Stochastic Modelling, Dependence and Catastrophe Risk

Introduction

Insurance losses are uncertain in their timing, frequency, severity and interaction. For ordinary portfolios, diversification may make aggregate experience relatively stable. Catastrophes challenge this principle because one event can affect thousands of policies, multiple business lines and geographically dispersed assets simultaneously. Dependence can therefore dominate marginal loss behaviour.

Stochastic models translate uncertainty into probability distributions that support pricing, reserving, capital assessment, reinsurance design and enterprise risk management. Their outputs are conditional on assumptions about hazards, exposure, vulnerability, coverage and future conditions. Consequently, process uncertainty, parameter uncertainty and model uncertainty must be distinguished rather than compressed into a single estimate.

This chapter develops frequency–severity models, dependent aggregate-loss models, copulas, catastrophe modelling, extreme-value methods, simulation and uncertainty analysis. Particular attention is given to the nonlinear application of insurance and reinsurance terms.

Learning Objectives

After studying this chapter, readers should be able to:

• construct and interpret stochastic frequency–severity models;
• derive moments of independent and dependent aggregate losses;
• distinguish correlation from tail dependence;
• use copulas to model multivariate insurance losses;
• explain the hazard–exposure–vulnerability–financial modelling chain;
• apply deductibles, limits and reinsurance layers correctly;
• estimate catastrophe risk using exceedance probabilities, VaR and TVaR;
• simulate dependent aggregate losses;
• incorporate process, parameter and model uncertainty; and
• evaluate governance, climate, data and ethical issues in catastrophe analytics.

1. Foundations of Stochastic Insurance Modelling

A stochastic model describes uncertain outcomes through random variables and probability laws. Let N be the number of losses in one year and let X_i denote the positive severity of loss i, measured in currency units. Aggregate annual loss is

with S = 0 when N = 0. This is a collective risk model. In its classical form:

1. X₁,X₂,… are independent and identically distributed;
2. severities are independent of N; and
3. parameters remain fixed during the modelled period.

These assumptions are useful but often unsuitable for catastrophe portfolios, where one event generates related losses and frequency may vary with climate or exposure conditions.

1.1 Sources of uncertainty

Separating these sources is important because conditional simulation using fixed fitted parameters captures process uncertainty but not necessarily parameter or model uncertainty.

2. Frequency, Severity and Aggregate Loss

2.1 Frequency models

The Poisson model is commonly used when events occur independently at a constant rate λ > 0:

Here E[N]=Var(N)=λ. The equality of mean and variance is restrictive. Overdispersion may arise from unobserved heterogeneity, contagion or changing hazard rates. Alternatives include:

• the negative binomial distribution for overdispersed counts;
• a mixed Poisson or Cox process with a random intensity;
• nonhomogeneous Poisson processes with time-varying intensity;
• Hawkes processes when an event temporarily increases subsequent intensity.

A catastrophe event count must be defined consistently. For example, a meteorological event may generate many insurance claims but remains one catastrophe occurrence.

2.2 Severity models

Positive and right-skewed distributions used for claim severity include gamma, Weibull, lognormal, Pareto and generalized Pareto distributions. Selection should consider tail fit, deductibles, policy limits and the purpose of the model. A distribution fitting ordinary claims well may still underestimate regulatory capital if its upper tail is too light.

For a lognormal severity, ln X ∼ N(μ,σ²), where μ and σ refer to log-losses. Its first two moments are

These expressions apply only when the logarithm of loss is normally distributed.

2.3 Moments of a compound aggregate loss

Conditional on N=n, independence gives

Using iterated expectation,

The law of total variance gives

For N∼Poisson(λ), this simplifies to

The simplification is a special property of Poisson frequency under frequency–severity independence.

Worked Example 1: Compound Poisson loss

Suppose annual claim count is Poisson with λ=3. Severity has mean £2 million and standard deviation £3 million.

Because E[X²]=3²+2²=13 in squared millions,

The standard deviation exceeds the mean because severity is highly variable.

3. Dependence in Insurance Portfolios

Dependence arises through common hazards, economic inflation, supply-chain disruption, litigation environments and policyholder behaviour. Zero linear correlation does not imply independence, while the same correlation can correspond to markedly different joint tails.

For losses S₁,…,S_m, total portfolio loss is T=∑_j=1^mS_j. Its variance is

Ignoring positive covariance understates volatility and may overstate diversification.

3.1 Common-event compound model

Suppose N∼Poisson(λ) catastrophe events occur. Event r causes vector loss (X_r1,…,X_rm), with independent and identically distributed event vectors that are independent of N. Annual loss in segment j is

For two segments, conditional covariance is

The covariance of conditional means is

Therefore,

For Poisson N,

This derivation reveals two dependence channels: losses are related within an event, and both segments share the random event count. Even conditionally independent event severities can produce positively dependent annual totals through common frequency.

Worked Example 2: Dependent aggregate losses

A common-event model has λ=0.4 per year. Per-event losses in two regions have means £20 million and £10 million, standard deviations £30 million and £15 million, and correlation 0.6.

The event covariance is 0.6×30×15=270 in squared millions. Thus,

Consequently,

The segment variances are

Hence,

Ignoring covariance would give a standard deviation of only √650=£25.495 million, a material understatement.

4. Copulas and Tail Dependence

By Sklar’s theorem, if a multivariate distribution H has marginal distribution functions F₁,…,F_m, then a copula C exists such that

If the margins are continuous, the copula is unique. Copulas separate marginal behaviour from dependence, allowing heavy-tailed insurance margins to be combined with a chosen dependence structure.

4.1 Common copula families

For continuous variables X and Y, upper-tail dependence is

when the limit exists. A positive value means that extreme losses remain associated at increasingly high thresholds.

For a bivariate Student-t copula with correlation ρ, degrees of freedom ν and -1<ρ<1,

where t_ν+1 is the cumulative distribution function of a standard Student-t variable with ν+1 degrees of freedom. Unlike a Gaussian copula with |ρ|<1, the t-copula has positive asymptotic tail dependence.

Copula choice should not be based solely on in-sample likelihood. Tail diagnostics, stress tests, parameter stability and economic plausibility are essential, especially because catastrophe datasets contain few truly extreme observations.

5. Architecture of a Catastrophe Model

A catastrophe model combines scientific and financial components. Its principal modules are hazard, exposure, vulnerability and financial loss.

Figure: Catastrophe modelling modules and validation feedback.

5.1 Hazard module

The hazard module generates event location, frequency, footprint and intensity. Depending on the peril, intensity may be wind speed, flood depth, ground acceleration or wildfire heat. Event sets may be historical, stochastic or hybrid. Spatial correlation in intensity is crucial because it determines how many exposures are affected together.

5.2 Exposure module

Exposure data identify location, insured value, occupancy, construction, height, age and other characteristics. Geocoding errors and outdated values can materially distort results. Replacement-cost inflation must not be confused with general consumer-price inflation.

5.3 Vulnerability module

A vulnerability function relates hazard intensity h and asset characteristics z to a damage ratio D, often through a conditional distribution:

If asset replacement value is V, the physical or ground-up loss is X=VD. Secondary uncertainty represents variation in damage among otherwise similar assets exposed to the same intensity.

5.4 Financial module: separating ground-up and insured loss

For event e and policy j, let X_je be ground-up physical loss before deductibles, limits, coinsurance and reinsurance. Portfolio ground-up loss is

It is academically incorrect to call a quantity ground-up loss after multiplying it by coverage or reinsurance adjustments. Financial terms are generally nonlinear and must be applied through contractual functions.

For a simple per-occurrence policy with deductible d_j, payout limit u_j and insurer share q_j, one possible indemnity function is

where (y)_+=max(y,0). This formula applies only to the stated contract: the limit caps payout after the deductible. Franchise deductibles, aggregate deductibles, sublimits and proportional limits require different functions.

Gross insured event loss before reinsurance may then be

where F_e applies any event-level deductible, limit or aggregation clause. Reinsurance is applied afterward according to treaty wording.

For an occurrence layer with attachment A and limit L, recovery is

The cedant’s net event loss, ignoring reinstatement premiums and other recoveries, is B_e-R(B_e). Annual aggregate covers require losses to be accumulated across qualifying events and cannot ordinarily be represented by an event-level multiplier.

5.5 Deductibles and layer wording

A £40 million layer attaching at £60 million pays

A full £40 million layer shifted to attach at £70 million pays

If instead a £10 million deductible erodes recovery under the original £40 million layer, the payout is

Here R₃(B)=min((B-£70 million)_+,£30 million). Its maximum recovery is £30 million, not £40 million. Thus it is not equivalent to a £40 million layer attaching at £70 million. For B=£120 million, R₂=£40 million but R₃=£30 million. Contract wording determines which interpretation applies.

6. Catastrophe Risk Measures

Let S be annual loss. The annual exceedance probability curve is

The occurrence exceedance probability is based on the maximum event loss M during the year:

A return period of T years conventionally corresponds to annual exceedance probability 1/T, assuming a stable annual probability model. It does not mean that events occur regularly every T years.

Expected annual loss satisfies the tail-integral identity

for a nonnegative loss with finite mean.

Value at Risk at confidence level p is

A general quantile-based definition of Tail Value at Risk is

provided the relevant tail expectation is finite. For a continuous distribution, this commonly equals E[S|S>VaR_p(S)]. TVaR reflects losses beyond the quantile and is subadditive under standard integrability conditions; VaR need not be subadditive.

7. Extreme-Value Modelling

Extreme value theory focuses on tail behaviour rather than the whole distribution. Under broad regularity conditions, exceedances over a sufficiently high threshold u can be approximated by a generalized Pareto distribution (GPD):

where β>0, and 1+ξy/β>0. The parameter ξ controls tail heaviness. When ξ=0, the limiting form is exponential.

Threshold selection creates a bias–variance trade-off. A low threshold may violate the tail approximation; a high threshold leaves too little data. Catastrophe observations may also be dependent, nonstationary or truncated by reporting and policy limits. Declustering or event-based modelling may therefore be necessary before applying standard extreme-value methods.

8. Simulation of Dependent Aggregate Loss

Analytical distributions are rarely available once dependence, policy terms and reinsurance are introduced. Monte Carlo simulation provides a flexible alternative.

8.1 Event-based algorithm

For each simulated year:

1. Draw parameters Θ if parameter uncertainty is included.
2. Draw annual event count conditional on Θ.
3. For each event, simulate its hazard footprint.
4. Generate spatially dependent asset damage, including secondary uncertainty.
5. Calculate ground-up losses X_je.
6. Apply policy terms to obtain I_j(X_je).
7. Apply event and treaty terms in contractual order.
8. Sum net or gross event losses to annual totals.
9. Repeat and estimate means, quantiles and tail measures.

Dependence can be generated using a common-event mechanism, multivariate distributions, copulas, shared latent factors or spatial random fields. Sampling marginal losses independently and correlating annual totals afterward generally fails to reproduce contractual nonlinearities.

8.2 Monte Carlo precision

For M independent simulated annual losses S⁽¹⁾,…,S^(M), the estimator of expected annual loss is

Its estimated standard error is s_S/√M, where s_S is the sample standard deviation. Quantile and TVaR estimation requires substantially more simulation in sparse tails. Importance sampling, stratification or quasi-Monte Carlo methods can improve efficiency, but likelihood weights and dependence structures must remain valid.

9. Parameter Uncertainty and Predictive Risk

Let Θ denote uncertain model parameters. Predictive loss risk integrates over their distribution:

where π(Θ|data) is a posterior distribution in a Bayesian analysis. A frequentist alternative may use bootstrap parameter draws. Plug-in simulation fixes Θ at an estimate and usually produces narrower predictive tails.

The law of total variance separates uncertainty:

The first term is average process variance; the second is parameter variance in the conditional mean.

Worked Example 3: Uncertain catastrophe frequency

Suppose N|Λ∼Poisson(Λ), while Λ has mean 0.4 and variance 0.08. Event severity is independent of Λ, with mean £25 million and standard deviation £40 million.

Conditional on Λ,

using units of £ million and squared £ million. Therefore,

Thus,

A fixed-rate model would report only √890=£29.833 million. The difference may be much greater at high quantiles because parameter mixing can thicken predictive tails.

Parameter uncertainty should be propagated jointly. Independently drawing parameters that are statistically correlated can create implausible scenarios. Model uncertainty can be explored through model ensembles, Bayesian model averaging or structured scenario weights, but weights require transparent justification.

10. Illustrative Case: Reinsurance and Predictive Capital

Background

Northshore Insurance, a fictional composite insurer, writes coastal property business. Management is reviewing catastrophe reinsurance after rapid exposure growth and worsening flood conditions.

Problem

The previous model independently aggregated regional losses and used fixed hazard parameters. It also approximated policy terms with average payout ratios. These practices understated common-event dependence and ignored contractual nonlinearities.

Analysis and approach

The revised model uses a shared stochastic event set, spatial hazard footprints, building-specific vulnerability and policy-level financial terms. Frequency parameters are sampled from their posterior distribution. Alternative climate-conditioned event rates are presented separately rather than blended invisibly into a historical baseline.

For one illustrative event, three policies have ground-up losses of £80 million, £50 million and £20 million. Their deductibles are £5 million, £2 million and £1 million, and payout limits are £50 million, £40 million and £15 million. Assuming full insurer shares, payouts are:

Policy-level insured loss is therefore £105 million. An event-level deductible of £10 million followed by a £90 million cap produces

A £40 million layer attaching at £60 million recovers

so the insurer retains £60 million before other arrangements. This calculation cannot be replicated reliably by multiplying the £150 million ground-up loss by one average coverage factor.

The simulation produces the following model outputs:

These are illustrative outputs, not universal industry benchmarks.

Outcome

Management uses predictive results for capital planning, tests several occurrence and aggregate reinsurance structures, improves geocoding controls and establishes risk limits for coastal concentration. Pricing retains a central expected-loss view but includes explicit catastrophe loadings, cost of capital and reinsurance cost.

Learning points

• Common-event dependence must be modelled before aggregation.
• Ground-up, insured, ceded and net losses are distinct quantities.
• Financial terms must follow contractual order.
• Parameter uncertainty can materially increase tail capital.
• Model output should support, not replace, expert judgment and stress testing.

11. Validation, Governance and Professional Practice

Catastrophe models are difficult to validate because extreme events are rare and historical conditions may not represent the future. Effective validation is therefore multi-layered:

• compare event frequencies and intensities with scientific evidence;
• back-test ordinary and moderate losses while recognising limited tail power;
• reconcile exposure totals and investigate geocoding quality;
• test vulnerability curves against engineering and claims evidence;
• reproduce policy and treaty calculations using unit tests;
• benchmark outputs across model versions and external models;
• perform sensitivity and reverse stress testing;
• document assumptions, limitations and expert adjustments.

Governance should establish model ownership, independent validation, change approval, version control and escalation thresholds. Regulatory frameworks such as risk-based capital and solvency regimes generally require firms to understand model limitations rather than accept vendor output as a black box.

Ethical issues include affordability, withdrawal of coverage from vulnerable communities, discriminatory proxies in granular pricing and communication of uncertain climate projections. Accurate risk classification must be balanced with applicable law, fairness objectives and access to essential protection.

12. Contemporary Developments and Limitations

Climate change challenges stationarity. Hazard frequency, intensity, spatial extent and correlations may change, while adaptation and building standards alter vulnerability. Models should identify their time horizon and distinguish observed trends, scientific scenarios and speculative assumptions.

Machine learning can improve geocoding, image-based damage assessment and vulnerability estimation. However, rare-event data scarcity, distribution shift and poor explainability limit unconstrained use. Physics-informed and hybrid models can combine scientific structure with statistical learning.

Other emerging challenges include interconnected infrastructure, cyber-induced physical losses, wildfire–flood sequences, demand surge and casualty accumulation. These effects create temporal and cross-peril dependence that a single-period copula may not adequately represent.

No catastrophe model provides a uniquely correct loss distribution. Best practice is to use transparent model ensembles, scenario analysis, uncertainty ranges and decision tests. Precision in presentation should never imply certainty in knowledge.

Chapter Summary

Stochastic insurance models combine frequency, severity and dependence to describe aggregate loss. Classical compound models provide tractable moments, but catastrophe portfolios require common-event, spatial and tail dependence. Copulas separate marginal distributions from dependence, although their tail assumptions demand careful validation.

Catastrophe models integrate hazard, exposure and vulnerability to produce ground-up physical loss. Policy and reinsurance terms are then applied through nonlinear contractual functions. Ground-up loss must not include deductibles, limits or reinsurance adjustments. Occurrence layers, shifted attachments and recovery-eroding deductibles are not automatically equivalent.

Simulation permits realistic treatment of dependence and financial terms. Predictive analysis should incorporate parameter uncertainty through posterior or bootstrap draws and distinguish it from process and model uncertainty. AEP, OEP, expected annual loss, VaR and TVaR provide complementary risk perspectives. Sound use ultimately depends on validation, governance, transparent communication and awareness of nonstationarity and social consequences.

Key Terms

Aggregate loss: Total loss accumulated across claims or events during a specified period.

Annual exceedance probability: Probability that annual aggregate loss exceeds a stated amount.

Catastrophe model: Integrated model of hazard, exposure, vulnerability and financial loss.

Copula: Function joining marginal distributions into a multivariate distribution.

Dependence: Statistical association among random outcomes beyond their marginal behaviour.

Event loss: Loss arising from one defined catastrophe occurrence.

Exposure: Assets, persons or values subject to loss.

Ground-up loss: Physical or economic loss before insurance and reinsurance financial terms.

Hazard: Potentially damaging event and its intensity at affected locations.

Occurrence exceedance probability: Probability that the largest event loss in a year exceeds a threshold.

Parameter uncertainty: Uncertainty about unknown model parameters.

Process uncertainty: Random variation conditional on a specified model and parameters.

Return period: Reciprocal of an annual exceedance probability under a stated probability model.

Secondary uncertainty: Conditional variability in damage given hazard intensity and exposure characteristics.

Tail dependence: Persistence of dependence at extreme quantiles.

Tail Value at Risk: Average quantile over the tail beyond a selected confidence level.

Value at Risk: Quantile of the loss distribution at a stated confidence level.

Vulnerability: Relationship between hazard intensity, asset characteristics and damage.

Review Questions

Short Answer Questions

1. Distinguish process uncertainty from parameter uncertainty.
2. Why can zero correlation not establish independence?
3. What is the difference between AEP and OEP?
4. Define ground-up loss and insured loss.
5. Why may a Gaussian copula understate joint catastrophe extremes?

Long Answer Questions

1. Derive the expectation and variance of a compound aggregate-loss model.
2. Explain the modules of a catastrophe model and the uncertainty introduced by each.
3. Critically compare Gaussian, Student-t, Gumbel and vine copulas for insurance applications.
4. Discuss the benefits and limitations of extreme-value theory in catastrophe risk estimation.
5. Develop a governance framework for the validation and use of a vendor catastrophe model.

Analytical / Application Questions

1. Construct a simulation algorithm for a two-region property portfolio exposed to common windstorm events, policy deductibles and occurrence reinsurance.
2. A treaty is described alternatively as £40 million attaching at £70 million and as a £10 million recovery-eroding deductible applied to £40 million attaching at £60 million. Analyse the difference across loss levels and explain the contractual implications.
3. Design an approach for incorporating parameter, model and climate-scenario uncertainty into a 99.5% capital estimate without double-counting uncertainty.

Multiple Choice Questions with Answers

Question 1

For a compound Poisson model with event rate λ and independent severity X, which expression gives aggregate-loss variance?

A. λVar(X)

B. λE[X²]

C. λ²E[X²]

D. E[X]/λ

Correct Answer: B. λE[X²]

Question 2

A Poisson claim count has mean 3. Severity has mean £2 million and standard deviation £3 million. What is the annual aggregate-loss variance?

A. 13 (£ million)²

B. 18 (£ million)²

C. 39 (£ million)²

D. 45 (£ million)²

Correct Answer: C. 39 (£ million)²

Question 3

In a common-event compound Poisson model, which formula gives covariance between two annual segment losses?

A. λE[X₁X₂]

B. λCov(X₁,X₂) in every case

C. Cov(X₁,X₂)/λ

D. E[X₁]E[X₂]

Correct Answer: A. λE[X₁X₂]

Question 4

Which statement about a Gaussian copula with correlation strictly between −1 and 1 is correct?

A. It always has positive upper-tail dependence.

B. It has no asymptotic upper-tail dependence.

C. It can only be used with normal marginal distributions.

D. It implies independence whenever correlation is positive.

Correct Answer: B. It has no asymptotic upper-tail dependence.

Question 5

Which quantity is correctly described as ground-up catastrophe loss?

A. Physical loss before deductibles, limits and reinsurance

B. Policy payout after deductibles and limits

C. Insurer loss after occurrence reinsurance

D. Reinsurance recovery from a catastrophe

Correct Answer: A. Physical loss before deductibles, limits and reinsurance

Question 6

An occurrence layer has a £60 million attachment and a £40 million limit. What is its recovery when subject loss is £90 million?

A. £20 million

B. £30 million

C. £40 million

D. £90 million

Correct Answer: B. £30 million

Question 7

A £10 million deductible erodes recovery under a £40 million layer attaching at £60 million. What is the maximum possible recovery?

A. £10 million

B. £30 million

C. £40 million

D. £50 million

Correct Answer: B. £30 million

Question 8

Which risk measure averages quantiles over the tail above confidence level p?

A. Occurrence exceedance probability

B. Expected annual loss

C. Value at Risk

D. Tail Value at Risk

Correct Answer: D. Tail Value at Risk

Question 9

In Var(S)=E[Var(S|Θ)]+Var(E[S|Θ]), what does the second term represent?

A. Average process variance

B. Sampling error from finite simulation only

C. Variance caused by uncertainty in the conditional mean

D. Expected policy deductible

Correct Answer: C. Variance caused by uncertainty in the conditional mean

Question 10

Which simulation practice best preserves catastrophe dependence and contractual nonlinearities?

A. Simulate each policy independently and correlate final totals afterward.

B. Simulate common events and apply financial terms in contractual order.

C. Replace every deductible and limit with one constant loss multiplier.

D. Simulate only the expected number of events each year.

Correct Answer: B. Simulate common events and apply financial terms in contractual order.