Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Property Insurance Frequency and Severity Decomposition

Abstract. Property insurance frequency–severity analysis decomposes the insurer's observed aggregate loss into an exposure denominator, a count of qualifying claims, and an average monetary amount per claim. The decomposition is an exact description of consistently defined historical experience, but its predictive and causal interpretation depends on the generation of events, insured exposures, contract terms, claim reporting, payment development, inflation, and dependence between insured properties. This monograph provides a source-grounded scientific examination of the complete transformation from physical damage to the insurance claim and its ultimate financial consequences. It distinguishes ordinary claim arrivals from catastrophe occurrences, develops elementary and conditional compound-loss mathematics, examines count regressions and positive-loss distributions, and explains deductibles, limits, truncation, censoring, mixed aggregate models, and severe-loss tails. Ratemaking, reserving, reinsurance, financial reporting, and capital analysis are differentiated by their respective valuation questions and time horizons. Publicly documented FEMA, NOAA, EIOPA, NAIC, OECD, IFRS, actuarial standards, and original scholarly sources establish the scope and limitations of empirical interpretation. Hypothetical calculations are explicitly reproduced as illustrations rather than presented as insurer observations. The final synthesis identifies the assumptions under which the decomposition is reliable, the mechanisms producing otherwise indistinguishable observed trends, and the evidence required for a reproducible international comparative study. The manuscript contributes original critical interpretation and mathematical exposition without claiming a new actuarial theory, a proprietary dataset, or independently measured empirical coefficients.

Keywords: Property insurance; claim frequency; claim severity; exposure; loss cost; collective risk theory; Poisson count; mixed Poisson; generalized linear models; negative binomial; lognormal; Gamma; Tweedie; catastrophe occurrence; spatial dependence; deductible; policy limit; censoring; truncation; severe losses; tail uncertainty; claims development; unpaid reserves; property ratemaking; loss inflation; protection gap.

Introduction: Scientific Object, Evidence, and Method

Object of the investigation

Property insurance starts with a physical interest at risk. An insured asset can be struck by fire, wind, water, earthquake, impact, or another peril; the resulting impairment has a physical description, a repair or replacement cost, a contractual coverage status, and an eventual financial settlement. A frequency–severity calculation begins further downstream, after losses have been converted into a chosen set of recorded claims. Its apparent simplicity is valuable because an aggregate loss total can be described by the count and size of those claims. Interpretation requires an explicit account of everything that made the claims observable.

For defined exposures during a defined period, let NN be the number of qualifying claims, XjX_j the final insured cost of each qualifying claim, and S=∑j=1NXjS=\sum_{j=1}^{N}X_j the aggregate. Where N>0N>0, observed claim frequency is N/EN/E for an agreed exposure denominator EE, and observed positive-claim severity is S/NS/N. Their product is S/ES/E. The identity describes a measured portfolio precisely. It does not establish why observed frequency or severity moved from one year to the next, nor whether a loss from one property is statistically independent of losses from another.

The principal scientific questions follow. What is the correct population and denominator for a property claim count? Which loss distributions and censoring mechanisms follow from the policy and the reporting process? How do models of frequency and severity contribute differently to future rates, unsettled claims and catastrophe accumulation? Which primary sources and independent measurements can separate physical hazard changes from changes in insurance contracts, insured value and recordkeeping?

The analytical boundary between damage and payment

An engineering damage observation and an insurance claim are different quantities. A house might suffer physical loss and remain uninsured. It might be insured for fire but damaged by a peril excluded from its particular contract. Damage can fall below a deductible, be disputed, or be measured at replacement rather than depreciated value. A valid claim may receive several installments months after the event. A business interruption loss may be covered under a separate time-element grant, subject to special conditions. Each of these mechanisms alters the observed monetary loss while leaving the physical event unchanged.

The published parent Science, Property Insurance, establishes the broad contractual, engineering and financial domain. This dossier studies the statistical and actuarial transformations within that domain. It includes residential, commercial and industrial portfolios and considers catastrophe-related property accumulation. Country-specific law and policy wording are treated as evidence requiring jurisdictional identification, not inferred from an actuarial equation.

Sources, method, and epistemic limits

The sources include original society actuarial texts, binding or professional standards within their stated scopes, peer-reviewed methods, international and national public documentation, and original governmental database descriptions. Werner and Modlin's Basic Ratemaking gives the conventional background on loss costs and exposures; Goldburd and colleagues' CAS GLM monograph supplies the predictive modeling setting; and the Actuarial Standards Board's ASOP 23 data standard establishes relevant data-quality expectations for its practitioners. The EIOPA, FEMA and NOAA sources are used to characterize their observed universes, not to manufacture a new joined dataset.

Mathematical statements are derived from definitions and probability assumptions stated beside them. Empirical claims are attributed to what the cited source actually measures. Institutional rules are limited to their stated regimes. The author has not obtained and fitted insurer claim-level microdata for this dossier; no coefficients, statistical confidence bounds, or causal effects are asserted as newly measured. Several peer-reviewed articles were checked through originating publisher abstracts and metadata, with that access limit preserved in EVIDENCE.json. The completed result is a fundamental documentary, mathematical and comparative synthesis.

Definitions, Historical Context, and Primary Source Foundations

A consistent dictionary of observational units

A peril occurrence is an event capable of damaging property. A damaged asset is an insured or uninsured physical unit experiencing impairment. A policy is a legal contract or group of coverages. A claim may be defined as one notified occurrence, one insured asset–occurrence pair, one coverage unit, or an administratively unique case number. A payment is a dated cash transaction. All are legitimate counting units, but counts between them cannot be substituted without a mapping.

For catastrophe analysis, one flood may cause ten thousand damaged buildings and several thousand reported claims. If event frequency is five floods a year, claim frequency cannot be inferred without exposure to the event footprint, physical vulnerability, insurance participation, coverage and reporting probability. Conversely, a reported claims extract may identify many building claims as one catastrophe code. The frequency denominator depends on the unit: catastrophe events per year, claims per insured site-year, or covered occurrence claims per policy-year.

A property-year measures one property exposed under a specified risk and coverage arrangement for twelve months. A property insured for six months may contribute 0.5 property-years. An insurance policy may cover several sites; a blanket contract counted as one policy-year has a different denominator from the number of building-years. Insured value or replacement value can be used to normalize monetary loss, but a dollar of insured value is a different exposure unit from one building-year. An internally consistent exposure register therefore records inception, expiry, cancellations, endorsements, locations, replacement values and perils.

The same monetary descriptor also requires specification. Ground-up loss means the loss before a particular insurance layer or deductible. Gross insured loss is the insurer's obligation before reinsurance under a defined gross/net convention. Net retained loss subtracts valid ceded recoveries and applies relevant accounting. Paid loss is cumulative cash disbursed by a valuation date. Incurred loss may mean paid plus case outstanding, depending on a reporting system. Ultimate loss estimates total eventual settlement under a chosen scope. A model must not label one of these as “claim severity” without indicating which is measured.

Frequency–severity identities and assumptions

Define the realized positive-claim average s=S/Ns=S/N for N>0N>0, and count rate f=N/Ef=N/E. Then fs=S/Efs=S/E, regardless of statistical independence. This identity can explain a historical experience reconciliation or a ledger audit. A prospective equality such as E[S]=E[N]E[X]\mathbb E[S]=\mathbb E[N]\mathbb E[X] is an additional probabilistic statement when XX is a common independent claim size and NN is an independent count. It also follows in appropriate more general settings from conditional expectations of total claim costs; the form of “average severity” must then be made explicit.

To see the distinction, suppose a tropical cyclone both increases the number of affected properties and produces more expensive repairs in the same affected area. The conditional severity of a claim within a high-claim-count year may exceed that of a claim within a low-count year. Multiplying an unconditional mean event count by an unconditional severity for a different population can misstate expected loss. The correct quantity is E[∑j=1NXj]\mathbb E[\sum_{j=1}^N X_j], possibly computed by conditioning on event environment, insured exposure and the event's spatial footprint.

Exposure itself may be stochastic. An insurer may nonrenew, expand, withdraw or change deductibles in response to weather or previous claims. A portfolio with higher prospective exposure in safer regions can exhibit falling aggregate claim frequency per insured site even while the risk in a particular hazard zone rises. Conditioning or standardizing on portfolio mix is required before a temporal change can be interpreted as a hazard change.

Development of actuarial statistical approaches

The collective-risk view became foundational as insurance organizations required distributions of total claims rather than only average rates. Tabular claims experience evolved toward probability models of claim counts and sizes, credibility, regression, and simulations of large loss. Modern actuarial ratemaking uses exposure normalization, prospective claim adjustment, development, trend, class relativities and expense provisions. The Werner–Modlin treatment explains these linked operations.

Regression methods permit property attributes and exposures to affect frequency and severity separately. The CAS practitioner GLM guide, by Anderson and colleagues (2004), describes data preparation, modeling and refinement. Generalized models have analytical value when a factor can plausibly alter a specific mechanism. Roof condition may affect damage likelihood and repair magnitude; construction cost inflation chiefly changes monetary severity, though its interaction with deductibles also affects the number of positive insurer payments. A coefficient obtained from all claims mixed together represents a statistical association under the chosen observation scheme.

Catastrophe modeling changes the counting level. A large weather event is one peril occurrence affecting many insured sites, often introducing spatial and temporal dependence. Portfolio aggregation must retain this connection. The two scientific questions, how many events occur and how much loss each event generates, remain related to frequency and severity but at the event level rather than the individual-claim level.

Primary sources and their admissible claims

A claims ledger documents recorded notifications, case changes and payments. An exposure inventory documents policy and site coverage during a time interval. A hazard catalog identifies weather or geophysical occurrences independently of insurance purchase. Engineering vulnerability data translate hazard intensity into physical damage, subject to uncertainty. A policy document governs the contractual conversion from damage to claim liability. An accounting standard defines how certain measured liabilities appear in financial statements. These sources provide distinct facts.

The NAIC homeowners report records market conduct information in the United States; the NOAA Storm Events Database describes weather events; and the EIOPA Catastrophe Data Hub describes selected European insured exposure and claims. Their measures cannot be pooled as though each were a complete register of the same risks. A source-to-claim audit must identify its country, date, peril, insured sample, unit and reporting conventions. The OECD 2026 market assessment offers a cross-market context using principally 2025 experience, without supplying a universal insured property claim panel.

Contract terms are statistical transformations

Suppose XX is the amount of an otherwise covered loss before a fixed deductible dd and an illustrative total-loss ceiling uu, with $0\leq d<u.Iftheinsurancecontractappliesthestateddeductibleandlimitconventions,paymentmaybewritten. If the insurance contract applies the stated deductible and limit conventions, payment may be writtenY=\min{(X-d)+,u-d},where, where(z)+=\max(z,0)$. This is an illustrative mathematical transformation of a carefully defined layer, not a representation of every wording. Some contracts apply a percentage deductible to insured value; others have sublimits, aggregate caps, coinsurance clauses, or limits interpreted on a different order of operations.

Once a deductible changes, the rate of insurer-paid claims changes even if the rate of physical property damage is identical. Once a policy limit changes, insurer-paid severity changes even if restoration costs and physical damage ratios are identical. A claims record is consequently an observation of a physical process transformed by a contract and an administrative system.

Exposure and Claim Counts

Count models and the Poisson baseline

Let NiN_i be the number of qualifying claims for risk unit ii during exposure EiE_i. A common baseline assumes Ni∣Ei,xi∼Poisson⁡(Eiλi)N_i\mid E_i,x_i\sim\operatorname{Poisson}(E_i\lambda_i), with λi>0\lambda_i>0 a conditional claim rate determined by covariates xix_i. The conditional mean and variance equal EiλiE_i\lambda_i, and the probability of no claim equals exp⁡(−Eiλi)\exp(-E_i\lambda_i). These are consequences of the Poisson model, not empirical laws applying to all property risks.

A log-link regression can specify log⁡E[Ni∣xi,Ei]=log⁡Ei+xi⊤β\log\mathbb E[N_i\mid x_i,E_i]=\log E_i+x_i^\top\beta. The term log⁡Ei\log E_i is an offset with fixed coefficient one. The model compares claims per unit of defined exposure and allows risk classes to have different conditional rates. The revised CAS GLM monograph supplies a professional context for this approach. It is important to verify that policy-year, building-year, and insurer system identifiers are represented consistently before fitting the model.

For a transparent hypothetical example, 10,000 equally defined property-years with a claim rate of 0.04 have 400 expected claims under the model. A further 2,500 identically defined property-years imply 100 expected claims. These expectations do not provide confidence intervals, estimate a real insurer's loss frequency, or account for correlated weather events. They simply demonstrate the count–exposure relationship.

Overdispersion and mixed rates

Property risks vary across construction materials, occupancy, age, maintenance, regional conditions, fire protection, elevation and hazard exposure. If the claim rate varies randomly across latent environments, a mixed Poisson specification N∣Λ∼Poisson⁡(Λ)N\mid\Lambda\sim\operatorname{Poisson}(\Lambda) yields E[N]=E[Λ]\mathbb E[N]=\mathbb E[\Lambda] and Var⁡(N)=E[Λ]+Var⁡(Λ)\operatorname{Var}(N)=\mathbb E[\Lambda]+\operatorname{Var}(\Lambda). The additional variance comes from variation in the conditional rate.

A negative binomial parameterization with Var⁡(N)=μ+αμ2\operatorname{Var}(N)=\mu+\alpha\mu^2 and α>0\alpha>0 accommodates extra-Poisson variation. Its empirical success does not uniquely identify what created the excess variance. Unmeasured risk heterogeneity, common weather exposures, claim handling, temporal changes or administrative counting can produce similar marginal count patterns. Overdispersion is thus a statistical diagnostic that triggers investigation, not a proof of any specific causal mechanism.

Zero-inflated and hurdle models introduce other mechanisms for zero counts. Structural protection from a covered peril, a policy that excludes the peril, or incomplete observation may create qualitatively different zeros. A report containing no paid claim may reflect no physical loss, an amount below the deductible, a rejected notification, a late payment or a zero-value reserve. A zero-inflated model fit to paid claim counts cannot automatically distinguish these conditions.

Event and claim frequency under catastrophe dependence

Let MM count catastrophe events during a year, and let KeK_e count insured claims generated by event ee. The total catastrophe-claim count is Ncat=∑e=1MKeN^{cat}=\sum_{e=1}^{M}K_e. If MM and KeK_e satisfy specified independence and stationarity conditions, a collective event-cluster model can be analyzed with familiar compounding identities. In practice, KeK_e depends on the event footprint, intensity, insured locations, vulnerability, policy coverage, and insured market participation. An event's geographic reach is a key source of dependence.

The number of storms making landfall is not the same as the number of property claims. A single flood can affect many insured and uninsured structures. Even within one insured portfolio, the claim rate per building is conditional on where the hazard lands. If a model predicts claims independently for each building while ignoring shared wind or flood events, its mean may be plausible while its variance and capital tail are substantially understated.

The NOAA event records and USGS earthquake impact descriptions illustrate two independent hazard-observation systems. NOAA's estimates of weather impacts and USGS rapid earthquake impact estimates are physical and economic proxies with specific methodologies. They do not directly count paid insurance claims or establish their ultimate contract liabilities.

Reported counts, incurred claims, and late notification

Define occurrence-period claims NaoccN_a^{occ} for loss period aa and Na,vrepN_{a,v}^{rep} claims reported by valuation date vv. For a consistent universe of occurrences with unique identifiers, reported counts usually develop toward an ultimate count as notifications arrive. Catastrophe access restrictions, loss discovery delays, and coverage uncertainty may prolong that development. A claim reopened after closure should be distinguished from a new loss occurrence.

A report-year series counts what enters the insurer's administrative system in the reporting period. An accident-year series counts events by loss date. Their rates can differ, particularly immediately following catastrophes. A researcher comparing frequency by period must specify which origin and maturity convention is used. A recent period may look artificially safe simply because many late notifications have not arrived.

An insurer might divide building and contents losses under one policy into one claim identifier, while another produces separate coverage records. A single fire can involve a property damage component, debris removal, emergency accommodation, and time-element costs. Counting each payment component as a separate claim inflates frequency relative to a case-based definition. Data reconciliation requires an explicit mapping of IDs to unique occurrences and insured entities.

Exposure registers and selection

An exposure file needs risk inception and termination dates, asset and occupancy characteristics, geographic coordinates, peril eligibility, deductible and limit versions, and the effective time of each change. A risk added in October contributes only a fraction of a full calendar year. A commercial policy covering many locations cannot automatically be treated as one building of exposure. Exposure amount based on total insured value must be distinguished from a count of insured structures, because a smaller number of high-valued properties may carry disproportionate severity.

Insurers may alter the portfolio in response to past losses, hazard forecasts, or reinsurance prices. Nonrenewal, cancellation, underwriting restrictions, policy changes and location-level premium increases can change the observed insured population. A falling observed frequency following withdrawal from flood-prone zones does not demonstrate a decline in underlying flood risk for the community.

The NAIC homeowner conduct data provide a U.S. source for policy-related market outcomes. Any causal claim about nonrenewals and physical risk would need linked portfolio and hazard measures, rather than relying solely on total market premiums.

Count-model validation and leakage

Count models should be assessed by risk class, geography, policy cohort and calendar period, not just overall likelihood or error. Residual overdispersion, extreme claims years and spatially concentrated events should be investigated. A temporally ordered holdout better approximates prospective ratemaking than randomly splitting multiple claims from the same catastrophe between training and validation. Event-based cross-validation should keep the entire event together when the purpose is to predict new events.

A claim adjusted and paid after the underwriting date may contribute useful information to an ultimate loss target, but its later claim-handling details were not available at the earlier underwriting decision. Including post-event inspections or settlements as covariates in a model intended for new-policy pricing produces information leakage. The observed score will then exaggerate the prospective predictive information actually available.

Loss Severity Tails and Censoring

Definition of claim severity

Claim severity may refer to final paid indemnity, ultimate gross claim cost including adjustment expense, ground-up cost before deductible, or the amount remaining after reinsurance. These different financial variables have different distributions and may differ by orders of magnitude. Each severity sample needs an inclusion rule, a currency and price date, a development maturity, a contractual basis, and a statement on whether zero-paid or denied claims are included.

A conditional positive-severity model may use building replacement value, occupancy, insured object type, cause of loss, geographic hazard intensity, vulnerability and settlement conditions. It should distinguish damage caused by the hazard from the price of remedying that damage. A building may become costlier to rebuild because construction costs rise even if its physical fragility is unchanged.

Compound-loss expectation and variance

Let NN be a claim count and X1,X2,…X_1,X_2,\ldots iid positive severity variables independent of NN, with μ=E[X]\mu=\mathbb E[X] and σ2=Var⁡(X)\sigma^2=\operatorname{Var}(X) finite. Set S=∑j=1NXjS=\sum_{j=1}^{N}X_j. The law of iterated expectation gives E[S]=E[N]μ\mathbb E[S]=\mathbb E[N]\mu. The law of total variance gives Var⁡(S)=E[N]σ2+Var⁡(N)μ2\operatorname{Var}(S)=\mathbb E[N]\sigma^2+\operatorname{Var}(N)\mu^2.

For a Poisson count with mean ν\nu, Var⁡(N)=ν\operatorname{Var}(N)=\nu, so Var⁡(S)=ν(σ2+μ2)=νE[X2]\operatorname{Var}(S)=\nu(\sigma^2+\mu^2)=\nu\mathbb E[X^2]. This derivation indicates why accurate mean severity alone is insufficient to describe aggregate uncertainty. It also shows why ignoring frequency variance matters outside the Poisson assumption.

Take an explicitly hypothetical portfolio with 400 expected claims, $20,000 mean severity, and $40,000 severity standard deviation. Expected aggregate loss is $8 million. Since E[X2]=2.0×109\mathbb E[X^2]=2.0\times10^9 in squared monetary units, the aggregate variance is $8.0\times10^{11}, and the modeled standard deviation is about \894,427$. The model calculation supplies no real market forecast or empirical confidence band. Dependence between claims would require adding appropriate covariance terms.

Positive-claim distributions and heterogeneity

The Gamma and lognormal distributions are common positive right-skewed working families. The Gamma is often analytically convenient with positive support. A lognormal can describe multiplicative financial and physical drivers. Pareto-type heavy tails may be used for unusually large losses, but a parametric family should not be accepted merely because its visual curve looks plausible in the central sample. Extreme outcomes are sparse and sensitive to the observation period and policy terms.

A mixture of attritional and major claims may be more transparent for pricing. Consider 600 small claims averaging $2,000 and 20 large claims averaging $100,000 in a hypothetical experience. The small claims cost $1.2 million, the large claims $2 million, and total loss is $3.2 million. The 20 large claims represent approximately 3.23% of claim count but 62.5% of cost. This is a numerical illustration of concentration, not an external estimate of homeowners or commercial insurance.

If the large-loss threshold is fixed in nominal money, inflation will cause more otherwise unchanged real losses to pass it. Threshold analysis should either convert claims to comparable real prices or disclose the extent of drift. An apparent increase in large claims may reflect currency values and contractual limits rather than a larger physical event distribution.

Deductible selection and left truncation

Let XX measure the otherwise covered pre-deductible loss for one occurrence and dd a fixed deductible. A stylized paid amount is (X−d)+(X-d)_+. The probability of a positive insurer payment is Pr⁡(X>d)\Pr(X>d), and the average paid amount conditional on a positive payment differs from the original mean damage. A larger deductible generally reduces the number of positive insurer payments and changes the conditional paid severity population.

If the insurer database excludes all losses below deductible, the original XX observations are left-truncated. The sample distribution observed above dd is the original severity conditional on X>dX>d. A likelihood fit must account for that selection. If damage amounts are recorded for every event but payments of zero below the deductible are retained, the payment variable instead has an atom at zero. These are distinct statistical datasets even if they produce the same positive paid claim list.

The stop-loss identity E[(X−d)+]=∫d∞Pr⁡(X>x) dx\mathbb E[(X-d)_+]=\int_d^\infty \Pr(X>x)\,dx holds for nonnegative losses with finite expectation. It shows that the expected deductible-layer cost depends on the integrated survival distribution above the deductible. Modeling the proportion of claims exceeding dd is therefore insufficient to price the layer; how far the excess losses extend also matters.

Policy limits and right censoring

Let uu denote an upper ceiling on an otherwise covered loss and apply the chosen deductible–limit conventions. A policy payment may be capped at a known maximum even when physical loss continues growing. If only the capped payment is recorded, the payment amount is an exact observation of what the insurer paid, but the underlying ground-up damage above the cap is only bounded from below. Treating the capped figure as an exact damage measurement biases an uncapped damage tail downward.

Some insurers retain a separate adjuster's estimate of ground-up damage beyond the policy limit; others do not. A model seeking the distribution of insurer payments should handle capped amounts as financial outcomes. A model seeking physical loss or the loss cost of a higher-limit policy must treat the information limit explicitly. Both are reasonable tasks, but they have different targets and likelihoods.

Percentage deductibles, named-peril sublimits, coinsurance and aggregate exhaustion complicate the simple example. A catastrophe deductible expressed as a percentage of total insured value can have a changing absolute amount as the insured value changes. Policy-level frequency–severity comparisons must therefore retain applicable contract revisions and valuation basis rather than infer risk change from a fixed nominal paid-loss threshold.

Statistical tail models and event clustering

Peaks-over-threshold methods use exceedances above a high threshold to model upper tails, often with a generalized Pareto approximation under an appropriate asymptotic regime. A tail model must report threshold choice, exceedance count, inflation and development treatment, policy limit censoring, and uncertainty about tail shape. A long period with no major catastrophe within a small observed portfolio is not proof that extreme losses are impossible or have negligible future probability.

Property catastrophe claims generated by one event may be correlated even after conditioning on geography. High hazard intensity may simultaneously increase the number of damaged structures and the severity of each damaged structure. Fitting independent tail severity distributions to all claims of one cyclone can understate uncertainty because the effective number of independent extreme events is much smaller than the number of claim rows.

For a combined portfolio loss S=∑pSpS=\sum_pS_p across perils, Var⁡(S)=∑pVar⁡(Sp)+2∑p<qCov⁡(Sp,Sq)\operatorname{Var}(S)=\sum_p\operatorname{Var}(S_p)+2\sum_{p<q}\operatorname{Cov}(S_p,S_q) when variances exist. The cross-peril covariances may reflect common weather or economic regimes. Frees and Valdez's copula research demonstrates a general actuarial approach to multivariate dependence. Whether a chosen copula is appropriate for severe storms, earthquake shocks or insurer aggregation remains an empirical question requiring validation.

Tweedie models and compound outcomes

A compound Poisson sum with positive Gamma severities belongs to the Tweedie family for an appropriate variance-power parameter. The original Tweedie claims study by Jørgensen and de Souza (1994) treats this construction in insurance data. With a standard mean–variance power relation Var⁡(S)=ϕμp\operatorname{Var}(S)=\phi\mu^p, the compound Poisson–Gamma setting corresponds to $1<p<2$. Its mass at zero and continuous positive component can be useful for loss cost per exposure.

A Tweedie aggregate model can estimate the combined cost without a separately fitted claim-count regression and conditional severity regression. It is less immediately informative when the scientific question requires attributing a change to frequency, conditional severity, claim reporting or deductible selection. Distributional flexibility does not automatically resolve physical catastrophe dependence or adverse selection in which policies appear in the sample.

Extreme-loss return periods and model boundaries

A modeled return period pertains to a clearly specified random variable, observation horizon and probability distribution. A 100-year event characterization concerns the selected annual exceedance probability under specified stationarity assumptions; it does not mean such events arrive once every 100 calendar years on schedule. A 100-year hazard footprint is also different from an insurer's 100-year gross aggregate loss and from a reinsured 100-year retained loss. Location selection, vulnerability, coverage and contract terms transform the distributions.

Estimating property tail loss requires separating model uncertainty from ordinary process variability. Sparse severe events, changing building stock, evolving climate regimes, and changes in market coverage can destabilize simple historical frequency extrapolation. The IPCC synthesis report documents evolving physical risk and adaptation uncertainty, while insurer-specific losses require further exposure and financial modeling.

Pricing and Reserving Decomposition

Pure premium and portfolio mix

For a homogeneous prospective risk group with claim frequency f∗f^* and conditional mean insured severity s∗s^* on a consistent basis, the expected pure loss cost is f∗s∗f^*s^* per exposure unit. For heterogeneous groups with exposure Ei∗E_i^*, rates fi∗f_i^* and severities si∗s_i^*, the weighted expected pure loss cost is (∑iEi∗fi∗si∗)/(∑iEi∗)(\sum_iE_i^*f_i^*s_i^*)/(\sum_iE_i^*). Multiplying an unweighted average frequency by an unweighted average severity of differently sized groups can give a materially incorrect result.

The final premium also involves allocated and unallocated claim expenses, acquisition and administrative expenses, capital and risk-related provisions, and applicable requirements. Werner and Modlin (2016) present the conventional ratemaking context. A pure premium equation is a loss-cost derivation, not a universal approved rate or a rate-filing recommendation. National price regulation, mandated coverages, tax treatments and distribution costs differ.

Suppose, for a fixed prospective class and financial coverage basis, claim frequency grows by 10% and mean severity by 20%. Expected pure loss cost grows by $1.10\times1.20=1.32$, or 32%. The additional two percentage points are the multiplicative interaction, not an extra independent trend. If the estimated changes refer to different portfolio mixes or policy limits, multiplying them is not justified until those definitions are aligned.

Claim inflation and on-level adjustments

Construction labor, materials, building code changes, replacement technology, emergency accommodation and contractor capacity can alter repair costs. Insured values may also be updated, altering deductibles or average-clause applications. A higher repair-cost index can move more claims above a nominal deductible, increasing reported positive-paid frequency while raising severity. A trend attributed to “severity inflation” may thus operate through the observation threshold as well.

The ASOP 13 trending standard addresses procedures for projecting historical property/casualty measures and for assessing trends. Historical claims development estimates the ultimate cost of past occurrence periods; economic trend projects a loss measure to a different price or exposure environment. Confusing them can generate duplicate adjustments. The analysis must state the origin date of claims, the valuation date of their estimates, and the prospective policy period to which they are projected.

Where perils differ, a single inflation factor is rarely adequate for research interpretation. Fire rebuilding, water damage remediation and commercial machinery replacement may respond to distinct price baskets and repair processes. A hazard-aware rate study would identify peril-specific loss cost, event activity, construction and labor inflation, exposure changes and contract updates with consistent definitions.

Claims reserves and development time

Reserving asks about obligations from events that have already occurred. Pricing asks about events expected during a future coverage period. The ASOP 43 unpaid-claims standard requires clarity about the intended unpaid claim measure and related data, assumptions and communication in the relevant professional practice. This includes defining gross versus net amounts, the included claims expenses, the measurement date and the purpose of the estimate.

For occurrence or accident period aa and development age dd, let Ca,dC_{a,d} be cumulative paid or incurred loss, consistently measured. Chain-ladder methods develop the observed loss history toward an estimated ultimate using selected age-to-age factors. The method relies on assumptions about stability and representativeness of development patterns. Mack's chain-ladder uncertainty paper dates from 1993 and supplies one recognized approach to estimating reserve variability under its specified assumptions. The original publication year remains 1993 even though the publisher's web system may show a later online posting date.

A hypothetical accident year has $5 million cumulative paid at a given maturity and a selected cumulative development factor 1.25. Projected ultimate is $6.25 million, implying $1.25 million remaining payment on that basis. This arithmetic does not estimate a real outstanding claim liability or its confidence interval. If payment processes changed because of catastrophic access, labor shortages or unusual litigation, the historical factor may cease to be representative.

Frequency development and severity development

The number of claims reported by age dd can differ from ultimate reported count because new notifications arrive. Reported claims may develop in value because adjusters revise estimates; paid claims develop because disbursements occur. One approach estimates ultimate reported claims and multiplies by expected ultimate claim severity. Another directly develops cumulative paid or incurred dollars. These approaches may reconcile under appropriate models; multiplying both by separate IBNR loadings without a consistent decomposition may double count late claims.

Large-loss claims are especially problematic. An individual industrial fire may be reported promptly but take years to settle, while many modest household claims close relatively quickly. Mixing these claim types in one development triangle can mask distinct settlement distributions. A catastrophe may create many simultaneously reported claims, delayed inspections, coverage disputes, and later supplemental payments. Segmentation by peril, event and claim size can improve interpretation but can also reduce credibility if cells become extremely sparse.

Accounting liabilities versus actuarial loss totals

IFRS 17 describes insurance accounting through categories such as liability for remaining coverage, liability for incurred claims, fulfilment cash flows and risk adjustment. The IFRS 17 key-terms documentation helps establish the accounting vocabulary; the full issued standard governs applicable reporting. The present-value and risk-adjustment requirements depend on the reporting framework and the underlying insurance-contract group. They cannot simply be equated with an undiscounted compound Poisson mean.

The IAIS Insurance Core Principles provide a wider international supervisory setting concerning risk management and insurer financial integrity. An internal property loss frequency model may inform pricing, claims reserving, reinsurance and risk capital, but those uses have distinct objectives, time horizons and governance standards. A numerical agreement in expected annual loss cost does not guarantee agreement in capital needs or contractual claims obligations.

Reinsurance and gross/net definitions

Reinsurance often aggregates many site claims into an event- or occurrence-level financial amount. A per-risk deductible or limit is applied to an individual insured claim according to its policy; an occurrence reinsurance attachment may then apply to the sum of many policies' gross claims. One large storm can reach the insurer's catastrophe retention even when no single property claim is extraordinary. Applying treaty attachment independently to each site would misrepresent that coverage.

Gross losses and net retained losses require separate variables, with recovery timing, credit quality, exclusions, reinstatement and aggregation rules represented as appropriate. Salvage and subrogation also affect net recorded severity. A fixed percentage of gross loss is generally inadequate to estimate net severity under nonproportional reinsurance. The claims distribution must be passed through the actual treaty structure to estimate recoveries.

A portfolio management consequence

Frequency–severity decompositions inform decisions only when their units and assumptions are documented. An underwriting model optimized for mean expected claims may underrepresent catastrophe accumulation. A reserving model that predicts paid claims rapidly may still have substantial tail uncertainty. An accounting measure may reflect discounting and risk adjustments absent from a nominal pure premium model. Shared terminology and traceable transformations allow these distinct decisions to exchange information without silently collapsing their scientific objects.

Empirical Verification, Uncertainty, and Competing Explanations

Public insurance data and their selection mechanisms

An insurer's claims database observes reported and processed insurance outcomes. A government hazard dataset observes certain events regardless of insurance purchase. An international market summary aggregates reported financial and market measures. Their samples and denominators differ; a valid joint analysis requires reconciliation of policies, covered hazards, asset values and time.

The OpenFEMA redacted claims catalog identifies a public dataset of National Flood Insurance Program claims. The official NFIP data questions explain limitations of privacy-protected policies and claims extracts, including restrictions on following individual properties. A researcher cannot infer a fully linked insured-building panel solely from these releases. Aggregate geographical comparisons may be possible with appropriate scope restrictions and a defined exposure proxy, but the resulting rates cannot automatically be described as per-property physical flood incidence.

The EIOPA catastrophe data hub provides observations concerning selected exposures and insured catastrophe losses, particularly for flood and windstorm risk. Its participating insurers are a sample rather than the entire European insurance market. Generalizing to all European properties requires a justified estimation of selection and coverage. This documentary fact is independent of whether a statistical model fits the available observations.

The NOAA Storm Events Database records hazard occurrences and selected damage information. It does not count every property insurance claim. An event catalog can support peril timing and spatial exposure reconstruction when joined with policy and asset records; it should not supply the insurer's claim denominator by assumption. Similarly the USGS PAGER approach offers rapid earthquake loss estimates that differ from finally adjusted and paid insurance amounts.

An insurance protection-gap example

EIOPA's 2025 protection-gap description documents a comparison between economic losses and insured losses for natural catastrophes. On aligned event and financial definitions, define an uninsured economic loss share as (Leco−Lins)/Leco(L^{eco}-L^{ins})/L^{eco} when Leco>0L^{eco}>0. If a hypothetical event has $100 million of total economic damage and $35 million of insured loss, the difference is $65 million, or 65% of the total. The calculation expresses an economic insurance share, not claim frequency, insurance take-up, contract availability or the probability of physical damage.

A jurisdiction can exhibit widespread policy ownership and still have a high uninsured loss share if the dominant peril is not covered or limits and deductibles reduce indemnification. Conversely, a low insured share may reflect uninsured assets outside the formal insurance market. Interpreting an observed protection gap requires construction of compatible economic and insured numerators and denominators.

Competing causes of apparent severity growth

Suppose a claims ledger shows a 30% increase in nominal mean paid claim size. Several mechanisms can create this observation: building material inflation, more expensive restoration, greater hazard intensity, shift toward commercial risks, lower attritional claims reporting, increased deductibles suppressing small claims, changes in policy limits, large one-off catastrophes, slower payment development, or changed claim identifiers. The observed ratio alone cannot identify which mechanism operated.

A valid analysis separates claims by stable peril and coverage definitions, fixes an exposure basis, converts monetary amounts to a declared currency-price level, identifies deductible and limit changes, and develops claims consistently. Where available, independent hazard intensity and engineering damage indicators help test whether the physical distribution changed. Event matching and geospatial exposures help distinguish a catastrophic year from a general shift in attritional claim behavior.

The NAIC homeowners report establishes a U.S. reporting context; the OECD global market report characterizes broader national and international insurance outcomes. Neither by itself identifies insurer-specific structural coefficients for property claim severity. In particular, global growth in premium volume can result from price changes, rising insured asset value, coverage penetration or portfolio mix rather than more frequent physical disasters.

Evolving hazards and climate uncertainty

The IPCC AR6 synthesis provides a scientific assessment of physical risks, changing exposure and limits to adaptation. It does not provide universal actuarial risk coefficients to be inserted directly into all national property books. Translating physical projections into insured loss requires peril-specific hazards, location and asset exposures, vulnerability, future construction and mitigation, insurance purchase, contract coverage and settlement mechanisms.

Historical stationarity may be implausible in an evolving climate or rapidly developing urban region. Yet replacing a historical model with an external hazard projection without validation introduces additional structural uncertainty. Researchers should test alternative hazard conditions, quantify sensitivity to the chosen climate and socioeconomic scenarios, and disclose spatial mismatch between projections and insured assets.

An insurer's property portfolio also changes endogenously. Better mitigation and new building codes may reduce loss ratios for selected insured assets even as the regional hazard intensifies. Geographic underwriting retreat may lower average insured claim frequency while raising societal uninsured economic loss. These observations require distinct denominators and mechanisms.

Tail dependence and catastrophe validation

An annual claim model may reproduce mean count and average severity while underestimating the probability of simultaneous large losses. Shared events generate clusters, so a random claim-level training/test split may place claims from the same hurricane or flood in both sets. A holdout by future catastrophe event, season or geography better measures the model's ability to generalize. Frees and Valdez (1998) provide a foundational source on modeling actuarial dependence, with applications requiring empirical selection and calibration.

Severe-loss estimates are sensitive to how limits, deductibles and reported losses are represented. An uncapped ground-up tail cannot be validated solely from capped insured payments without addressing censored losses. A reinsurance tail must use occurrence aggregates and policy/treaty financial terms. Tail diagnostic checks should include alternative thresholds, extreme-event exclusions, inflation conventions, event clusters and uncertainty in the severity second moment.

Data quality and methodological audit

The ASOP 23 quality standard is relevant to selection, review and disclosure of data. For a property claims study, the audit should establish an exact observation cutoff; reconcile records to accounting totals; identify duplicate and reopened claims; verify effective coverage; remove inconsistent currency conventions; map losses to unique peril events; record the maturity of ultimate estimates; and document whether claim reporting thresholds create selection. A reproducible research output should record these transformations with versioned code or an executable analysis specification.

Validation should compare observed and predicted counts across genuine out-of-sample periods, positive severity quantiles, calibration of total loss cost, and behavior of tail and clustered losses. Compare frequency and severity changes separately across risk classes, and recompute a standardized portfolio aggregate. A model with good overall aggregate performance may still have serious bias in one vulnerable region or in a catastrophe tail.

The ASOP 43 reserving standard is also relevant when incurred or ultimate loss rather than paid-to-date is used. A frequency model based on mature report periods and a severity model based on immature incurred estimates can yield misleading trends. Documenting ultimate-loss methods and development factor choices allows their contribution to the forecast to be evaluated independently.

Classes of uncertainty

Process uncertainty concerns random occurrences and individual loss outcomes under a given model. Parameter uncertainty concerns estimation from finite data. Model uncertainty concerns distributional form, tail selection, covariates and dependence assumptions. Measurement uncertainty concerns misclassification, reporting lags, contract terms and missing exposures. Structural uncertainty concerns future changes in hazard, economic conditions, market participation, building stock and law. These sources should be stated separately because they lead to different responses.

A claimed uncertainty interval must identify the object it covers: prospective pure premium, ultimate unpaid claims, annual aggregate loss, catastrophe return-period estimate or something else. Different intervals are not interchangeable. In particular, a fitted confidence interval for average annual cost does not by itself represent a solvency buffer for the rare joint loss of many insured sites.

Limits of the present documentary research

The source record for this monograph verifies the existence and documented scope of primary standards and public databases, together with bibliographic records and selected publisher abstracts for original statistical literature. The mathematical results and numerical scenarios follow from stated assumptions and reproducible arithmetic. The research does not assert independent estimation of contemporary property frequency, modeled protection-gap coefficients, or insurer reserving parameters from unprovided proprietary data. The distinction is scientifically material: no simulated illustration is presented as a measured real-world finding.

Integrative Comparative Scientific Synthesis

The exact identity and the conditional model

The product of observed frequency and observed average severity equals observed loss cost per exposure only when numerator, denominator and monetary loss definitions are consistent. This is an accounting identity; it needs no assumption of Poisson events, independence, stable climate or particular claim-size distribution. Its value lies in transparent reconciliation. For prospective inference, an additional set of probability and exposure assumptions is indispensable.

Consider two hypothetical portfolios each expecting $8 million of insured aggregate claims. One expects 400 claims at $20,000; the other expects 200 claims at $40,000, under a comparable exposure basis. The expected totals match but the distributions, catastrophe sensitivity, claim administration workload and financial tail can differ. The decomposition reveals this distinction. Even two models with the same count rate and mean severity can have sharply different aggregate uncertainty if severity variance or event dependence differs.

Three scientific layers of property claim generation

The underlying physical layer consists of hazards, asset location, engineered vulnerability and damage. The contractual layer transforms physical and economic damage through insured interest, covered peril, valuation, exclusions, deductibles and limits. The administrative and financial layer transforms insurer obligation into reporting, case reserves, settlements, reinsurance recoveries and paid amounts over time. These layers explain why a single observed claims total can be consistent with several different causes and interventions.

For example, a fall in recorded paid-claim frequency can follow improved building protection, higher deductibles, reporting delays or an insurer's withdrawal from vulnerable locations. Each has different implications for physical risk, insurance affordability, consumer protection and reserves. Independent records from the three layers are needed to distinguish them. A frequency–severity model is most informative when its explanatory variables are traced to the layer in which they operate.

Conditional comparison across classes and perils

For a widely dispersed portfolio of ordinary residential fires, a count model and positive severity model may be an adequate working approximation after conditioning on building and exposure. For flood, wind and wildfire, shared hazard footprints introduce spatial clusters and selection by insurance participation. For high-value industrial assets, a small number of large losses, engineering characteristics, limits and complex adjustment may dominate severity estimation. For coastal commercial property, a catastrophe occurrence can produce a large collection of correlated physical and interruption claims.

These are analytical propositions about suitable observational units, not claims that one international insurer model is legally or empirically universal. The appropriate frequency unit may be a site-year, covered occurrence, named event or reinsurance occurrence. The appropriate severity target may be ground-up damage, gross property claim, net retained insurer cost or completed paid claim. Each pair answers a different question.

Why naïve factorization fails under change

If a deductible grows while physical loss rates stay constant, fewer paid claims may remain and average positive paid severity can rise because small claims disappear. If rebuilding inflation increases physical repair costs, the same nominal deductible is less restrictive in real terms and more physical claims may trigger payments. If a cyclone produces both many claims and unusually large claim sizes, a model assuming count–severity independence may understate variance. If coverage take-up changes, observed insured severity and economic damage distributions may diverge. These interactions establish practical limits on interpreting observed component movements as independent drivers.

The EIOPA protection-gap analysis and the NFIP data guidance illustrate why sampling, coverage and reporting affect insured statistics. The data universe is part of the scientific proposition. A model developed for a selected insured population must disclose that population before its parameters are transferred elsewhere.

International inference and comparable evidence

An international study of property frequency should harmonize insured asset units, exposure duration, claims identifiers, perils and coverage definitions. An international severity study should also harmonize claim-development maturity, currency and construction-price base, insured values, deductible/limit basis, catastrophe codes and gross/net conventions. National financial summaries can support comparisons of market development, while detailed claim statistics require more precise administrative microdata.

The OECD 2026 report is useful as a macro-institutional comparison; the EIOPA catastrophe sample supplies selected peril-specific exposure and loss observations; and the NOAA storm archive identifies weather events in the United States. Each illuminates a different component. Their combination into a single comparative frequency metric would require a separate explicit research design, access to compatible data and uncertainty modeling.

Reproducibility and further empirical work

A fully testable property-claim study should archive a dictionary of covered exposures, claim units and peril groups; document policy financial terms, nominal and real currency conventions and data cutoffs; link insured sites to hazard events where lawful and feasible; estimate reporting and settlement development; and compare independent count, severity and aggregate models on temporally ordered validation samples. The source archive should identify which conclusions derive from a contract, standard, observed dataset, academic method or mathematical identity.

The intellectual contribution of this monograph lies in the integrated critical reading of these established relationships, in transparent decomposition of observational selection, and in a comparative account of the evidence required to distinguish competing mechanisms. The work makes no claim to establish a novel actuarial distribution, a new theory of risk, a proprietary model or a measured global causal effect. It specifies a coherent approach to using existing actuarial knowledge without confusing mathematical identity with empirical explanation.

Conclusion

Property insurance frequency and severity measure observed outcomes within a defined insured, contractual and reporting population. The historical identity of count per exposure multiplied by average claim cost reconstructs loss per exposure, while its prospective application requires a model of changing risk, coverage and reporting. Compound-loss theory makes explicit how count variance, severity variation and event dependence affect aggregate uncertainty.

The review yields five principal scientific conclusions. Claim frequency differs from catastrophe-event frequency and physical damage incidence. Deductibles, limits, claim selection and settlement maturity directly reshape observed frequency and severity. Extreme loss and geographic dependence determine tail accumulation beyond average loss cost. Ratemaking, claims reserving, reinsurance and financial reporting require different temporal and contractual valuations. Finally, international comparative inference depends on aligning policy, asset, event, monetary and administrative denominators before modeling trends.

The practical research agenda is to build longitudinal event–asset–policy–claim cohorts with verified contract transformations, mature loss development, transparent inflation adjustments and external hazard information. Such datasets can test whether changing insured losses reflect peril intensity, exposure, vulnerability, insured penetration, contract changes or settlement dynamics. The present monograph completes a documentary and mathematical synthesis of the subject and preserves the distinction between verified facts, conventional model assumptions, hypothetical arithmetic and empirical questions requiring new observations.

References

  • Werner, Geoff, and Claudine Modlin. (2016). Basic Ratemaking. 5th ed. Casualty Actuarial Society. Original source

  • Goldburd, Mark, Anand Khare, Dan Tevet, and Dmitriy Guller. (2025 revision). Generalized Linear Models for Insurance Rating. 2nd ed. Casualty Actuarial Society Monograph No. 5. Original source

  • Jørgensen, Bent, and Marta C. Paes de Souza. (1994). Fitting Tweedie's Compound Poisson Model to Insurance Claims Data. Scandinavian Actuarial Journal, 1994(1), 69–93. https://doi.org/10.1080/03461238.1994.10413930. Original source

  • Actuarial Standards Board. (2009). Actuarial Standard of Practice No. 13: Trending Procedures in Property/Casualty Insurance. Original source

  • Actuarial Standards Board. (2016). Actuarial Standard of Practice No. 23: Data Quality. Original source

  • Actuarial Standards Board. (2007). Actuarial Standard of Practice No. 43: Property/Casualty Unpaid Claim Estimates. Original source

  • Mack, Thomas. (1993). Distribution-Free Calculation of the Standard Error of Chain Ladder Reserve Estimates. ASTIN Bulletin, 23(2), 213–225. https://doi.org/10.2143/AST.23.2.2005092. Original source

  • Frees, Edward W., and Emiliano A. Valdez. (1998). Understanding Relationships Using Copulas. North American Actuarial Journal, 2(1), 1–25. Original source

  • European Insurance and Occupational Pensions Authority. (2025, October 6). Technical Description: Dashboard on Insurance Protection Gap for Natural Catastrophes. EIOPA-BoS-25/565. Original source

  • European Insurance and Occupational Pensions Authority. (2023; current online description). Catastrophe Data Hub: Insured Exposure and Loss Data. Original source

  • Federal Emergency Management Agency, National Flood Insurance Program. (2026 access). FAQs About NFIP Policies and Data. Original source

  • Federal Emergency Management Agency. (2026 access). FIMA NFIP Redacted Claims (OpenFEMA). Data.gov Catalog. Original source

  • National Oceanic and Atmospheric Administration, National Centers for Environmental Information. (2026 access). Storm Events Database. Original source

  • National Association of Insurance Commissioners. (2026 access). MCAS Homeowners Insurance Report, 2018–2024. Original source

  • U.S. Geological Survey. (2026 access). PAGER OnePAGER Information. Original source

  • IFRS Foundation. (2026 access). IFRS 17 Insurance Contracts: Key Terms. Original source

  • Organisation for Economic Co-operation and Development. (2026). Global Insurance Market Trends 2026. OECD Publishing. https://doi.org/10.1787/c8be6585-en. Original source

  • Anderson, Duncan, Sholom Feldblum, Doris Y. Schirmacher, Ernesto Schirmacher, Neeza Thandi, and Claudine H. Modlin. (2004). A Practitioner's Guide to Generalized Linear Models. CAS Discussion Paper Program, 1–116. Original source

  • Intergovernmental Panel on Climate Change. (2023). Climate Change 2023: Synthesis Report, Longer Report. IPCC. Original source

  • International Association of Insurance Supervisors. (2024 revisions; accessed 2026). Insurance Core Principles and ComFrame Online Tool. Original source

  • Casualty Actuarial Society. (2026 access). Bibliographic Record: Thomas Mack (1993), Distribution-Free Calculation of the Standard Error of Chain Ladder Reserve Estimates. Original source