Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Quantile and Distributional Regression for Insurance Losses

Abstract. Insurance loss models are decision-relevant only to the extent that their predicted statistical object corresponds to the economic object under consideration. Conditional means support expected-cost calculations; conditional quantiles describe specified portions of a heterogeneous loss distribution; full conditional distributions permit the evaluation of probabilities, moments, tail functionals, retention structures, and aggregate exposure. This monograph examines the mathematical foundations, identification conditions, estimation methods, insurance data-generating mechanisms, diagnostics, and actuarial uses of quantile and distributional regression. It gives particular attention to losses with an atom at zero, exposure-dependent frequency, highly skewed positive severity, heavy tails, evolving claim development, censoring under contract limits, and the non-additivity of quantiles. A self-contained numerical comparison demonstrates that portfolios with the same expected individual loss can have sharply different high quantiles and expected shortfalls. A second derivation establishes the exact mapping between positive-claim severity quantiles and unconditional policy-loss quantiles in a two-part model. The analytical synthesis distinguishes estimation, predictive validation, contractual transformations, aggregation, and ultimate institutional use. It concludes that the evidentiary unit for a loss-distribution model is a specifically defined random variable, horizon, exposure population, information set, and forecast target, together with tested calibration in the relevant tail. The manuscript reviews established research and develops its own comparative deductions; it proposes no new named theory or doctrine.

Keywords: Quantile regression; conditional loss distribution; distributional regression; insurance losses; pure premium; claim frequency; claim severity; GAMLSS; Tweedie model; zero-inflated losses; quantile crossing; heavy tails; value at risk; expected shortfall; probabilistic calibration; actuarial validation; risk adjustment; solvency.

Introduction

An insurance loss is the financial result of a stochastic process mediated by exposure, claims occurrence, event severity, contractual coverage, reporting, settlement, and observation. Each stage can change the probability law of the ultimately recorded amount. Consequently, a regression that estimates the average paid amount and a regression that estimates an upper quantile answer different questions even when their inputs and training observations are identical. The choice of estimand is an essential part of actuarial interpretation.

The foundational regression-quantile formulation of Koenker and Bassett (1978) extends the optimization problem defining an ordinary sample quantile to a conditional regression setting. Koenker's systematic treatment of conditional quantiles shows how covariates can affect more than location, including distributional scale and shape. In a different but complementary tradition, the GAMLSS framework of Rigby and Stasinopoulos (2005) allows multiple parameters of a response distribution to vary with predictors. These approaches provide alternative representations of heterogeneous conditional distributions rather than interchangeable estimators of one universal parameter.

The insurance literature gives concrete motivation. Smyth and Jørgensen's insurance dispersion-modeling analysis explains why aggregate claims cost can require a dispersion regression in addition to a mean regression. Heras, Moreno, and Vilar-Zanón's two-stage ratemaking study explicitly separates the probability of a positive policy loss from its conditional positive amount. Subsequent neural quantile estimation for insurance claims applies this two-part logic to nonlinear predictors. These studies concern particular research designs and datasets; their comparative performance findings do not by themselves establish universal superiority across lines of business.

This monograph treats distributional regression as a family of procedures whose validity depends on the insurance loss being modeled, the loss-generating process, the sample selection mechanism, and the eventual decision. It distinguishes conditional policy losses from claim severity, accident-year totals from accounting observations, paid losses from incurred ultimate losses, and separate exposure quantiles from a portfolio's aggregate distribution. The same distinction governs evaluation: pinball loss measures accuracy for a chosen quantile; a proper distributional score evaluates a complete probabilistic forecast; observed exceedance proportions and calibration curves test consequences not guaranteed by in-sample optimization.

The method is conceptual and mathematical, supported by research-paper and institutional source verification. The explicit numerical examples are original hypothetical constructions, not empirical estimates from an insurer. Source abstracts, publisher records, accessible article text, and official regulatory guidance are distinguished in the accompanying evidence ledger. The international scope concerns transferable statistical methods; product design, censoring, insurance supervision, and accounting obligations must be interpreted under the applicable jurisdiction and contractual form.

Conditional Quantiles and Loss Distributions

The response variable is the first model decision

Let (Y\geq 0) denote a precisely defined monetary loss and (X=x) the information available at prediction time. The conditional cumulative distribution function is (F_x(y)=P(Y\leq y\mid X=x)). For (0<\tau<1), its lower conditional quantile is

[ Q_\tau(x)=\inf{y:F_x(y)\geq\tau}. ]

This definition also works for an insurance loss distribution with atoms. At a discontinuity, the condition (P(Y\leq Q_\tau\mid x)\geq\tau) holds while exact equality may fail. The conditional mean, if it exists, is (\mu(x)=E[Y\mid X=x]). The quantile is a feature of the order of outcomes; the mean is an integral over their values. A complete predictive law permits computing both when the corresponding integrals are finite.

The identity of (Y) matters. It may represent total annual loss for one policy, positive loss on an individual claim, claim count, a limited and deductible-adjusted indemnity, ultimate cost on an accident-year cohort, or one-year change in an insurer's basic own funds. These variables have different supports, dependence structures, reporting processes, and risk uses. A severe claim is typically sampled conditional on a claim having occurred; a randomly selected policy commonly produces zero paid loss. Quantiles learned from only reported positive claims cannot automatically be labeled unconditional policy-loss quantiles.

Observation horizons also matter. A 12-month prospective policy-loss forecast belongs to a specific exposure cohort and rating period. The amount recognized on the financial statements at a valuation date may include past underwriting years and claim development. A modeled annual net loss after reinsurance is not the same random variable as the annual gross loss. Neither a confidence level nor a model class resolves these differences. An explicit estimand must fix the risk unit, currency, inflation basis, loss basis, gross/net perspective, observation and settlement horizon, and permitted information.

A mass at zero and its quantile mapping

A convenient nonnegative policy-loss representation assigns conditional probability (p_0(x)=P(Y=0\mid x)) to zero and uses a strictly positive conditional distribution (G_x(y)=P(Y\leq y\mid Y>0,x)) otherwise. For (y>0),

[ F_x(y)=p_0(x)+[1-p_0(x)]G_x(y). ]

The lower quantile therefore obeys

[ Q_\tau(x)= \begin{cases} 0,&\tau\leq p_0(x),\ G_x^{-1}!\left(\frac{\tau-p_0(x)}{1-p_0(x)}\right),&\tau>p_0(x), \end{cases} ]

when the generalized inverse of (G_x) is used. The mapping is an algebraic consequence of the mixture CDF and requires no additional parametric assumption. It is the central calculation behind two-stage models of unconditional quantiles. The logistic first stage in Heras et al. (2018) is one implementation of the claim-occurrence component. Importantly, its probability error is transmitted nonlinearly to the second-stage quantile, especially when (\tau) is close to (p_0(x)).

Suppose a class of annual policies has an 80% probability of zero payment and, conditional on a positive loss, an exponential distribution with mean 1,000 currency units. Then (E[Y]=0.2\times1{,}000=200), whereas (Q_{0.95}=-1{,}000\log(0.25)\approx1{,}386.29) and (Q_{0.99}=-1{,}000\log(0.05)\approx2{,}995.73). A second hypothetical class that always produces exactly 200 units has the same mean, but both its 95th and 99th quantiles equal 200. The contrast demonstrates the insufficiency of mean equality for risk-ranking purposes. These numbers are exact theoretical results rounded to two decimals; they are not market loss statistics.

For the mixture, (E[Y^2]=0.2\times2\times1{,}000^2=400{,}000) and (\operatorname{Var}(Y)=400{,}000-200^2=360{,}000). For the constant-loss class, the variance is zero. Moreover, the conditional distribution of the amount above (Q_{0.95}) retains the exponential mean-excess of 1,000, so the continuous-distribution expected shortfall is approximately (1{,}386.29+1{,}000=2{,}386.29). This last result depends on the memoryless exponential assumption. It cannot be transferred to lognormal or Pareto severities by analogy.

Frequency, severity, and aggregate cost

An alternative data-generating representation is (S=\sum_{j=1}^{N} Z_j), where (N) is a claim count and (Z_j) are positive claim amounts. Under conditional independence of frequency and iid severities given (X=x), with finite moments,

[ E[S\mid x]=E[N\mid x]E[Z\mid x] ]

and

[ \operatorname{Var}(S\mid x)=E[N\mid x]\operatorname{Var}(Z\mid x)+\operatorname{Var}(N\mid x){E[Z\mid x]}^{2}. ]

These equations separate frequency uncertainty from severity uncertainty, but real insurance portfolios can violate the independence assumptions through shared events, inflation, claims handling, reporting processes, and risk selection. A model for (N) and a model for (Z) can produce an aggregate distribution by compounding or simulation; its quantiles must be computed from that resulting law. The sum of component quantiles is generally different from the quantile of a sum, even with independent components.

Tweedie compound-Poisson/gamma representations offer an alternative for nonnegative, semicontinuous total costs. Smyth and Jørgensen (2002) specifically analyze dispersion in aggregated insurance costs, and Dunn and Smyth (2005) examine numerical evaluation of Tweedie densities. A Tweedie mean-variance relationship (\operatorname{Var}(Y\mid x)=\phi\mu(x)^p), with compound-Poisson/gamma behavior for (1<p<2), is a modeling choice to be tested. Its zero mass, positive-tail shape, and suitability for claim limits cannot be inferred from a reasonable fit to the mean alone.

Conditioning, selection, and identifiable populations

Conditional quantiles are population-specific. A dataset of only closed claims selects on reporting and closure; a dataset of large litigated claims selects on legal escalation; a deductible filters observable payment amounts; a capped policy payment replaces the true underlying severity above the limit with a mass at the cap. Recovering an unobserved ground-up loss distribution requires a justified observation model, exposure data, external information, or explicit partial identification. An apparently stable estimate of paid-loss quantiles may be an excellent model of the recorded payment and a poor model of underlying economic damage.

Another distinction is temporal. At underwriting, only information then available may enter (X). An ultimate loss, settlement date, litigation result, or final reserve cannot be included as a known predictor of a prior-year underwriting decision. At valuation, a broader information set is legitimate, but the target and validation dates change accordingly. Define a forecast as a pair ((\widehat F_{x,t},Y_{t+h})) to preserve this information chronology. Such temporal discipline is a prerequisite for interpreting quantile calibration as predictive skill rather than hindsight.

Quantile Regression Principles

Why the check loss targets a quantile

The classical regression-quantile objective for quantile level (\tau) minimizes the sum of asymmetric absolute deviations. Define the check loss (\rho_\tau(u)=u{\tau-\mathbf{1}(u<0)}), so a positive residual is weighted by (\tau), a negative residual by (1-\tau), and the loss is nonnegative. For a linear specification (q_\tau(x)=x^\top\beta_\tau), estimate

[ \widehat\beta_\tau=\arg\min_{\beta}\sum_{i=1}^{n}w_i\rho_\tau(y_i-x_i^\top\beta), ]

where (w_i\geq0) may represent defensible observation weights. Koenker and Bassett (1978) establish the original linear construction; Koenker (2005) develops its statistical and computational structure. With a continuously distributed response and interior solutions, differentiation of expected check loss with respect to a proposed quantile (q) gives (F_x(q)-\tau). Thus its population minimizer is a conditional quantile. At mass points, subgradient inequalities select valid quantiles, and a flat interval of minimizers can occur.

The shape of the objective matters operationally. Increasing (\tau) makes underprediction more costly than overprediction under check loss. But this is a forecast-evaluation weighting, not a statement about the economic cost of an insurer's pricing error or its utility function. Choosing the 99th conditional quantile because it minimizes 99th-quantile check loss makes a correct statistical target; using that quantile as a premium or capital number requires a further actuarial decision rule.

Identification, specification, and interpretation

A linear quantile relation (Q_\tau(Y\mid X=x)=x^\top\beta_\tau) is a substantive assumption about the quantile surface, not an identity inherited from the mathematical definition. Different (\tau) values may have different coefficients because location, dispersion, and shape can all depend on exposure characteristics. When a predictor's coefficient rises from a central to an upper quantile, one defensible interpretation is a stronger conditional effect in the upper part of the modeled loss distribution. A causal effect would require separate identification assumptions and design: observed rating covariates can reflect risk selection, coverage choice, and reporting rather than exogenous treatment.

The distribution of (X) matters. Rare industrial policies and catastrophe-exposed risks may have little overlap with the common policies on which a flexible quantile learner performs well. Tail extrapolation outside observed support requires assumptions that cannot be validated merely by apparent in-sample fit. Interactions, nonlinear terms, and granular grouping can improve fit but increase variance for small risk cells. A model selected for a quantile at one population mix can fail after underwriting strategy changes that alter the covariate composition.

Observation weighting is especially sensitive to the estimand. Weighting each policy-year equally targets the distribution among policy-years. Weighting proportionally to earned exposure can target a unit-exposure population under suitable sampling assumptions. Weighting by past claim amount intentionally emphasizes severity and changes the population measure. All three can be statistically coherent, but the interpretation of quantile level (\tau) depends on the choice. An insurer should state whether its predicted quantile describes a randomly drawn policy, one unit of insured exposure, one claim, a policy with a selected risk profile, or its aggregate book.

The complete curve and noncrossing restrictions

Estimating several quantile levels independently can yield (\widehat Q_{0.90}(x)>\widehat Q_{0.95}(x)), an impossible ordering for genuine quantiles. Crossing is evidence of incompatible estimated functions at that (x), not a property of the true distribution. Chernozhukov, Fernández-Val, and Galichon's monotone rearrangement method transforms an estimated curve to respect quantile order and establishes formal properties under the paper's conditions. Joint constrained estimation, monotonic quantile-network architectures, and distribution-derived quantiles are other approaches. Rearrangement repairs order, but it does not establish that the repaired tail is calibrated to future extreme losses.

An entire estimated quantile curve, properly nondecreasing over (\tau), can represent the inverse CDF for a fixed (x). Integrals of (Q_\tau(x)) can then recover expectation when integrability holds. In particular, for nonnegative integrable (Y), (E[Y\mid x]=\int_0^1Q_u(x)du), with the usual generalized inverse. In actual estimation, integrating an observed grid of quantiles requires numerical interpolation and extrapolation near 0 and 1, where uncertainty may be greatest. Inferring a mean from an upper-tail-only quantile model is therefore generally underidentified.

Nonlinear and dependent-data estimators

Flexible quantile estimators include spline and additive regressions, gradient boosting with check loss, quantile forests, and neural networks trained on weighted asymmetric deviations. Meinshausen's quantile regression forests use forest-induced weights to estimate conditional distributions and quantiles, reducing reliance on a globally linear predictor. For longitudinal observations, Koenker's penalized panel-quantile analysis addresses the proliferation of individual effects. Hierarchical policy histories, clustered insureds, repeated claims, and geographic catastrophe clusters require inference procedures reflecting dependence, whether through explicit models, block resampling, or robust cluster-based uncertainty estimates.

Quantile-regression machinery developed for continuous responses must be used carefully for insurance claim counts and mixed outcomes. Machado and Santos Silva's quantiles-for-counts study discusses the special discreteness problem and a smoothing strategy. Directly fitting a continuous check-loss model to an atom-heavy response may provide a valid conditional quantile under the generalized definition, while conventional smooth-density asymptotics or differentiability assumptions can fail. Reporting a zero 50th percentile for a zero-heavy portfolio is often statistically correct, although that percentile alone says little about expected cost or the occurrence of severe claims.

Limits of the statistical objective

Quantile regression supplies a conditional distributional functional. It does not guarantee a valid probability distribution over all possible observations, a causal interpretation of covariate differences, stationarity after deployment, or compliance with insurance pricing constraints. Its advantage over an expected-value regression depends on whether the decision uses a quantile, a tail functional, or the whole distribution. A distributional model may be more efficient when many quantiles are needed and the family is well specified; a direct quantile model may be more robust to a misspecified complete density when a particular quantile is the only scientifically relevant target. These are consequences of the different assumptions and objectives, not a universal ranking of model families.

Scale and Dispersion Effects

Conditional heterogeneity beyond the mean

A central empirical problem is that covariates can affect the dispersion and tail thickness of losses while leaving their expected value almost unchanged. Consider a positive loss model (\log Z=m(x)+s(x)\varepsilon), where (\varepsilon) follows a standardized reference distribution. Its quantiles are (Q_\tau(Z\mid x)=\exp{m(x)+s(x)Q_\tau(\varepsilon)}), provided (s(x)>0). The location function (m(x)) shifts log severity; the scale function (s(x)) also changes the separation between lower and upper quantiles. Even simple parametric specifications therefore imply different covariate sensitivities at different levels of (\tau).

For illustration, if (\varepsilon) is standard normal, (Q_\tau(Z\mid x)=\exp{m(x)+s(x)\Phi^{-1}(\tau)}). For (\tau>0.5), increasing (s) at fixed (m) increases the corresponding quantile, although the same change can lower a quantile with (\tau<0.5). Holding the arithmetic mean constant while changing (s) requires changing (m), because (E[Z\mid x]=\exp{m+s^2/2}). The common intuition that increased dispersion must raise every high quantile can fail near the center after such a mean-preserving adjustment; signs should be derived for the particular family and quantile level.

A more flexible formulation writes (Y\mid X=x\sim D{\theta_1(x),\ldots,\theta_K(x)}) with separate link functions (g_k{\theta_k(x)}=\eta_k(x)). The (\eta_k) may combine linear effects, splines, spatial structure, and justified random effects. Rigby and Stasinopoulos's generalized additive distributional models were developed to model location, scale, and shape, subject to the chosen family's meaning and constraints. A recent distributional-regression methods primer reviews the extension to full conditional distributions and diagnostic choices. Modeling shape parameters separately is valuable when skewness, tail thickness, or positive-loss probability differs between risk groups.

The parameterization must remain actuarially interpretable. In a gamma model, linking the shape and scale separately changes how mean and variance depend on predictors. In a lognormal model, a predictor that increases log-scale variance also changes the arithmetic mean unless the location is adjusted. In a generalized Pareto tail, an estimated shape near the moment-existence boundary changes whether a finite mean or variance is mathematically available. Coefficients on a link scale are not direct currency changes, and the meaning of a one-unit covariate increment can be contingent on all other parameters.

Two-part and compound representations

A two-part model may estimate (\pi(x)=P(Y>0\mid x)) by logistic regression and (G_x) for positive losses by a distributional regression or direct quantile regression. Its complete CDF is (F_x(y)=1-\pi(x)+\pi(x)G_x(y)) for (y>0). A fitted high quantile is then a coupled consequence of both components. The decomposition allows the covariates that increase claim probability to differ from those that increase severity, as in the actuarial methods examined by Heras et al. (2018) and Baione and Biancalana (2021). The latter investigate a parametric quantile approach as an alternative to repeatedly fitting many second-stage quantile equations across risk classes.

To see the effect of the first stage, retain an exponential positive-claim severity with mean 1,000. If (\pi=0.20), then (p_0=0.80) and the 95th quantile is 1,386.29. If the occurrence probability falls to (\pi=0.10), then (p_0=0.90) and the same quantile equals (-1{,}000\log(0.50)\approx693.15). The positive-claim severity distribution is identical; the unconditional policy-loss quantile is nearly halved solely by the occurrence probability change. A regression fit only to positive severities could not produce this policy-level distinction without a frequency or occurrence component.

A Tweedie model can encode a point mass at zero and a positive continuous distribution within a single exponential dispersion family. Its compound-Poisson/gamma characterization makes it a natural benchmark for aggregate cost with potentially several claims. Nevertheless, identifying it as a compound law does not solve model comparison. The assumed variance power, distributional fit near zero, probability of multiple claims, potential policy limit, and upper-tail extrapolation require separate checks. Smyth and Jørgensen (2002) emphasize that claims-cost dispersion deserves direct modeling; Dunn and Smyth (2005) document nontrivial Tweedie likelihood computation. A score improvement in the central body of the data may coexist with a high-quantile failure.

Distributional forests, boosting, and neural networks

Prediction methods that estimate conditional distributions can be separated by what they assume. Parametric distributional regression assumes a conditional family and learns its parameters. Quantile forests produce localized empirical outcome weights; each resulting distribution is discrete at training observations unless smoothed. Quantile boosting or neural networks can directly target selected levels via check loss. Other learners optimize a likelihood or proper score for a predictive density, mixture, or discrete-continuous law. The algorithm name does not identify the target: a neural network can output a conditional mean, one quantile, a full parametric law, or an ensemble of laws.

Meinshausen's conditional-distribution forest method illustrates data-adaptive weighting. Laporta, Levantesi, and Petrella's insurance Quantile-CANN research combines an actuarial neural-network approach with quantile targets and a two-part treatment of claims. Empirical improvements reported for their health-insurance case study are a useful research result, but a different insurer may face sparse extreme losses, exposure shifts, and claims inflation not represented by that sample. Any algorithmic comparison must use the same split, target, loss definition, hyperparameter selection protocol, and observed information set.

Regularization is an essential component of valid comparison. If a flexible model learns separate shape functions and numerous interactions from a small upper-tail sample, its in-sample quantiles may be unstable. Penalties, smoothing choices, constrained tails, monotonicity restrictions, hierarchical pooling, and low-dimensional distributional families can reduce variance at the cost of bias. This trade-off should be reported with repeated-sample or blocked validation, not inferred from the training likelihood alone. A large data table can still contain few independent extremes if claims share catastrophe events or a single inflation regime.

Coverage and policy transformations

The insurer often observes an indemnity rather than a ground-up economic loss. For a simple per-claim deductible (d) and payment limit (\ell), payment is (P=\min{\max(Z-d,0),\ell}). This nondecreasing transformation implies the exact quantile identity (Q_\tau(P)=\min{\max(Q_\tau(Z)-d,0),\ell}) for the lower quantile when the same conditional law and deterministic contract parameters are used. Yet the identity does not permit recovery of ground-up quantiles beyond the cap from capped payments: distinct underlying severities can produce the same observed payment distribution above (\ell).

For a policy aggregate with reinstatements, annual aggregate deductibles, shared limits, attachment points, occurrence definitions, or coinsurance, the payment functional acts on a collection of claims and events. Individual-claim transformations cannot simply be applied to an aggregate quantile. More generally, even a monotone indemnity mapping requires correct alignment of occurrence, insured interest, coverage status, limits and currency. Regression on paid loss can incorporate contractual features as covariates, but the resulting model forecasts the recorded payment conditional on those features, not an invariant natural-loss law.

The interpretive consequence

Distributional regression permits a controlled decomposition of risk heterogeneity: occurrence probability, positive severity, dispersion, and tail shape can be assigned to different but connected conditional components. The resulting interpretation is strongest when each component corresponds to an identifiable data process. An apparent insurance-class effect on the 99th quantile may arise from higher claim frequency, a wider severity distribution, a thicker tail, a higher limit, or a reporting difference. An actuarial explanation should identify which mechanism is supported by the evidence before attaching pricing or solvency meaning to the quantile.

Heavy Tails and Limits

Tail behavior is a separate empirical question

Many insurance loss distributions are skewed, but skewness and heavy-tail asymptotics are distinct features. A lognormal distribution has finite moments of every positive order and a tail that decays more slowly than any exponential. A regularly varying Pareto-type tail may have moments that fail to exist above a critical order. Catastrophe losses may combine rare event occurrence with a finite contractual maximum, making the observed insured payment bounded even if an underlying damage process is modeled as unbounded. The observed quantile in one range cannot reveal which asymptotic regime is true without information and assumptions beyond that range.

For a generalized Pareto exceedance model above threshold (u), let (H(z)=1-(1+\xi z/\sigma)^{-1/\xi}) for (z\geq0) on its valid support, with scale (\sigma>0), shape (\xi), and exponential limit when (\xi\to0). Let (p_u=P(Y>u)). For a continuous unconditional tail and (\tau>1-p_u), the threshold-exceedance approximation gives

[ Q_\tau(Y)\approx u+\frac{\sigma}{\xi}\left[\left(\frac{p_u}{1-\tau}\right)^\xi-1\right],\quad \xi\ne0, ]

with the logarithmic limit (u+\sigma\log{p_u/(1-\tau)}) when (\xi=0). This formula is a consequence of solving (1-\tau=p_u{1-H(Q_\tau-u)}). Its reliability depends critically on the adequacy of the approximation above (u) and the stability of estimated (p_u,\sigma,\xi). It does not become an exact nonparametric result by expressing the answer with a quantile symbol.

The insurance-finance treatment of extremes in Embrechts, Klüppelberg, and Mikosch (1997) provides an established mathematical basis for tail modeling. If (\xi>0), the generalized Pareto tail has finite positive mean excess only for (\xi<1) and finite variance only for (\xi<1/2). At or beyond these boundaries, sample averages or normal-theory intervals may be unstable or theoretically undefined for the unbounded model. Real contracts with deterministic limits can restore finite payment moments, but that is a property of the paid-loss variable after its cap, not of the postulated ground-up tail.

Connecting body and tail models

A distributional regression for the body and an exceedance model for the upper tail need a coherent joining rule. At each covariate profile (x), choose a threshold (u(x)), estimate (p_u(x)) consistently with the body CDF, and condition the tail on exceedance. The complete CDF must be nondecreasing, right-continuous, and normalized. Tail quantiles should not jump solely because two independently fitted components assign inconsistent exceedance probabilities at the threshold. One can also fit a full conditional heavy-tailed parametric family when evidence supports it; a separate tail model is not obligatory.

A changing threshold can be scientifically meaningful when the loss scale varies strongly across covariates, but the selected threshold is then itself part of the inferential design. Threshold diagnostics, sensitivity to covariate restrictions, enough independent exceedances, uncertainty in the frequency of exceedances, and cluster dependence must all be reported. An empirical return period of a catastrophe loss depends on event frequency and the exposure period; it cannot be obtained by interpreting a high conditional percentile of one policy as a geographic event return period.

Hou, Kang, Lo, and Peng's multi-step insurance tail-inference study explicitly combines an occurrence model, a dynamic quantile threshold, and semiparametric tail inference for semicontinuous claims data. The published abstract establishes this three-step research strategy; it is not independent evidence that every tail-threshold estimate remains stable across insurers. A related two-stage high-risk insurance model also addresses the semicontinuity of claim losses and theoretical inference. These contributions make clear that a high quantile is jointly determined by the zero mass, the positive body, and the extrapolated far tail.

Estimation uncertainty and event dependence

A nominal 99.5th percentile of a distribution can be computed from an estimated model even if the sample contains far fewer than 200 independent observations. Its precision, however, is then heavily model-dependent. For an iid sample of 10,000 observations, the expected number above the true 99.5th percentile is 50; at 100,000 it is 500. These counts follow (n(1-\tau)) and assume identical conditional exceedance probabilities and effective independence. Real catastrophe or liability datasets may exhibit strong clusters, and their effective tail sample can be substantially smaller than the row count.

Under threshold uncertainty, inflation-regime shifts, or parameter instability, predictive uncertainty should include both stochastic process variation and estimation error when relevant to the decision. A single point estimate of the far quantile obscures uncertainty associated with the choice of tail family. Profile likelihood, bootstrap under dependence, Bayesian posterior predictive distributions, model averaging, and explicit scenario stresses can address different parts of the problem. The appropriateness of each depends on data-generating assumptions; none substitutes for examining whether the historical observation window is representative of the targeted future exposure.

The division between aleatory and epistemic uncertainty also has an operational limit: if contract wording, legal outcomes, or claims inflation regime changes, the target distribution itself may have changed. A beautifully calibrated model over a stable preceding decade can fail because a future legal or physical environment moves beyond its experience. Quantile regression measures conditional patterns in specified information; prospective stress analysis addresses plausible departures from the learned regime.

Portfolio aggregation changes the question

A portfolio total (L=\sum_{i=1}^{m}Y_i) requires a joint distribution of the components, not merely (m) marginal conditional quantile curves. If losses share a catastrophe event, correlated court rulings, common repair inflation, or cyber infrastructure dependency, dependence can dominate aggregate tail risk. Two books can have identical marginal quantiles and different aggregate quantiles under different dependence structures. Under comonotonicity and suitable continuity, summing component quantiles can be correct; under general dependence, it is not.

An illustrative comparison involves two Bernoulli losses, each paying 100 with probability 0.05 and zero otherwise. Under independence, the sum is 0 with probability 0.9025, 100 with probability 0.095, and 200 with probability 0.0025; therefore its lower 95th percentile is 100. Under perfect positive dependence, the sum is 0 with probability 0.95 and 200 with probability 0.05; its lower 95th percentile is 0. Both portfolios have expected total loss 10, and each individual loss has a 95th percentile of 0. This counterexample demonstrates simultaneously that equal means and equal marginal quantiles leave aggregate quantiles underdetermined. The difference at the 95th percentile exploits the lower-quantile convention and the atom at the boundary; the full distributions convey more information than a single percentile.

The claim that conditional policy quantiles can directly determine regulatory capital is therefore too strong. A portfolio capital model must address dependence, claim development, contractual risk transfer, asset/liability interaction, and the precise prudential capital definition. The quantitative and institutional reasons are developed in the final chapter.

Distributional Diagnostics

Predictive validation precedes institutional inference

An estimated quantile is a probabilistic prediction about future observable outcomes under defined conditioning information. Its validation must therefore be out of sample in a manner consistent with the time and population of use. Random cross-validation of policy rows may leak the same insured, event, catastrophe, or policy renewal history into both training and test sets. Predictors based on ultimately settled losses, post-loss information, or data-cleaning decisions made using the full outcome sample can introduce further leakage. An evaluation protocol should use an underwriting or valuation cutoff, a fixed information set, development-consistent outcomes, and a test period that includes available relevant regimes.

At quantile level (\tau), the empirical pinball score is (n^{-1}\sum_i\rho_\tau(y_i-\widehat q_{\tau,i})). A smaller held-out score indicates better performance for that quantile objective across the chosen population and weighting scheme. The score has units of currency if (Y) is measured in currency, and its value depends on the scale and population of losses. Scores across different lines, currencies, deductibles, or horizons must not be pooled indiscriminately. For insurer comparisons, use consistent exposure normalization and clearly specify whether large contracts receive the same weight as small contracts.

A calibrated continuous conditional quantile satisfies (P{Y\leq\widehat Q_\tau(X)\mid X=x}\approx\tau) for relevant conditioning groups, recognizing that empirical tests are noisy. For atomic distributions, the appropriate inequalities at a lower quantile are (P(Y<q\mid x)\leq\tau\leq P(Y\leq q\mid x)). Thus an observed frequency of outcomes strictly exceeding a fitted quantile need not be exactly (1-\tau) when many observations equal the quantile. One must use randomized tie handling, interval calibration diagnostics, or conditional mass tests appropriate to the discrete component. Treating a zero median as a continuous percentile can generate a spurious calibration failure.

Proper scoring of full distributions

Proper scoring rules reward accurate probability forecasts for the entire selected distributional target. The scoring-rule theory of Gneiting and Raftery (2007) supplies the conceptual basis. A conventional negatively oriented logarithmic score is (-\log f(y)) for a continuous density forecast when the realized outcome has positive density under it. For mixed insurance distributions, the log score must use a dominating measure that accounts for the atom at zero as well as the continuous positive component. Ignoring the zero probability can assign artificial zero density to observed zero losses and make the score meaningless.

The continuous ranked probability score, when the relevant expectation is finite, can be written as

[ \operatorname{CRPS}(F,y)=\int_{-\infty}^{\infty}{F(z)-\mathbf{1}(y\leq z)}^{2},dz. ]

It integrates discrepancies between predictive and realized cumulative indicators and evaluates an entire CDF. There is also a quantile-score representation through integrated pinball losses under appropriate integrability conditions. CRPS and log score emphasize different aspects of predictive misspecification: a model can rank well under one and less well under the other. Choice of score should be made before inspecting the test results and aligned with the intended decision, with additional diagnostics of regions that a global average score may underweight.

Discrete-count evaluation has its own tools. Czado, Gneiting, and Held's probabilistic assessment of count data discusses calibration and scoring in discrete ordered outcomes. Their work supports a general methodological warning: continuous probability integral transform diagnostics are not automatically uniform for discrete observations. A randomized PIT can be defined by (U=F(Y^-)+V{F(Y)-F(Y^-)}), where (V\sim\operatorname{Uniform}(0,1)) is independent. If the forecast distribution is correct, (U) is uniform under the relevant conditioning assumptions. A PIT histogram can reveal aggregate errors yet conceal opposite errors in different risk segments; segmentwise reliability remains necessary.

Quantile coverage and tail sensitivity

A responsible validation report includes unconditional coverage, conditional coverage across risk bands, score by quantile level, calibration after changes in deductible and limit, and sensitivity to the largest independent losses. For high quantiles, report the number of effective exceedances and interval uncertainty. A 99.9th percentile checked on 2,000 independent observations would have only two expected exceedances under a correctly calibrated continuous forecast; its empirical coverage estimate is too noisy to justify tight declarations of accuracy without additional model assumptions. A lower pinball loss at the 95th percentile does not prove adequate 99.9th-percentile calibration.

Joint assessment of predicted value at risk and expected shortfall requires special scoring logic. Fissler and Ziegel's higher-order elicitability results establish an important framework for scoring multivariate functionals, including the VaR–expected-shortfall pair under stated conditions. Expected shortfall alone cannot be compared by casually applying ordinary quantile check loss; it is a different functional. In particular, (\operatorname{ES}\tau={1/(1-\tau)}\int\tau^1Q_u,du) for an integrable loss under a generalized-quantile definition. Near a modeled moment boundary, the very existence of this integral must be considered before selecting a scoring procedure.

Model-uncertainty and data-quality diagnostics

Model comparisons should combine distributional calibration with data-generation audits. Relevant records include exposure start and end dates, earned exposure, geographic and industry coding, currency and inflation adjustments, retention and limit terms, risk period, reporting and closure dates, settlement amounts, negative recoveries, reserve revisions, litigation status, and recoveries from reinsurers. Some paid or net loss variables can be negative after recoveries; the nonnegative derivations in earlier chapters apply only once the precise variable is confirmed nonnegative. Using a nonnegative family on signed net payments requires a new target definition or an appropriate signed distribution.

Conditional independence of claim observations is rarely a safe default. Diagnostic residuals should be examined by shared event, location, policyholder, line, year, and claims handler where relevant. Catastrophe years can dominate a random-split validation, while a single unexpectedly benign year can flatter a tail model. A useful research design includes blocked temporal validation and scenario analysis across plausible changes in inflation, frequency, limits, and reporting speed. Validation of a model's distributional skill is an ongoing property of a forecasting process, not an immutable certification of a fitted algorithm.

A reproducible evaluation protocol

The following sequence operationalizes the scientific requirements without prescribing an insurer-specific model. First, record the exact response variable and its economic meaning, including exposure basis and treatment of zero and negative amounts. Second, identify the observation and reporting process, including censoring, truncation, and missing values. Third, select baseline mean, two-part, direct quantile, and full distributional alternatives corresponding to the research question. Fourth, keep model selection and hyperparameter tuning inside training or validation partitions with time-consistent feature availability. Fifth, freeze a genuinely out-of-sample test and report multiple proper scores, quantile-specific check losses, zero-mass calibration, and tail exceedances. Sixth, quantify uncertainty using dependence-aware procedures and evaluate sensitivity to model family, tail threshold, and economic regime. Seventh, record which proposed actuarial decisions survive the observed forecast uncertainty. Every step produces interpretable evidence, not merely a pass/fail metric.

A model can be accurate for central quantiles and deficient in extreme-loss prediction; it can forecast positive claim severity while misestimating claim occurrence; it can fit the policy population while failing for the largest risks. These are distinct diagnostic findings and should be reported separately. More elaborate methodology is justified when it resolves an identified failure relevant to a decision and remains reproducible under the observed data constraints.

Insurance Financial Decisions

Expected-cost pricing and risk-sensitive tariffs

The pure expected-cost component of a conventional premium is related to (E[Y\mid x]) for the relevant contract and exposure, possibly with a frequency-severity decomposition. Expenses, commissions, capital costs, reinsurance costs, taxes, cross-subsidies, underwriting objectives, and legal pricing constraints require additional inputs. The upper conditional quantile can inform risk differentiation and safety loading but is not automatically an adequate price. Kudryavtsev's quantile-based ratemaking analysis and Heras et al.'s quantile premium study investigate quantile alternatives within specified actuarial settings.

The earlier hypothetical distributions demonstrate the substantive tension. Two classes have the same pure mean loss of 200. Charging each a premium based only on that expectation makes no distinction between their 95th percentiles of 200 and 1,386.29. Yet setting a premium equal to each 95th quantile is also a choice needing justification: the riskier class's quantile could be more than six times the common mean, while expected losses remain equal. A loading principle must specify how tail risk, portfolio diversification, capital scarcity, market competition, and applicable fairness rules enter the decision. Empirical quantile estimates offer evidence about risk heterogeneity; they do not dictate the social or commercial allocation of the resulting costs.

Baione and Biancalana's parametric quantile ratemaking extension addresses implementation complexity when many risk classes need estimated quantiles. An operational implication is that the relative simplicity of a parametric quantile curve must be judged against potential misspecification of upper-tail behavior. A highly flexible learner might differentiate risk classes better in one historical dataset, while a stable calibrated distributional model may provide more coherent tail probabilities in another. The evaluation criterion should follow the premium rule and the chosen observation horizon.

Retentions, limits, and reinsurance design

Distributional regression has immediate use in analyzing indemnity structures. The probability that a loss exceeds an attachment (a) is (1-F_x(a)) for a strictly greater-than event; expected excess payment is (E[(Y-a)_+\mid x]=\int_a^\infty{1-F_x(t)},dt) for a nonnegative integrable loss. For a layer of width (\ell), expected payment is the difference between expected excess above (a) and above (a+\ell). These formulas use the entire survival curve in the layer region. A single percentile can indicate a threshold at one probability level, but it does not identify the expected layer cost or the shape of loss experience inside the layer.

A parametric distributional model permits direct integration or simulation for a deductible, limit, quota share, or excess-of-loss layer. If positive losses follow a heavy-tailed family, uncertainty at the attachment and upper limit requires special scrutiny. The contract's actual reinstatement, aggregation, occurrence, and exhaustion provisions determine the payment functional. A different reinsurance purchase can alter the net-loss distribution without changing ground-up losses. Gross and net quantiles therefore require separate definitions and, where necessary, a stochastic representation of counterparty or reinstatement risk.

Reserving and insurance accounting

Claims reserving asks for amounts and timing of future cash outflows for past insured events, considering reported and unreported claims, claim development, inflation, case reserves, and claims handling. Conditional quantile regression can characterize uncertainty in ultimate losses, development factors, or reserve errors. Full distributional models can support stochastic cash-flow simulations. But a distribution over future paid payments for a valuation date differs from a distribution over new business policy losses. Calendar-year effects, reporting delays, and claims development must be included appropriately; otherwise, an apparently rigorous loss quantile may be applied to the wrong liability.

The IFRS 17 insurance-contract framework measures groups of insurance contracts with risk-adjusted present values of future cash flows and other prescribed components. Its risk adjustment for non-financial risk concerns compensation for bearing insurance risk. A predicted percentile of an isolated claim is not itself an IFRS 17 risk adjustment. A quantile or confidence-level technique may inform an implementation, but the accounting target includes grouping, timing, discounting, compensation for uncertainty, and the entity's measurement objectives. The applicable accounting standards and their local adoption control the final calculation.

Solvency and portfolio capital

Under the European Solvency II framework, Article 101 specifies that the Solvency Capital Requirement corresponds to 99.5% one-year Value at Risk of the basic own funds, taking the prescribed risks and future business treatment into account. This is an institution-level one-year change in basic own funds, not an unconditional 99.5th percentile of one motor claim or a single policy-year cost. A direct model for aggregate claims can be a component of the underwriting-risk assessment, but the regulatory target also depends on assets, liabilities, risk transfer, dependencies, and the specified capital model. Distinguishing the loss sign from an own-funds change is essential: a high loss quantile may correspond to a low own-funds quantile.

Another implication concerns diversification. Summing policy-level expected losses preserves linearity, but summing policy-level Value-at-Risk estimates generally does not recover portfolio VaR. Estimation of aggregate portfolio tails therefore needs a joint dependence structure, at least sufficiently detailed for the intended capital measure. A hierarchy from exposure to peril to event and then financial result is often analytically necessary. The extreme-loss methodology for insurance and finance provides the underlying mathematical warning about reliance on scarce tail information; it cannot replace a jurisdiction's formal regulatory instructions.

Catastrophe, climate, and systemic uncertainty

Catastrophe insurance is particularly sensitive to the distinction between conditional risk and distributional shift. Geographic attributes, building characteristics, mitigation, and event exposure may inform the conditional distribution of loss. Physical hazard changes, new construction, litigation standards, and insurer portfolio migration can change both the hazard and the portfolio to which the model is applied. A probabilistic regression estimated from claims alone may also reflect sparse insured participation and historical coverage patterns rather than the full physical loss distribution.

A catastrophe loss model that predicts the marginal 99th quantile of individual locations is not a complete simulation of region-wide correlated losses. An event-based representation must connect hazard footprint, vulnerability, insured exposure, contract transformations, and aggregation. Distributional regression can estimate conditional vulnerability or component loss distributions within such a system. Its empirical validity is constrained by event coverage, model error at unobserved intensities, and possible environmental nonstationarity. Heavy-tail sensitivity, dependence assumptions, and scenario analysis should be reported as separate uncertainty sources.

Cyber accumulation, mass tort, and systemic infrastructure failures create analogous problems. Losses can share causes while being recorded under distinct policies and years. Historical policy-level quantile calibration does not imply correct simultaneous event aggregation. Stress tests and structural dependence scenarios are therefore complements to distributional regression rather than optional embellishments when the decision concerns portfolio survival under concentrated events.

Governance and reproducible actuarial judgment

An operational model inventory should state the prediction target, historical training window, distribution family or quantile algorithm, contract transforms, exposure and development adjustments, treatment of zero and negative observations, tail threshold if any, and limitations of the available source data. It should also preserve the actual fitted version, validation split, software environment, training parameters, calibration tables, reproducible tests of material calculations, and the business decision rule consuming the forecast. The model owner's interpretation should identify assumptions whose failure would change pricing, reserving, reinsurance, or solvency decisions.

The distinction between statistical quality and institutional authorization remains important. A model may be well calibrated yet used for prohibited risk classification in a particular jurisdiction, or applied without adequate disclosure and governance. Conversely, a legally permitted tariff can remain statistically unstable in an unusual risk cell. These questions require separate empirical, legal, governance, and financial analyses. The dossier establishes the statistical research basis for them; it does not supply individualized insurance or regulatory advice.

Integrated analytical synthesis

The evidence supports a coherent sequence of inference. First specify the insurance random variable and conditioning information. Second represent occurrence, severity, dependence, and payment transforms with appropriate model structures. Third estimate targeted quantiles or the complete law, taking account of atoms, censoring, shape, and tail behavior. Fourth test held-out calibration and uncertainty at the loss levels relevant to the decision. Fifth apply the contractual, aggregation, and institutional mapping that translates predicted losses into premiums, reserves, layer costs, or capital. This sequence is an original synthesis of established statistical and actuarial material, not a newly claimed theory or universal mandatory protocol.

A useful substantive conclusion follows. A distributional predictor's most important output may be a diagnostic finding that the available evidence does not identify an economically meaningful far-tail quantile with sufficient precision. The appropriate next step can be to broaden exposure data, model an event mechanism, procure additional claims experience, simplify a distributional assumption, or report a range of outcomes instead of a single point. Scientific adequacy consists in matching the claimed precision and scope to the information actually present in the data.

Conclusion

Quantile and distributional regression expand insurance loss analysis from conditional expected cost to the structure of outcomes. Quantile regression estimates specified order statistics under an asymmetric-loss objective. Full distributional regression models a conditional probability law from which occurrence probabilities, central tendencies, high quantiles, expected layer losses, and other functionals can be computed if the required assumptions and moments hold. The choice between them depends on the estimand, data-generating process, and decision, with the accuracy of each established empirically rather than by method label.

Six substantive findings follow from the research. A loss model must define its unit, contractual basis, exposure horizon, and prediction information set before interpretation. Zero-heavy policy-loss distributions require explicit treatment of occurrence and positive severity; the unconditional quantile can change substantially while the positive-severity law stays fixed. Covariates can influence scale and shape independently of their mean effects, and multiple fitted quantiles must respect distributional monotonicity. Extreme quantiles require tail-specific assumptions and enough independent evidence, with moment and threshold limitations made explicit. Proper scoring and out-of-sample conditional calibration test different aspects of forecast quality. Financial application requires a further mapping from conditional losses to contracts, portfolios, liabilities, and supervisory or accounting objectives.

The original mathematical examples establish that identical conditional means do not imply equivalent upper-tail risk and that identical marginal quantiles do not determine portfolio VaR. They also demonstrate how the probability of no claim enters policy quantiles and why a layer cost requires more than one quantile. These conclusions follow directly from probability and expected-value identities. They make no claim to scientific priority or to a novel authored theoretical system.

The empirical research agenda is to compare a fixed, transparent set of mean, two-part, direct quantile, and full-distribution benchmarks across insurance products while preserving exposure, policy, time, and source-data definitions. Such comparisons should evaluate sparse-tail reliability, dependence, censoring, development, calibration under distribution shift, and the financial sensitivity of downstream decisions. The standard of evidence is the reproducibility of the complete route from recorded claims to the stated actuarial conclusion. The Research Dossier is an international research synthesis, with all numerical illustrations labeled as hypothetical and all regulatory statements attached to their specific institutional sources.

References