RESEARCH DOSSIER
Loss Severity and Heavy-Tailed Distributions
Abstract. Loss severity modeling concerns the probability distribution of the monetary size of a loss conditional on an appropriately defined loss event. Insurance records frequently contain a more complicated object: payments selected by deductibles, capped by limits, affected by claims handling, and observed only when reporting and settlement conditions are satisfied. The distinction between the underlying positive loss and the recorded payment is essential for inference. This monograph derives and compares gamma, lognormal, Pareto, and related heavy-tailed models; explains tail equivalence, regular variation, subexponential behavior, and moment existence; and establishes likelihoods for left-truncated, right-censored, and interval-observed amounts. It then derives limited expected values, quantiles, expected shortfall, and layer costs under actual contract transformations. Worked calculations show how a Pareto model with a finite mean and infinite variance can produce materially different tail estimates from an otherwise plausible light-tailed fit, and how an insurance limit converts an unbounded loss into a bounded payment without making the gross loss distribution bounded. The original analytical contribution is an integrated comparison of three distinct mathematical objects—gross severity, observation-conditioned severity, and coverage-transformed payment—together with explicit consistency checks connecting their likelihoods and expectations. These are applications and syntheses of established probability and actuarial methods, not a proposed new theoretical system. The conclusions specify which estimands and assumptions each model can support, and why reporting thresholds, inflation, exposures, portfolio change, tail sample size, and model uncertainty govern the reliability of catastrophic-loss extrapolation.
Keywords: Loss severity; actuarial risk models; heavy-tailed distributions; Pareto distribution; lognormal distribution; gamma distribution; regular variation; subexponential distributions; loss truncation; right censoring; deductibles; policy limits; limited expected value; tail index; quantile; expected shortfall; excess-of-loss reinsurance; extreme value theory; model uncertainty.
Introduction
An insurance loss may be reported as a damage estimate, a claim incurred amount, an insurer payment, a paid amount net of salvage, a recovery-inclusive net amount, or a reinsurer's participation. Each measure answers a different question. A valid statistical model requires an explicit monetary unit, loss definition, sampling population, exposure period, valuation date, and contractual transformation. The central actuarial task is to connect the modeled random variable to the financial quantity that will actually be priced, reserved, insured, retained, transferred, or disclosed.
The mathematical foundations of this subject are developed in Klugman, Panjer, and Willmot's Loss Models: From Data to Decisions, whose treatment integrates individual distributions, payment modifications, and aggregate claims. Embrechts, Klüppelberg, and Mikosch's Modelling Extremal Events provides foundational results on extremes and insurance tail risk, while Resnick's Heavy-Tail Phenomena develops the probability and statistical theory behind regular variation and tail inference. The following analysis uses that established body of work and makes the assumptions of each claim explicit.
The object of this Research Dossier is individual loss severity. The theory of block maxima and generalized extreme-value limits receives a bounded treatment as a tool for identifying upper severity behavior. Comprehensive asymptotic extreme-value methodology, joint tail dependence, count models, and compound aggregate loss models belong to their own Science-linked Research Dossiers. Their interfaces are identified here because severity results are often misapplied when passed into aggregate-loss or solvency calculations.
Every formula in this study refers to a hypothetical probability model, unless an external source is explicitly named. No proprietary claims records, insurance-company experiences, calibration data, or new empirical findings are represented as having been observed. All numerical examples use transparently specified hypothetical distributions. This distinction is indispensable: a mathematically exact sensitivity example demonstrates consequences of assumptions, whereas empirical confirmation requires independently observed exposure and claims information.
Individual Positive Losses
Definition of the severity random variable
Let (X) be a nonnegative gross monetary loss measured in a fixed currency at a specified price level. When an event may cause no economic damage, (X) can have a probability mass at zero. For a strictly positive severity model, define (X^+=X\mid X>0). The corresponding distribution function is (F(x)=\Pr(X\leq x)), the survival function is (S(x)=1-F(x)=\Pr(X>x)), and a density, where it exists, is (f(x)=F'(x)). Conditional and unconditional probabilities must be distinguished whenever zero-loss events or reporting thresholds are present.
A severity distribution is a model of the size of an identified kind of loss; it is not a complete model of how often such losses arise. If (N) is the number of events and (X_1,X_2,\ldots) their sizes, a collective aggregate model may write (T=\sum_{i=1}^{N}X_i). The severity law supplies one ingredient. Assuming that the (X_i) are identically distributed, independent, and independent of (N) is an additional modeling decision. A shared catastrophe, latent environmental driver, or contract-induced selection may invalidate one or more of these assumptions.
Loss can also be defined at several observational levels. A single physical event may produce several coverage claims, each covered claim may contain multiple payment transactions, and a reinsurance treaty may aggregate claims before the attachment point applies. Therefore, “individual claim severity” identifies the stochastic unit only after the study specifies whether it concerns event, coverage, claim, payment, occurrence, or treaty aggregation.
Exposure, inflation, mixture, and dependence
A portfolio has an exposure distribution as well as a loss distribution. Property value, sum insured, revenue, bodily injury jurisdiction, claims inflation, policy wording, and catastrophe hazard all influence potential severities. If a nominal loss is (X_t) at time (t) and a defensible monetary inflation index is (I_t), an illustrative adjustment to base period (t_0) is (X_{t\to t_0}=X_t I_{t_0}/I_t). This adjustment is an accounting or modeling convention; a general price index may fail to represent legal damages inflation, building replacement costs, medical-cost inflation, or catastrophe-demand surge.
When loss populations are mixed, the unconditional distribution may be written (F(x)=\sum_j w_j F_j(x)) for component weights (w_j\geq0) summing to one. A mixture of claim types can show a slowly decreasing empirical survival curve even when no single component has been separately identified. Conversely, a portfolio that adds high-limit policies can show an apparent tail increase because the insured exposures have changed. An observed large-loss trend cannot be attributed to the physical severity mechanism until the exposure and contractual composition are reconciled.
For a positive scale factor (A), multiplying a loss by (A) changes its scale and may change its tail if the multiplier is itself random and sufficiently extreme. Fixed inflation factors and deterministic unit conversions preserve the distinction between light and regularly varying heavy tails. Across decades, inflation, policy limits, and reporting thresholds may change simultaneously, making direct comparisons of raw amounts unreliable.
Empirical observation and estimand design
The observed variable is usually better represented as (O=h(X,D,L,R,V)), where (D) is a deductible or attachment, (L) is a policy limit or exhaustion point, (R) captures reporting and claims selection, and (V) represents valuation and development. The function (h) may be discontinuous, may map a range of gross losses to the same capped payment, and may suppress losses under a threshold. Statistically, the investigator must establish whether the available values are exact gross losses, exact payments, censored claims, or selected observations.
An actuarial dataset with a reporting minimum is often left-truncated: values below the minimum are not recorded at all. A record showing a payment at an applicable policy limit is often right-censored for the underlying gross severity: the model learns that gross loss was at least as large as the exhaustion threshold, but not its exact magnitude. An observed capped payment is nevertheless known exactly as a payment. The same row can therefore be an exact observation for a payment distribution and a censored observation for a gross-loss distribution.
This distinction is the premise of the empirical investigation by Zhao, Upretee, and Yu (2023), whose model selection study of truncated and censored losses examines how claim thresholds and policy limits alter parametric comparison. Their setting illustrates why estimating parameters from recorded paid amounts as if they were complete gross severities can systematically misstate the tail.
Scope of inference
A likelihood is conditional on a specified sampling mechanism. If only paid claims are available, it can identify the distribution of positive payments conditional on selection under explicit assumptions. Recovering the distribution of all underlying events additionally requires information about unreported or zero-payment claims, exposure counts, and the thresholds generating those omissions. When different severity models fit the observed region similarly but diverge above the highest observed or insured value, their far-tail predictions remain model-dependent.
The appropriate object of inference is consequently a documented pair: a target quantity and a data-generating mechanism. Examples include the gross-loss 99.5th percentile for exposures of a given kind, expected insurer payment per reported claim under a specified deductible and limit, or expected cost of a particular nonproportional reinsurance layer. The calibration procedures and permissible conclusions differ across these examples.
Lognormal and Gamma Families
Gamma severity and exponential tails
For shape (k>0) and scale (\theta>0), a gamma density is
[ f(x;k,\theta)=\frac{x^{k-1}e^{-x/\theta}}{\Gamma(k)\theta^k},\qquad x>0. ]
Its mean is (k\theta), its variance is (k\theta^2), and its moment-generating function is (M_X(t)=(1-\theta t)^{-k}) for (t<1/\theta). Thus a gamma distribution has some finite positive exponential moments and an exponentially declining upper tail. Its shape can capture positive skewness in moderate losses without implying Pareto-like catastrophic probabilities. The NIST gamma reference independently confirms the shape-scale parameterization, density, mean, and variance.
For integer (k), the gamma survival is a finite exponential-polynomial expression. More generally the upper tail behaves, up to a positive constant, like (x^{k-1}e^{-x/\theta}). This distinction becomes practically important at high attachment points. Small or medium losses can be fitted closely by a gamma and a heavy-tailed alternative while the implied probability of an exceptionally large loss differs by orders of magnitude. Calibration at the center of the distribution does not guarantee accurate tail extrapolation.
The exponential family is convenient for actuarial estimation because many likelihoods and expected payments are tractable. Convenience is not a factual reason to impose an exponential tail on a liability or catastrophe portfolio. The correct choice depends on the data-generating process, the size of the relevant insured layer, and the available evidence about high severities.
Lognormal severity and all finite polynomial moments
For (\mu\in\mathbb{R}) and (\sigma>0), let (\log X\sim N(\mu,\sigma^2)). Then
[ f(x)=\frac{1}{x\sigma\sqrt{2\pi}}\exp!\left[-\frac{(\log x-\mu)^2}{2\sigma^2}\right],\qquad x>0, ]
[ F(x)=\Phi!\left(\frac{\log x-\mu}{\sigma}\right),\quad E[X^r]=\exp!\left(r\mu+\frac{r^2\sigma^2}{2}\right)\quad(r>0). ]
The median is (e^\mu), the mean is (e^{\mu+\sigma^2/2}), and the variance is ((e^{\sigma^2}-1)e^{2\mu+\sigma^2}). These identities agree with the NIST lognormal distribution formulas. Its quantile is (Q(p)=\exp[\mu+\sigma\Phi^{-1}(p)]) for (0<p<1).
The lognormal has finite moments of every strictly positive polynomial order but lacks a finite moment-generating function at any positive argument. Under the standard exponential-moment criterion it is a heavy-tailed distribution. It is not regularly varying with a finite Pareto index: its tail eventually decays faster than any fixed power (x^{-a}), yet slower than an exponential (e^{-cx}). The words “heavy tail” and “infinite variance” are therefore not synonyms. A lognormal distribution can produce very large observations and substantial upper-quantile sensitivity even though its variance and all other positive integer moments exist.
An important operational implication follows. If two candidate models both fit a limited range of observed claim severities, one lognormal and one Pareto, a tail index or a finite-moment conclusion cannot be transferred from one candidate to the other. Choosing lognormal for the body and Pareto for the far tail requires an explicit threshold, a continuity or normalization treatment, and a credible basis for extrapolating the tail component.
Weibull and other intermediate cases
A Weibull distribution with survival (S(x)=\exp[-(x/\lambda)^a]), (x\geq0), illustrates that the term “heavy” depends on a mathematical definition. At (a=1) it is exponential; at (a>1) its tail is lighter than exponential; and at (0<a<1) its stretched-exponential tail is heavy according to the absence of a positive exponential moment. Yet every positive polynomial moment exists for every (a>0). A Weibull with (a<1) and a lognormal may both be broadly called heavy-tailed without possessing a common Pareto tail exponent.
This classification matters for stress testing because positive-moment existence determines the legitimacy of means and variances, while the full survival curve determines the cost of limited coverages and upper layers. Different tail labels compress information needed for pricing. A useful model register therefore records the survival formula, support, parameterization, data basis, and targeted extrapolation range rather than only a light/heavy indicator.
Model selection for distinct insurance purposes
A severity distribution can be calibrated for individual paid-loss prediction, unlimited gross-loss reconstruction, reserve simulation, high-attachment pricing, or stress testing. These purposes imply different loss functions. If pricing depends primarily on (\int_a^b S(x),dx), a good fit near the median supplies limited evidence about the integral over a remote layer. If the use is claim-by-claim predictive scoring in a region containing most records, a model selected solely on the largest observations may perform poorly for ordinary payments.
Good practice combines mathematical admissibility with targeted empirical diagnostics: log-scale survival plots, threshold exceedance plots, quantile-quantile comparisons, held-out predictive performance where feasible, likelihood evaluation under the correct observation process, and stability analyses across contract cohorts. Model evaluation should include material parameter and distributional uncertainty. The U.S. Actuarial Standards Board's ASOP No. 56 on modeling establishes relevant professional expectations concerning intended model use, data, assumptions, validation, limitations, and reporting for actuaries within its jurisdiction. It is a methodological governance reference here, not a worldwide legal rule.
A lower information criterion on a pooled fit does not establish that a candidate correctly characterizes a far tail. With thousands of ordinary payments and only a few severe claims, the central likelihood may dominate selection. The analysis should therefore report both overall fit and the consequences of model choice for the specifically insured layers, while avoiding an arbitrary claim that either model must be universally preferred.
Pareto and Heavy Tails
Definitions of tail weight
A nonnegative distribution is heavy-tailed in the exponential-moment sense when (E[e^{tX}]=\infty) for every (t>0). This broad mathematical class includes lognormal, some Weibull, and Pareto laws. A stronger, more structured class is regularly varying survival: for an index (\alpha>0),
[ \lim_{x\to\infty}\frac{S(tx)}{S(x)}=t^{-\alpha}\qquad\text{for each fixed }t>0. ]
Equivalently, (S(x)=x^{-\alpha}L(x)) with a slowly varying (L), meaning (L(tx)/L(x)\to1). Regular variation identifies a tail scaling structure, not a full parametric distribution across all claim sizes. It provides strong asymptotic conclusions under its assumptions, while finite observations can only test compatibility with those assumptions over an observed range. Resnick (2007) develops the distinction between such probability limits and their statistical estimation in probabilistic and statistical heavy-tail modeling.
Under regular variation with positive index (\alpha), moments of order (r<\alpha) exist, while moments of order (r>\alpha) diverge. At (r=\alpha), existence depends on the slowly varying factor and requires separate analysis. For a pure Pareto survival (S(x)=(x_m/x)^\alpha), by contrast, the boundary (r=\alpha) also diverges. That distinction prevents the accidental extension of a Pareto-specific boundary statement to all regularly varying distributions.
Regular variation, Pareto structure, and moments
Consider the Pareto type I distribution with minimum loss (x_m>0), shape (\alpha>0), and survival
[ S(x)=\begin{cases}1,&0\leq x<x_m,\(x_m/x)^\alpha,&x\geq x_m.\end{cases} ]
Its density is (f(x)=\alpha x_m^\alpha x^{-\alpha-1}) for (x\geq x_m), and its quantile is (Q(p)=x_m(1-p)^{-1/\alpha}). Direct integration gives
[ E[X^r]=\frac{\alpha x_m^r}{\alpha-r}\quad\text{if }0<r<\alpha; \qquad E[X^r]=\infty\quad\text{if }r\geq\alpha. ]
Thus the mean exists precisely when (\alpha>1), and the variance exists precisely when (\alpha>2). For (1<\alpha\leq2), expected gross severity is finite but variance is infinite; for (\alpha\leq1), even the expected unlimited gross loss is infinite under the idealized unbounded Pareto assumption. This is a property of the chosen mathematical model, not an assertion that physical insured property or human harm must possess literally unlimited attainable values.
The mean residual loss above (u\geq x_m) is (E[X-u\mid X>u]=u/(\alpha-1)) for (\alpha>1). Its linear growth with threshold is a distinctive feature of the pure Pareto model. The mean excess plot can be diagnostically useful, but changes in policy limits and sampling selection can also distort its apparent slope.
Subexponential losses and catastrophic concentration
The subexponential property expresses the tendency of a large sum to be driven asymptotically by one extreme summand. For independent identically distributed nonnegative claims with a subexponential severity law, (\Pr(X_1+X_2>x)\sim2\Pr(X>x)) as (x\to\infty). Under appropriate extensions, (\Pr(\sum_{i=1}^nX_i>x)\sim nS(x)) for fixed (n). The class includes regularly varying positive losses and the lognormal. This asymptotic relation concerns extreme thresholds and does not justify replacing full aggregate calculations by a one-large-claim approximation at ordinary practical percentiles.
For a Pareto with (\alpha>1), the fraction of total expected gross loss attributable to the largest fraction (q) of events is (q^{1-1/\alpha}). This follows from (E[X\mathbf{1}{X>Q(1-q)}]/E[X]) by direct integration. At (\alpha=1.5), the largest one percent accounts for (0.01^{1/3}\approx21.54%) of expected gross loss. At (\alpha=3), it accounts for (0.01^{2/3}\approx4.64%). This comparison demonstrates sensitivity to the Pareto exponent; it does not infer a specific portfolio's concentration without data.
Composites, mixtures, and asymptotic boundaries
A body-tail model may use a lognormal form below a threshold (u) and a Pareto-type tail above it. Cooray and Ananda (2005) investigated a composite lognormal–Pareto model for actuarial loss payments, motivated by the competing merits of lognormal central fit and Pareto upper-tail behavior. Scollnik (2007) subsequently analyzed limitations and modifications of composite models, emphasizing that joining attractive distributional pieces requires valid normalization, smoothness decisions, and estimation properties. These publications document an established model family, not a newly introduced proposal in this dossier.
If the body density is (f_B) on ((0,u]), the tail is (f_T) on ((u,\infty)), and the tail mass is (\pi=\Pr(X>u)), a properly normalized piecewise model can be written (f(x)=(1-\pi)f_B(x)/F_B(u)) below (u) and (f(x)=\pi f_T(x)/S_T(u)) above (u). Continuity or differentiability at (u) can impose further constraints. If parameters are fitted separately without carrying through the mixing mass and normalization, the result may fail to integrate to one or may alter the intended tail probability.
The choice of (u) is consequently substantive. A low threshold improves sample size but can bias the tail approximation; a high threshold improves asymptotic plausibility but increases parameter uncertainty. Threshold diagnostics, sensitivity over a range of choices, and the number of independent exceedances should accompany any reported tail estimate.
High-threshold excess modeling
For a high threshold (u), define the conditional excess (Y=X-u\mid X>u). A generalized Pareto distribution (GPD) has survival (\Pr(Y>y)=(1+\xi y/\beta)^{-1/\xi}), for scale (\beta>0) and arguments satisfying (1+\xi y/\beta>0). The limit (\xi\to0) yields (e^{-y/\beta}). For (\xi>0), its tail is regularly varying with index (\alpha=1/\xi). Its mean is (\beta/(1-\xi)) when (\xi<1); its variance exists when (\xi<1/2) and equals (\beta^2/[(1-\xi)^2(1-2\xi)]).
The threshold-excess basis is associated with Balkema and de Haan's residual-life limit analysis and Pickands's extreme order-statistics inference. McNeil (1997) applied this foundation to the tails of insurance loss severity distributions, particularly the estimation of high-excess reinsurance layers. These sources delimit the asymptotic result: an approximate GPD at a high threshold is justified by domain-of-attraction assumptions, not by the mere presence of some large claims.
The GPD should be fitted to excess observations under an appropriate conditional likelihood, including relevant deductibles and censoring. An extrapolated expected excess is undefined under the fitted unbounded model when (\xi\geq1), while an actual capped policy layer may still have a finite expected payment. This contrast becomes central in the discussion of coverage modifications.
Truncation and Censoring
How insurance data are selected
Four observation mechanisms are regularly conflated. Complete observations give the actual value of (X). Left truncation at (d) makes a value observable only if (X>d); the sample lacks every omitted loss. Left censoring instead includes a record that the value was at most a known boundary. Right censoring at (u) includes the information (X\geq u) without the exact value above it. Interval censoring supplies only (a<X\leq b). A contractual payment transformation may combine selection with a point mass at the maximum payment.
The distinctions are about information. If a register contains only values above a deductible, the event count below the deductible is unavailable from that file. If a record contains a limit payment with a contractual maximum, the payment is exactly known while the gross loss is generally known only to exceed the exhaustion threshold. If gross losses above a records-retention cutoff were deliberately excluded, the file is right-truncated, not right-censored; no row remains to signal their occurrence. The distinction determines the normalizing denominator in the likelihood.
The practical implications of such mechanisms are analyzed by Zhao, Upretee, and Yu (2023) in risk-model selection under truncation and censoring. Their contribution is particularly relevant because common goodness-of-fit procedures applied to the visible payment numbers can be misleading when the underlying observation mechanism is ignored.
Conditional likelihoods
Let (f_\theta), (F_\theta), and (S_\theta) denote a parametric gross severity law. For an exactly observed loss (x>d) from a sample left-truncated below (d), the conditional likelihood contribution is
[ \ell(\theta;x,d)=\log f_\theta(x)-\log S_\theta(d). ]
The denominator is the probability of inclusion. It is not an optional weight. Omitting (-\log S_\theta(d)) incorrectly treats selected large losses as if they were independent unselected observations from the complete population.
If observations are left-truncated at (d) and right-censored at (u>d), let (\delta_i=1) denote an exact value in ((d,u)), and (\delta_i=0) a known exceedance (X_i\geq u). Then, conditional on entry,
[ L(\theta)=\prod_{i=1}^{n}\left[\frac{f_\theta(x_i)}{S_\theta(d)}\right]^{\delta_i} \left[\frac{S_\theta(u)}{S_\theta(d)}\right]^{1-\delta_i}. ]
For heterogeneous contracts, replace (d,u) by the individually known (d_i,u_i). This likelihood assumes that the entry and censoring thresholds are known and that their mechanism is noninformative conditional on modeled covariates. If policy terms were selected in response to unobserved risk characteristics, or if claims handling caused dependent censoring, that assumption needs explicit reconsideration.
If the sample instead contains only actual (X) values in ((d,u]) and all values above (u) are absent, its doubly truncated likelihood uses (f_\theta(x)/[F_\theta(u)-F_\theta(d)]). A censored observation contributes survival probability; a truncated-out observation contributes no row but changes the sample's conditional normalization. The two operations cannot be substituted without changing the implied data-generating process.
For interval observation (a<X\leq b), the unconditional contribution is (F_\theta(b)-F_\theta(a)), subject to relevant truncation normalization. For a positive payment known only within a reporting band, this probabilistic interval should be modeled explicitly instead of replaced mechanically by its midpoint. Such replacement can erase upper-tail information and misrepresent the number of threshold exceedances.
Distinguishing capped observation from truncation
Suppose a policy has per-loss deductible (d) and maximum insurer payment (c>0). Write its gross exhaustion threshold (u=d+c) and payment (P=\min{(X-d)_+,c}). A payment register containing only positive payments observes (P\mid P>0), corresponding to (X\mid X>d), transformed by the contract. It has a continuous density on ((0,c)) and a point mass at (c). Specifically,
[ g(p)=\frac{f(d+p)}{S(d)},\quad 0<p<c, \qquad \Pr(P=c\mid P>0)=\frac{S(d+c)}{S(d)}. ]
The conditional distribution has a proper total mass: ([F(d+c)-F(d)]/S(d)+S(d+c)/S(d)=1). In the unconditional payment distribution, a mass (\Pr(P=0)=F(d)) must also be included, if the gross-loss distribution includes all nonnegative events. These relationships give an exact internal consistency check for any implemented payment-severity model.
If the maximum payment occurs on a claim because the insurer pays the full limit while a separately recorded gross loss is known, the gross-loss likelihood may be exact for that record. If the file records only the capped amount, treating it as a precise gross severity of (d+c) creates an artificial spike at the exhaustion boundary. That spike is generated by the contract, not by a singularity in the underlying gross loss distribution.
An important qualification concerns claims with multiple coverages, aggregate deductibles, reinstatements, and proportional coinsurance: the simple one-deductible one-limit payment formula must be replaced by the contract's actual loss-to-payment mapping. The principles of conditional likelihood and state-dependent truncation remain applicable, but the payment density may be more complicated or contain additional atoms.
Inference under heterogeneous contracts
For observation (i), the relevant variables may include gross severity (X_i), deductible (d_i), payment cap (c_i), claims inflation factor, exposure class, policy date, jurisdiction, and settlement development. With known exact loss below exhaustion, the likelihood contribution is (f_\theta(x_i\mid z_i)/S_\theta(d_i\mid z_i)). If the claim is limited, it is (S_\theta(d_i+c_i\mid z_i)/S_\theta(d_i\mid z_i)). A single pooled threshold is valid only if contracts share the corresponding selection rule or the transformation has been standardized correctly.
Statistical dependence can arise when high-risk insureds choose higher deductibles, when insurers impose lower limits on high-hazard property, or when claims are allocated across several policies. Correctly incorporating observed contract covariates may reduce bias; it does not automatically identify unobserved selection. A credible study should record the selection process, underwriting vintage, and reasons thresholds changed over time.
Where thresholds vary, one can compare the observed fraction of exact and limit-censored payments against each fitted model's conditional probability (\Pr(X\geq d_i+c_i\mid X>d_i)). An apparent match to the uncensored amount histogram is weak evidence if the model predicts a systematically wrong fraction of limit claims. Evaluating that mass directly links model criticism to the information actually retained in the database.
Model criticism and validation
At least four checks should be distinguished. Data reconciliation verifies that counts, gross-to-net transformations, currencies, missingness, and policy terms agree with the source systems. Statistical diagnostics compare the fitted conditional distribution with observed exact and censored outcomes. Tail diagnostics focus on survival and exceedance behavior over a documented observable range. Financial diagnostics compare the model-implied expected payments and layer costs against the quantities for which the model is intended.
A conventional Kolmogorov–Smirnov statistic on a sample with a mass at the insurance limit is not directly interchangeable with a continuous-gross-loss test. Parameters estimated from the same sample further change the null distribution of such tests. A parametric bootstrap can simulate the full observation process—including each policy's thresholds—refit the model, and compare appropriate statistics; this procedure must state its computational and identification assumptions.
Clauset, Shalizi, and Newman (2009) show in power-law model evaluation why fitting a straight line to a log-log plot can produce unreliable tail exponents and why likelihood, goodness-of-fit, and alternatives should be considered together. This finding applies strongly to insurance samples, where observed sizes are often truncated, censored, inflated, or pooled across changing contracts. No simple goodness-of-fit test alone establishes the correct extrapolation far beyond the observed maximum.
Validation should also examine gross loss versus paid loss, attritional versus severe loss, claims with and without limits, and separate exposure cohorts. Simpler models may be appropriate in well-observed regions, while a tail-sensitive application may require a documented ensemble or interval of plausible distributions. Actuarial communication standards such as ASOP No. 41 emphasize reporting relevant assumptions and limitations within their jurisdiction.
Moment Existence and Quantiles
Existence of moments and valid pricing functionals
For (X\geq0), the survival integral identities are
[ E[X]=\int_0^\infty S(x),dx, \qquad E[X^r]=r\int_0^\infty x^{r-1}S(x),dx\quad(r>0), ]
where each identity is valid as an extended nonnegative integral, including the possibility of infinity. For a Pareto survival (S(x)\propto x^{-\alpha}), the integrand for the (r)-th moment scales as (x^{r-\alpha-1}), which is integrable at infinity exactly when (r<\alpha). This derivation establishes the previously stated moment conditions directly from the tail.
An infinite model variance prevents variance-based standard errors or normal approximations from being justified under their ordinary finite-variance premises. It does not mean that every probability or expected insurance payment is undefined. A gross Pareto with (\alpha=1.5) has finite (E[X]) and infinite (E[X^2]), while every per-occurrence insurer payment capped at (c) obeys (0\leq P\leq c), and hence (E[P^r]\leq c^r<\infty) for every (r>0). The bounded payment is still connected to a potentially extremely heavy-tailed gross loss.
A compound annual aggregate made up of capped payments is not necessarily bounded if the annual claim count is unbounded. If (N) is independent of identically distributed capped payments (P_i) and (E[N]<\infty), then (E[\sum_{i=1}^N P_i]=E[N]E[P]). For its variance one additionally needs corresponding finite second-moment conditions on (N); the usual independent compound formula is (E[N]\operatorname{Var}(P)+\operatorname{Var}(N)(E[P])^2). These are assumptions about aggregate risk and frequency, not properties of an individual capped claim alone.
When a model implies infinite unlimited expected loss but the insured contract has a finite limit, the finite contract cost is mathematically meaningful. Whether the unbounded gross-loss tail is a useful extrapolation is a separate empirical question. The selection of an unbounded distribution should be reported as an assumption about the mathematical support rather than evidence of unbounded physically realizable losses.
Tail quantiles and expected shortfall
The quantile at probability level (p\in(0,1)) is (\operatorname{VaR}_p(X)=\inf{x:F(x)\geq p}). For continuous distributions with integrable upper tails, expected shortfall can be expressed as (\operatorname{ES}_p(X)=E[X\mid X>\operatorname{VaR}_p(X)]); more generally, the quantile-integral definition ((1-p)^{-1}\int_p^1\operatorname{VaR}_q(X),dq) handles atoms without ambiguity. For an unbounded Pareto type I with (\alpha>1),
[ \operatorname{VaR}_p(X)=x_m(1-p)^{-1/\alpha},\qquad \operatorname{ES}_p(X)=\frac{\alpha}{\alpha-1}\operatorname{VaR}_p(X). ]
At (\alpha\leq1), the Pareto expected shortfall is infinite under the unbounded assumption. The quantile is nevertheless finite for every fixed (p<1). A high quantile gives a threshold, whereas expected shortfall evaluates the severity of outcomes beyond that threshold; the latter is more sensitive to assumptions about the far tail.
For a bounded payment (P=\min{(X-d)_+,c}), every quantile is at most (c), and (\operatorname{ES}_p(P)\leq c) regardless of the gross Pareto moment boundary. A point mass at the limit may cause (\operatorname{VaR}_p(P)=c) over an entire interval of confidence levels. If an analyst reports a smooth unbounded payment quantile through that interval, the model may be describing gross losses rather than the contractually bounded payment.
McNeil, Frey, and Embrechts (2015) systematically connect quantile and tail risk measurement in Quantitative Risk Management. Their broader treatment supports careful separation of distributional tail behavior from risk-measure definitions. It does not remove the need to verify that the inputs actually represent gross or net losses as intended.
Reproducible Pareto sensitivity analysis
Set (x_m=1) monetary unit and consider three hypothetical Pareto type I shapes (\alpha\in{0.9,1.5,3}). Each model has the same minimum and median-based structure within its own parameterization but radically different moment and layer behavior. The table reports results obtained from the displayed analytic formulas, not estimates from historical claims.
| Pareto shape (\alpha) | Gross mean | Gross variance | (\Pr(X>10)) | Gross 99th percentile | Gross 99% expected shortfall |
|---|---|---|---|---|---|
| 0.9 | Infinite | Infinite | 0.125893 | 166.810054 | Infinite |
| 1.5 | 3.000000 | Infinite | 0.031623 | 21.544347 | 64.633040 |
| 3.0 | 1.500000 | 0.750000 | 0.001000 | 4.641589 | 6.962383 |
For (\alpha=3), (E[X^2]=\alpha x_m^2/(\alpha-2)=3), so (\operatorname{Var}(X)=3-(1.5)^2=0.75). For (\alpha=1.5), the mean (3) is exact while (E[X^2]) diverges. These values are analytically generated. Reporting them as observed actuarial loss statistics would incorrectly convert model implications into empirical facts.
The 99th percentile grows from about 4.64 to 21.54 monetary units when the exponent changes from 3 to 1.5, although the underlying minimum remains one. The expected shortfall grows still more sharply and is undefined at (\alpha=0.9) under the unbounded model. This is why selection and uncertainty in the exponent are economically decisive for remote attachment points, even when model differences around the typical claim look modest.
Estimating the upper tail and the limits of inference
For an independent exact sample from a Pareto type I tail above a known minimum (u), the maximum-likelihood estimate of the shape parameter is
[ \widehat\alpha=\frac{n}{\sum_{i=1}^n\log(x_i/u)}. ]
If the sample consists of exact exceedances above (u) but some are right-censored at known (v>u), write (m) for the number of exact exceedances, (n_c) for censored observations, and (x_j) for exact losses. Under a common censoring point and the conditional Pareto law, the estimate becomes
[ \widehat\alpha=\frac{m}{\sum_{j=1}^{m}\log(x_j/u)+n_c\log(v/u)}, ]
provided the denominator is positive and (m>0). This follows by differentiating a log likelihood in which exact losses contribute (\log\alpha-\alpha\log(x/u)), up to terms independent of (\alpha), and censored losses contribute (-\alpha\log(v/u)). The formula shows directly how treating capped payments as exact gross observations would miscount the number of uncensored observations and misestimate the exponent.
Hill's (1975) tail-index inference method uses upper order statistics. In a regularly varying model, the Hill estimate of (\gamma=1/\alpha) can be written (\widehat\gamma_k=k^{-1}\sum_{j=1}^k\log(X_{n-j+1:n}/X_{n-k:n})), where (k) is the chosen number of upper observations. Its meaning depends on independent observations, an appropriate upper-tail regime, and selection of (k). Standard Hill estimation applied naively to limit-censored payments or a lognormal body as though they were exact Pareto extremes can be seriously misleading.
The empirical literature provides both evidence for tail-focused methods and warnings about interpretation. McNeil's large fire-loss study is grounded in a well-known Danish insurance experience. Documentation for the Danish reinsurance claim dataset describes 2,167 recorded fire losses for 1980–1990 in one standardized dataset representation. Other analyses use filtered subsets, so observations counted in one publication cannot automatically be equated with the count in a later software package. This monograph draws no fitted parameter or claimed replication from those records.
Finite claim samples do not establish that a moment is truly infinite. A fitted tail exponent is model-dependent and uncertain, and an extrapolation far beyond the top observed loss relies on structural stability not observable in those records alone. Robustness analysis should therefore compare alternative threshold choices, plausible distribution families, censoring assumptions, exposure inflation, and the resulting expected layer costs. The existence of an accurate likelihood on the observed region is necessary but insufficient for reliable far-tail risk claims.
Coverage-Modified Severity
Deductibles, limits, and exact expected payments
Let a positive gross loss (X\geq0) have survival (S). For a per-occurrence deductible (d\geq0) and a maximum insurer payment (c>0), define
[ P=\min{(X-d)+,c}= (X-d)+-(X-d-c)_+. ]
Using Tonelli's theorem for nonnegative integrands and the layer-cake representation, the unconditional expected insurer payment per gross-loss event is
[ E[P]=\int_d^{d+c}S(x),dx. ]
If the insurer instead pays all excess above (d), without a limit, the expected payment is (\int_d^\infty S(x),dx), which may diverge. For a reinsurance occurrence layer attaching at (a) and exhausting at (b>a), the expected layer loss is
[ E[(X-a)+-(X-b)+]=\int_a^b S(x),dx. ]
These identities do not require a density. They apply to continuous, mixed, and discrete nonnegative gross-loss laws, with the appropriate event unit and contract definition. A second-moment identity for a bounded payment is (E[P^2]=2\int_0^c yS(d+y),dy), and (\operatorname{Var}(P)=E[P^2]-E[P]^2). Both are finite for finite (c), even if the corresponding gross moments diverge.
If the dataset consists only of claims with positive payment, the expected positive payment is
[ E[P\mid P>0]=\frac{\int_d^{d+c}S(x),dx}{S(d)},\qquad S(d)>0. ]
If a reporting or coverage system excludes claims with (P=0), using the conditional mean as the unconditional expected cost per event will inflate the model-implied cost unless the positive-payment probability is also taken into account. Conversely, using the unconditional mean to price per positive claim will understate that conditional quantity. The appropriate frequency and severity pairing depends on whether the count variable represents gross-loss events, reported claims, or positive payments.
A distinctive analytical result is that policy transformation changes several risk descriptors at once. A deductible introduces a mass at zero, a limit introduces a mass at the maximum payment, and both change the positive-payment conditional distribution. Yet the underlying gross survival remains the determinant of every limited expected value. A model can therefore fit the distribution of paid amounts while leaving the far gross tail poorly identified beyond the highest contractual limit. The fitted payment distribution and inferred unlimited gross distribution should be reported separately.
Illustrative layer valuation
For a hypothetical Pareto type I with (x_m=1) and (\alpha=1.5), consider a gross-loss reinsurance layer attaching at (a=10) and exhausting at (b=100), all amounts in arbitrary consistent monetary units. Because (S(x)=x^{-1.5}) for (x\geq1),
[ E[(X-10)+-(X-100)+] =\int_{10}^{100}x^{-1.5},dx =2(10^{-1/2}-100^{-1/2}) \approx0.432456. ]
The unlimited attachment loss has expectation (E[(X-10)+]=2/\sqrt{10}\approx0.632456). The expected part above the layer's exhaustion is (E[(X-100)+]=0.2). Their difference is the finite layer cost shown above. This is a decomposition of expected severity, and it excludes claim frequency, exposure, expenses, catastrophe clustering, reinsurer credit risk, and loading for uncertainty or required capital.
The probability the attachment is reached is (S(10)=10^{-1.5}\approx0.031623), while the probability of complete exhaustion is (S(100)=100^{-1.5}=0.001). Conditional on a claim reaching the layer, the expected layer payment is (0.432456/0.031623\approx13.675) monetary units. The unconditional event expectation and conditional-on-attachment mean are different estimands; neither is an annual treaty premium without an associated exposure and occurrence process.
Consider changing the upper limit from 100 to 1,000 while retaining (a=10). The expected layer loss becomes (2(10^{-1/2}-1000^{-1/2})\approx0.569210), substantially higher than the 10-to-100 layer because the heavy upper tail retains material mass. A lighter-tailed alternative could imply a different increase. This is a mathematical demonstration of how a gross tail assumption propagates into the marginal cost of additional policy or reinsurance capacity.
For (\alpha=0.9), the unlimited Pareto expectation is infinite; nevertheless, the finite 10-to-100 layer has finite expected cost ((100^{0.1}-10^{0.1})/0.1\approx3.259678). The proper inference is that unlimited gross loss and finite insured layer cost respond differently to the same mathematical tail. If physical exposures impose a true maximum, that limit should enter the gross-loss support itself and be distinguished from a contractual payment maximum.
Reporting frequency and aggregate loss
Let (N_g) denote the count of all gross loss events and suppose (X_1,\ldots,X_{N_g}) are severity observations for those events. The count of positive insurer payments is (N_p=\sum_{i=1}^{N_g}\mathbf{1}{X_i>d}). When gross events are independent and identically distributed, and the count process is suitably independent of their sizes, (E[N_p]=E[N_g]S(d)). The expected aggregate insurer payment is then (E[N_g]E[P]=E[N_p]E[P\mid P>0]), provided these expectations are finite. This identity is a powerful reconciliation test linking frequency definitions, severity selection, and expected cost.
If contracts have heterogeneous deductibles, (S(d)) must be replaced by a policy-specific or appropriately weighted exposure average. If large losses produce correlated claim counts, as with a catastrophe affecting many insured locations, a product of separate marginal averages may be inadequate for the full aggregate distribution even if the expectation identity can be established conditionally. Contract aggregation clauses and occurrence definitions may change both the loss and the relevant count.
A working insurer may have claim counts for paid claims but exposure data only for policies. A proper rating or reserving study should map policy exposure to reported event incidence, positive-payment frequency, and paid severity, rather than inserting a conditional positive-payment severity into an aggregate process that counts all events. The resulting bias can persist even if the selected severity family is otherwise correctly estimated.
Reinsurance and portfolio consequences
At a high excess-of-loss attachment, the expected cost is a survival integral across the layer. For large-claim modeling, tail evidence is therefore relevant not simply because “large losses matter” but because the premium-bearing functional assigns all its weight to (S(x)) over precisely the insured interval. McNeil's (1997) high-excess loss-severity analysis motivates threshold methods for this setting. A gross-loss model calibrated only against many small claims can produce reasonable total mean loss while materially underestimating a rare upper layer.
The same distinction shapes retained risk. A primary insurer with a policy limit (c) may have bounded per-claim payments but exposure to a large and stochastic number of claims, inflation, litigation escalation, catastrophe concentration, and reinsurance collectability. A reinsurer may have a different policy transformation at the treaty level, and an aggregate stop-loss treaty introduces an additional cap or deductible on a compound sum rather than on each (X_i). The relevant loss operator must be applied at the contract's stated aggregation level.
For a parametric gross law (S_\theta), layer cost sensitivity can be written (m(\theta;a,b)=\int_a^bS_\theta(x),dx). When a parameter is differentiable under the integral, (\partial m/\partial\theta=\int_a^b\partial S_\theta(x)/\partial\theta,dx). This identity shows exactly which section of the survival curve transmits parameter uncertainty into the expected layer cost. Large uncertainty in regions beyond the exhaust point need not dominate a fully limited layer, whereas uncertainty across the layer itself is directly financial.
Different stakeholders therefore face distinct model risks. The policyholder is concerned with retained loss and adequacy of protection; the insurer with bounded payment plus the distribution of claim frequency and balance-sheet obligations; the reinsurer with attachment-specific losses and contract language; supervisors with capital adequacy, governance, and externalities. The severity model becomes meaningful only when linked to the financial decision and to the contract actually underwritten.
Integrated model governance and reproducibility
A useful severity study produces separate tables of (1) gross severity assumptions, (2) observation mechanisms, and (3) net contract transformations. Those are categories of documented inputs and outputs, not a new theoretical framework. The tables reveal which conclusions derive directly from the chosen probability law and which require assumptions about exposure, reporting, or coverage. A scientifically honest model report then traces each financial output to an integral or conditional probability whose definition remains stable across methods.
The first check is mathematical: the density or mixed distribution integrates to one, the survival is nonincreasing, and moments claimed finite meet their convergence conditions. The second is statistical: the likelihood accounts for known truncation and censoring, fitted tail parameters respect their valid domains, and independent exceedance assumptions are supported or appropriately limited. The third is financial: gross severity, positive payments, limits, and reinsurance layers use the right event unit, price level, and contract mapping. The fourth is communicative: model uncertainty, scope restrictions, benchmark outcomes, and decision sensitivity are disclosed.
For reproducibility, the synthetic Pareto values in this dossier can be recalculated with (S(x)=x^{-\alpha}) for (x\geq1), (Q(p)=(1-p)^{-1/\alpha}), (\operatorname{ES}_p=\alpha Q(p)/(\alpha-1)) for (\alpha>1), and (m(a,b)=(b^{1-\alpha}-a^{1-\alpha})/(1-\alpha)) for (\alpha\ne1), or (\log(b/a)) for (\alpha=1). These formulas specify all inputs, units, moment conditions, and computational transformations. Rounded display values should never replace the exact expressions in independent verification.
A research audit should retain the reference record and exact claim supported by each external link. Source titles alone are insufficient if they do not support the nearby proposition. The scholarly and editorial controls in this work therefore require short contextual links in the chapter prose, author-year attribution, complete bibliography entries, and a separate machine-readable mapping of checked claims to sources. This reporting structure follows the approved Published by Aisentica Development manuscript-intake convention without implying that external peer review or empirical validation has occurred.
The limitations are concrete. This study does not fit a severity distribution to insurer records, recover a uniquely identified tail beyond observed limits, supply full national-law comparison, estimate a real portfolio's solvency capital, model dependence among catastrophe claims, or turn a synthetic layer example into a premium recommendation. Its results are mathematical derivations and evidence-based methodological synthesis with explicit assumptions. A substantive application would require contract-level datasets, observation-process validation, current exposure measures, sensitivity testing, and independent domain-specific governance before its estimates were used operationally.
Conclusion
Loss severity analysis begins with the stochastic unit and the information actually observed. Gamma distributions provide an exponentially declining tail; lognormal distributions combine substantial skewness with all finite positive polynomial moments and no positive exponential moment; Pareto and regularly varying laws permit explicit tail scaling and, depending on their parameters, the nonexistence of expected gross severity or variance. These distinctions determine which analytical quantities are finite and how strongly remote-layer costs depend on model assumptions.
A reporting deductible changes the sampling population, a payment limit changes the observed gross-loss information, and both reshape the paid-loss distribution. The exact truncated and censored likelihoods and the mixed payment law make those effects explicit. Limited expected payments are integrals of the gross survival function across the covered interval, which explains both their dependence on tail estimation and their mathematical finiteness under a finite per-claim limit. A model implying infinite unlimited gross loss can still imply a finite contractual layer cost.
The central analytical synthesis is that gross severity, selected observation, and net payment are separate probability objects joined by explicitly specified operators. Reliable actuarial work preserves those distinctions from data ingestion through likelihood estimation to financial reporting. Tail index estimates, quantiles, expected shortfalls, and reinsurance layer costs should be understood as conditional on their model class, threshold, available range of observations, contract terms, and temporal exposure conditions.
The insurance significance extends beyond choosing a named distribution. A defensible severity analysis records what is observable, what can be identified from it, which tail assumptions are extrapolations, how uncertainty affects relevant financial layers, and how each model is validated against its declared purpose. These findings follow from established stochastic and actuarial theory, verified primary methodological literature, and the reproducible calculations developed in this monograph. They provide a complete methodological foundation for future empirical investigations without claiming a novel probability theory or a portfolio-specific empirical discovery.
References
Actuarial Standards Board. (2010). Actuarial Standard of Practice No. 41: Actuarial Communications. Adopted December 2010; effective May 1, 2011. Official standard.
Actuarial Standards Board. (2019). Actuarial Standard of Practice No. 56: Modeling. Adopted December 2019; effective October 1, 2020. Official standard.
Balkema, A. A., & de Haan, L. (1974). Residual life time at great age. The Annals of Probability, 2(5), 792–804. https://doi.org/10.1214/aop/1176996548.
Clauset, A., Shalizi, C. R., & Newman, M. E. J. (2009). Power-law distributions in empirical data. SIAM Review, 51(4), 661–703. https://doi.org/10.1137/070710111.
Cooray, K., & Ananda, M. M. A. (2005). Modeling actuarial data with a composite lognormal-Pareto model. Scandinavian Actuarial Journal, 2005(5), 321–334. https://doi.org/10.1080/03461230510009763.
Dutang, C. (n.d.). Danish reinsurance claim dataset [CASdatasets reference documentation; publicly available documentation for a widely studied dataset]. Dataset documentation.
Embrechts, P., Klüppelberg, C., & Mikosch, T. (1997). Modelling Extremal Events for Insurance and Finance. Stochastic Modelling and Applied Probability, Vol. 33. Springer. https://doi.org/10.1007/978-3-642-33483-2.
Hill, B. M. (1975). A simple general approach to inference about the tail of a distribution. The Annals of Statistics, 3(5), 1163–1174. https://doi.org/10.1214/aos/1176343247.
Klugman, S. A., Panjer, H. H., & Willmot, G. E. (2008). Loss Models: From Data to Decisions (3rd ed.). Wiley. https://doi.org/10.1002/9780470391341.
McNeil, A. J. (1997). Estimating the tails of loss severity distributions using extreme value theory. ASTIN Bulletin, 27(1), 117–137. https://doi.org/10.2143/AST.27.1.563210.
McNeil, A. J., Frey, R., & Embrechts, P. (2015). Quantitative Risk Management: Concepts, Techniques and Tools (revised ed.). Princeton University Press. Publisher catalog.
National Institute of Standards and Technology. (n.d.). Gamma. In NIST/SEMATECH e-Handbook of Statistical Methods, Section 8.1.6.5. Official handbook.
National Institute of Standards and Technology. (n.d.). Lognormal Distribution. In NIST/SEMATECH e-Handbook of Statistical Methods, Section 1.3.6.6.9. Official handbook.
Pickands, J., III. (1975). Statistical inference using extreme order statistics. The Annals of Statistics, 3(1), 119–131. https://doi.org/10.1214/aos/1176343003.
Resnick, S. I. (2007). Heavy-Tail Phenomena: Probabilistic and Statistical Modeling. Springer. https://doi.org/10.1007/978-0-387-45024-7.
Scollnik, D. P. M. (2007). On composite lognormal-Pareto models. Scandinavian Actuarial Journal, 2007(1), 20–33. https://doi.org/10.1080/03461230601110447.
Zhao, Q., Upretee, S., & Yu, D. (2023). Model uncertainty and selection of risk models for left-truncated and right-censored loss data. Risks, 11(11), Article 188. https://doi.org/10.3390/risks11110188.