Monetary amount of an individual claim.

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

Compound distribution: Distribution formed by summing a random number of random severities.

Exposure: Unit measuring the amount of insurance risk observed.

Deductible: Amount borne by the insured before insurer payment begins.

Policy limit: Maximum insurer payment under a policy or coverage.

Coinsurance: Proportion of a covered loss paid by the insurer.

Ground-up loss: Full loss before insurance modifications.

Inflation adjustment: Development of historical amounts to a common valuation basis.

Poisson distribution: Count model with equal mean and variance.

Negative binomial distribution: Flexible count model commonly used for overdispersion.

Lognormal distribution: Positive skewed distribution whose logarithm is normal.

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

Stop-loss expectation: Expected payment above a deductible or attachment point.

Convolution: Operation combining distributions of independent sums.

Monte Carlo simulation: Repeated random sampling used to estimate a distribution or risk measure.

Excess-of-loss reinsurance: Cover responding to losses above a retention.

Value at Risk: A specified quantile of a loss distribution.

Model risk: Risk that a model or its implementation is unsuitable or incorrectly used.

Review Questions

5 Short Answer Questions

1. What are the two principal random components in a frequency–severity model?
2. State the mean and variance of a Poisson claim-count variable.
3. Define an ordinary deductible mathematically.
4. What is the distinction between per-risk and aggregate excess-of-loss reinsurance?
5. Why can a Pareto severity model create substantial capital uncertainty?

5 Long Answer Questions

1. Explain the compound aggregate-loss model and derive its mean and variance using conditional expectation and conditional variance.
2. Compare Poisson and negative binomial frequency models, including the circumstances in which the latter is preferable.
3. Explain how deductibles, limits, coinsurance and inflation alter observed claim severity.
4. Compare convolution, Panjer recursion, approximation and Monte Carlo simulation as aggregate-distribution methods.
5. Discuss the governance and ethical requirements for using advanced analytics in non-life insurance.

3 Analytical / Application Questions

1. A portfolio has expected claim count 2,000, mean payment per claim 3,000 and severity variance 16,000,000. Frequency variance is 2,400. Calculate expected aggregate loss and aggregate variance under the standard independence assumptions.
2. A claim severity is 80,000, the policy deductible is 10,000 and the insurer limit is 50,000. Calculate the insurer payment and explain how the result changes if the limit is interpreted as a ground-up limit rather than a payment limit.
3. Design a simulation study to compare no reinsurance with per-risk excess-of-loss reinsurance. Specify inputs, transformations, output risk measures, validation checks and at least three sensitivity scenarios.

Multiple Choice Questions with Answers

Question 1

Under the standard independent frequency–severity model, which expression gives aggregate loss?

A. S = N + X

B. S = ∑ᵢ₌₁ᴺ Xᵢ

C. S = N ÷ X

D. S = X − N

Correct Answer: B. S = ∑ᵢ₌₁ᴺ Xᵢ

Question 2

For a Poisson claim count with mean λ, which statement is correct?

A. Its variance is zero

B. Its variance is λ²

C. Its variance is λ

D. Its variance is always larger than its mean

Correct Answer: C. Its variance is λ

Question 3

An ordinary deductible d produces insurer payment:

A. min(X,d)

B. (X − d)⁺

C. X + d

D. d − X

Correct Answer: B. (X − d)⁺

Question 4

If annual inflation is 5% for two years, the development factor is:

A. 1.05

B. 1.10 exactly

C. 1.1025

D. 0.9070

Correct Answer: C. 1.1025

Question 5

For independent frequency and severity, aggregate mean 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 6

Which frequency model is especially associated with overdispersion relative to Poisson?

A. Negative binomial

B. Uniform

C. Bernoulli with one trial

D. Exponential

Correct Answer: A. Negative binomial

Question 7

A per-risk excess-of-loss treaty with retention r pays:

A. min(X,r)

B. X + r

C. min((X − r)⁺, l) when a limit l applies

D. r − X for all claims

Correct Answer: C. min((X − r)⁺, l) when a limit l applies

Question 8

If E[N] = 100 and E[X] = 2,000, the expected aggregate loss is:

A. 2,000

B. 50,000

C. 100,000

D. 200,000

Correct Answer: D. 200,000

Question 9

Which method repeatedly samples claim counts and severities to estimate aggregate quantiles?

A. Monte Carlo simulation

B. Simple averaging only

C. Book depreciation

D. Deterministic ranking without sampling

Correct Answer: A. Monte Carlo simulation

Question 10

Why may a normal approximation be unsuitable for catastrophe losses?

A. It cannot calculate a mean

B. It may understate skewed heavy-tail risk and assign negative losses

C. It always overstates frequency

D. It requires no variance estimate

Correct Answer: B. It may understate skewed heavy-tail risk and assign negative losses