RESEARCH DOSSIER
Nonstationarity and Time-Varying Risk
Abstract. Nonstationarity changes the relationship between observed losses and the future distribution that an insurer, public authority, or risk-bearing enterprise must evaluate. Its consequences reach beyond changing average claim levels: event frequency, severity, temporal clustering, cross-risk dependence, distributional tails, exposure composition, observation systems, and the decision horizon can evolve at different rates. This monograph establishes a precise statistical treatment of nonstationarity in risk science and develops its implications for actuarial estimation, risk capital, pricing, risk transfer, and governance. It distinguishes strict and weak stationarity, ergodicity, conditional heteroskedasticity, deterministic trends, structural breaks, latent regimes, and uncertainty about future transitions. Drawing on primary statistical scholarship and international insurance supervisory documents, it derives the loss implications of time-varying point-process intensity, regime mixtures, covariance changes, and tail-probability thresholds. Three transparent synthetic numerical exercises show how pooled historical averages can misstate a current loss rate, how a shared latent state generates a material variance floor despite conditionally independent risks, and how a small increase in catastrophe probability can abruptly change a high-quantile capital estimate. It then evaluates diagnostics, chronological out-of-sample validation, predictive scoring, calibration, scenario analysis, and limitations of inference from short or censored loss histories. The central analytical conclusion is that risk estimates acquire decision meaning from an explicitly dated forecast origin, exposure definition, conditional information set, prediction horizon, and specified future loss law. Statistical stationarity can hold for a suitable transformed or regime-augmented process even when raw observations drift; conversely, a stable-looking historical sample can conceal changes of direct consequence for future losses. The contribution is a source-grounded analytical synthesis and reproducible exposition, without the institution of a new authorial theory, model, or doctrine.
Keywords: Nonstationarity; time-varying risk; strict stationarity; weak stationarity; ergodicity; structural breaks; change points; regime switching; latent states; nonhomogeneous Poisson process; claim frequency; claim severity; heteroskedasticity; changing correlation; tail risk; extreme value theory; exposure normalization; chronological backtesting; predictive calibration; insurance capital; own risk and solvency assessment; model uncertainty.
Introduction
A prediction about future insured loss is a statement about a probability law indexed to a particular decision date. A historical average is an estimator constructed from earlier observations. When the law governing those observations changes, the estimator's relationship to the next period becomes a question of statistical identification, forecast horizon, and causal structure. The monetary consequences can be direct: a premium derived from an obsolete frequency level may underfinance expected claims, a reserving development pattern may fail after claim-reporting practices change, and a diversified portfolio may accumulate joint losses if common conditions alter dependence.
The subject is broader than climatic trends. Changes in law, medical practice, insured demographics, cyber vulnerabilities, technology, macroeconomic conditions, inflation, portfolio composition, reporting procedures, and catastrophe intensity can each generate temporal changes in a risk measurement. Some are continuous, some abrupt, and others emerge through an unobserved state. Their empirical footprints differ. A mean shift may be visible in a relatively short claim-count series; an evolving 99.5th-percentile event probability may remain undetectable with decades of annual observations. A robust analysis therefore treats change detection, forward projection, and decision consequences as separate inquiries.
The classical statistical literature supplies essential distinctions. The unit-root framework developed by Dickey and Fuller (1979), and the complementary stationarity-null test of Kwiatkowski et al. (1992), interrogate different null hypotheses. Perron's structural-break analysis (1989) demonstrates why an unmodeled break can distort such classifications. Hamilton's regime-switching approach (1989) makes changing, unobserved economic conditions part of a probability model rather than an after-the-fact explanation. The current research develops these insights for losses and insurance portfolios while retaining their original assumptions and limits.
The scientific object of this dossier is the time variation of loss distributions, structural regimes, and dependence relationships. The separate dossier on Dynamic Risk and Time-Dependent Exposure addresses changing insured objects and population composition as its principal object; here exposure enters as a necessary control for identifying changes in the underlying loss law. The separate climate-specific dossiers investigate physical climate processes and particular protection systems. Here climate is one comparative source of time-varying risk among others. The approved parent Science, Risk Theory, supplies the foundational probabilistic setting.
The analysis proceeds through formal definitions, independently derived results, synthetic counterexamples, a comparison of scholarly methods, and a bounded examination of supervisory expectations. No insurer's proprietary claims data have been supplied; numerical illustrations are exactly specified hypothetical constructions rather than empirical estimates. Citations identify the relevant publication or original guidance in the sentence where the evidentiary proposition is discussed. Mathematical identities are derived under explicit assumptions and do not acquire empirical validity merely because the algebra is correct.
Stationarity and Ergodicity
Probability laws and information sets
Let L_t denote a monetary loss or loss rate associated with a specified portfolio and observation interval t, and let F_t be the information available at the forecast origin. The decision-relevant quantity is a conditional predictive distribution P(L_{t+h} ≤ x | F_t, E_{t+h}, C_{t+h}), where E describes the prospective exposure and C the relevant contract definition, coverage, and claims-recording conditions. This notation is a bookkeeping statement, not a proposed new stochastic model. It makes explicit three ordinary sources of apparent distribution change: the population at risk may change; the insured or reported definition of a loss may change; or the conditional distribution of the underlying events may change.
A sequence can be studied through its unconditional joint distributions or through predictive laws conditional on information. Those are different mathematical objects. For example, a volatility process may have a time-invariant unconditional distribution while its conditional variance changes following an unusually large observation. The original ARCH analysis by Engle (1982) establishes this conditional-variance distinction; the GARCH extension by Bollerslev (1986) describes autoregressive conditional-variance dynamics and their stationarity conditions. Neither should be treated automatically as evidence that the unconditional probability law permanently changes with calendar time.
A change in forecast P(L_{t+h} | F_t) can reflect Bayesian learning about a stable unknown mechanism, a stable mechanism with evolving state, or actual change in the underlying data-generating law. One forecast revision can therefore be entirely appropriate without implying structural instability. The empirical task is to distinguish sources of revision sufficiently for the specific risk decision, not to attach a single categorical label to every time series.
Strict and weak stationarity
A process {X_t} is strictly stationary when every finite-dimensional joint distribution is invariant to time translation: for any positive integer k, times t_1,...,t_k, and permitted shift s, (X_{t_1},...,X_{t_k}) has the same joint distribution as (X_{t_1+s},...,X_{t_k+s}). Strict stationarity is a property of the entire stochastic law. It is stronger than the constancy of a sample mean or a variance chart.
A second-order, or weakly, stationary process has a finite constant expectation μ, a finite variance, and autocovariance Cov(X_t,X_{t+k}) that depends on lag k rather than calendar date t. Strict stationarity with finite second moments implies weak stationarity, whereas weak stationarity does not establish invariance of every higher-order joint distribution. Heavy-tailed insurance losses may fail to possess a finite variance even when a strictly stationary process is mathematically well defined; a variance-based diagnostic is then inappropriate for that object.
Consider X_t = a + bt + ε_t with ε_t a zero-mean stationary error. For b ≠ 0 the raw process has a time-varying mean, but the detrended residual can be stationary if the assumed trend is correct and stable. A seasonal process may have a stable pattern within calendar position, which requires explicit seasonal adjustment or a periodically stationary description rather than pretending that all months are exchangeable. Conversely, if L_t appears level after deflation while the share of high-limit policies and the claim-reporting delay change, apparent stability of the monetary index provides little assurance of risk stability.
Unit-root hypotheses ask a more specialized question. In the simple autoregression X_t = ρX_{t-1} + ε_t, with suitable innovations, |ρ| < 1 permits a stationary solution, while ρ = 1 describes a stochastic trend under the model. The Dickey–Fuller estimator distribution explains why conventional regression critical values are unsuitable at a unit root. The KPSS stationarity-null framework reverses the hypothesis orientation and can support complementary diagnostics. Neither test supplies a universal certificate that a multivariate insurance tail distribution will remain fixed in future decades.
Ergodicity and historical estimation
Ergodicity addresses whether suitable time averages along a sufficiently long realization converge to corresponding ensemble expectations. Its meaning depends on the transformation and integrability conditions in question. Stationarity alone does not ensure the requisite ergodic property. A simple counterexample is X_t = Θ for every t, where Θ is sampled once from a nondegenerate distribution. The process is strictly stationary; the sample mean along a realized history equals Θ at every sample size, while the ensemble mean equals E[Θ]. One portfolio's indefinitely repeated observation cannot reveal the population expectation without additional information about the latent variable.
For independent identically distributed integrable observations, a law of large numbers supplies an ordinary consistency route. Insurance observations rarely have perfectly interchangeable units: policies enter and leave, legal coverage changes, claims develop after the occurrence date, and clustered catastrophes jointly affect numerous contracts. Increasing the number of administrative records does not automatically increase the number of independent realizations of the underlying common hazard. If the dependence is driven by a persistent latent regime, a long sequence confined to one regime estimates that regime's experience rather than an unconditional mixture.
The distinction has a practical implication for credibility: a large exposure denominator can yield a precise estimate of the historical period's frequency while leaving the forecast under a structural change highly uncertain. Sampling error within a stable segment and uncertainty about extrapolating to a future segment must be represented separately. A narrow confidence interval for the first task is not a narrow prediction interval for the second.
Conditional volatility, seasonality, and aggregation
Insurance volatility can vary within a stationary unconditional law. ARCH and GARCH processes give a particularly clear example: recent squared innovations affect conditional variance; under their appropriate restrictions a time-invariant marginal distribution may nevertheless exist. A volatility spike then increases the next-period conditional tail forecast even though the invariant joint model has not changed. In contrast, a deterministic secular trend in frequency or a permanent parameter break creates calendar-dependent distributions unless an augmented description explicitly absorbs that change.
Aggregation also changes apparent stationarity. Suppose individual policies maintain a fixed claim probability conditioned on age, while the portfolio ages. The unadjusted population-average claim rate rises even without a shift in the age-specific claims mechanism. Likewise, a rise in insured replacement values can increase nominal severity without a rise in the probability of physical damage. A rigorous diagnostic reports claim counts per earned exposure unit, a stable claim definition, real and nominal loss indices, contractual limits, and changes in the composition of risks. Standardization can reveal a stable within-group mechanism, although it may also obscure the economically relevant actual total-loss trajectory if the changed composition is omitted from the final forecast.
Identification and diagnostics
A first diagnostic package checks data provenance, exposure denominators, descriptive moments across appropriate windows, claim-count dispersion, autocorrelation, seasonal effects, and the stability of the definition of a loss. Cumulative-sum plots or control charts may flag an abrupt change, while formal tests must use an error structure consistent with the data and account for repeated surveillance. The continuous inspection schemes of Page (1954) are an early foundation for cumulative monitoring. Modern change-point inference supplies methods for retrospective segmentation, but the stopping rules and false-alarm behavior of real-time monitoring are different from the evidence required to date a break in a completed sample.
The correct diagnostic conclusion has a bounded form: an identified feature is inconsistent with a specified stationary null under a specified sampling and observation model. Failure to reject such a null may reflect low power, short histories, aggregate heterogeneity, or censoring. A fitted time trend can be an informative summary, but its apparent significance in an autocorrelated series may be inflated by naive independence assumptions. Diagnostic uncertainty should follow the data into prospective forecasts rather than disappearing at the model-selection step.
Time-Varying Loss Frequency
Intensity, exposure, and counting processes
Let N(t) count occurrences of events meeting a fixed insurance definition up to calendar time t. The intensity of events may depend on current conditions, past events, seasonal position, or the number and type of insured exposures. For an ordinary nonhomogeneous Poisson process with deterministic, nonnegative, locally integrable intensity λ(u), increments over disjoint intervals are independent and
E[N(b) − N(a)] = ∫_a^b λ(u)du,
with Var[N(b) − N(a)] equal to the same integral. The probability of observing no event in [a,b] is exp(−∫_a^b λ(u)du). These identities follow directly from the nonhomogeneous Poisson assumptions. The statistical literature on point processes, including Cox's analysis of event series (1955), provides a broader framework for examining event counts, randomness, and dependence. The present equations describe a narrower specified process, not all claim streams.
For an insurance portfolio, an operational representation is λ(u) = E(u)r(u), where E(u) is earned exposure in consistent risk units and r(u) is the event intensity per risk unit at time u. A change in λ may arise wholly from growth in E, wholly from a change in r, or from both. Estimating a rate without the proper denominator confuses portfolio growth with per-unit risk change. Changes in policy limits may affect the threshold at which a reported claim enters the data, thereby changing observed counts even when underlying incident rates remain stable.
Calendar time and development time require separate indexing. The date an incident occurs can differ from the date it is reported, settled, or statistically recognized. A dataset defined by report date is not necessarily a sample from the same occurrence-year process used in pricing a prospective policy period. Triangular claims data and delayed reporting must therefore be reconciled with the loss occurrence process; truncation near the observation cutoff can otherwise imitate a favorable recent frequency trend.
Nonhomogeneous Poisson derivation
Suppose an insurer observes a portfolio for one unit of future time, with intensity λ(u) that increases from its present value. For an expected event count over horizon h, the correct expectation is Λ(t,t+h) = ∫_t^{t+h} λ(u)du. Substituting λ(t)h is a local approximation that requires λ to change slowly over the horizon. If λ(u) = λ(t)exp(g(u−t)) with constant growth rate g, then for g ≠ 0,
Λ(t,t+h) = λ(t)[exp(gh)−1]/g.
For g = 0 the continuous limit equals λ(t)h. The relative difference grows with gh. This demonstrates a horizon effect: even an accurately measured current annualized intensity does not determine expected claims over an extended contract or bond term unless the trajectory is also specified.
For a time-varying stochastic intensity, one must distinguish event randomness conditional on a realized intensity path from the randomness of the path itself. If N conditional on integrated intensity Λ is Poisson, then E[N] = E[Λ] and Var(N) = E[Λ] + Var(Λ) by the laws of total expectation and variance. The second term is absent in a deterministic-intensity Poisson calculation. It can be economically material where annual hazard intensity is subject to shared climatic, technological, epidemic, or macroeconomic states. An unconditional claims process that is Poisson conditional on a random state is generally overdispersed relative to a Poisson law with the same marginal mean.
The label “annual event probability” needs care. For a Poisson arrival process, the probability of at least one event in a period is 1−exp(−Λ), while Λ is the expected number of events. They coincide only approximately for small Λ. A hazard-based statement about annual exceedance probability cannot be substituted without transformation for an expected annual event count. More fundamentally, when intensity changes with calendar time, 1/p_t is at most an instantaneous shorthand sometimes called a return period; it is not a generally valid waiting-time expectation. The first-event survival function depends on the integral of the entire future intensity trajectory.
Frequency–severity separation
Write S_t = Σ_{j=1}^{N_t}Y_{tj}, where Y_{tj} is the covered loss amount associated with event j during period t. Conditional on a specified state, suppose N_t is independent of independent identically distributed severities with finite mean m_t and variance v_t. Then
E[S_t | state] = E[N_t | state]m_t,
Var(S_t | state) = E[N_t | state]v_t + Var(N_t | state)m_t².
Under a conditional Poisson frequency model with mean Λ_t, this simplifies to Var(S_t | state) = Λ_t E[Y_t² | state]. If frequency and severity share a state, the unconditional expectation remains E[Λ_t m_t] rather than E[Λ_t]E[m_t] unless their relevant dependence is absent. The total variance additionally contains variation of conditional aggregate loss means. An apparent stable frequency series is thus compatible with worsening aggregate-loss risk when inflation, deductibles, legal awards, or heavy-tail severity changes.
For claim severity, a stationary normalized or inflation-adjusted distribution need not imply an unchanged insured loss distribution. Repair costs can grow faster than a general price index, coverage can extend to larger limits, and deductibles can alter the share of claims recorded. The tail of insured loss also depends on an insurer's policy attachment and exhaustion points. A time-varying coefficient in a severity regression is a statistical claim about that particular data-generating and measurement specification; broader interpretations must be justified separately.
Threshold exceedances warrant special attention. The high-threshold models of Davison and Smith (1990) examine exceedance sizes, occurrence, dependence, seasonality, and model diagnostics. Their extreme-value framework is compatible with time-dependent covariates, but asymptotic tail forms do not independently establish a stable threshold or constant tail shape across future regimes. A claim that a former 100-year loss has become a 30-year loss requires explicit, comparable definitions of the event threshold, exposure, and changing annual exceedance probability, with the horizon interpretation stated.
Worked illustration: a changing rate
Consider a synthetic portfolio of 10,000 homogeneous earned exposure-years per annual observation period. For an earlier period, suppose the insured event intensity equals 0.08 events per exposure-year. For the current future period, the assumed rate is 0.14. Under a Poisson frequency model the earlier expected count is 800 and the current expected count is 1,400. A historical model averaging equal exposure volumes from those two periods returns the mean count 1,100 and rate 0.11. Applied to the current portfolio, it understates the current expected count by 300, or 300/1,400 = 21.43% of the current expected count.
Now suppose the conditional mean covered severity is fixed at 2.0 million under the former rate, 2.75 million under the two-period pooled estimate. The resulting expected-loss gap is $750,000 per current year. These values are exact consequences of the stated hypothetical assumptions; they are not a statement about any country's actual claim inflation or insurance market. With deterministic intensity, Poisson count standard deviations are √800 ≈ 28.28, √1,400 ≈ 37.42, and √1,100 ≈ 33.17. Consequently, increasing the number of exposures can make the historical rate estimate sharply precise while failing to remedy the structural bias of using 0.11 to price an assumed current rate of 0.14.
Three forecasting approaches can be distinguished. An equal-weight full-history estimator has lower sampling volatility if the entire history truly shares a common rate, but can lag a persistent change. A recent-only estimator reacts more quickly, but is noisier and may mistake a transient anomaly for a new level. A structured time-varying estimator uses evidence from time, exposure, covariates, and possible change points, requiring stronger assumptions that themselves must be validated. The correct selection depends on actual predictive performance at the desired horizon, the costs of under- and overestimation, and the possibility that an observed change reverses.
Development delay and observation bias
Insurance records are a measurement system as well as a record of events. Suppose a claim occurring near the end of period t is filed in period t+1 with some probability, and the filing delay distribution itself becomes shorter because of digital reporting. A report-year frequency increase may then occur with no increase in occurrence-year incidents. Similarly, an expanded automated detection system can expose losses previously unrecorded; a new fraud-screening rule can alter settlement and denial patterns. Interpreting these shifts as changes in risk without examining the administrative mechanism biases the scientific object.
A defensible longitudinal dataset therefore records occurrence, report, payment, and closure dates as distinct time variables; earned units, coverage limits, deductibles, cancellations, and geographic mix; evolving definitions of claim; and records of system migration or incomplete years. Modeling can then separate the observation mechanism from the underlying insured events to the extent data permit. When historical definitions cannot be reconciled, the period should be flagged as noncomparable and modeled with an explicit measurement break or excluded from a particular estimator with an audit trail. The cost of data exclusion is reduced information; the cost of unqualified pooling is possible systematic bias.
Structural Breaks
Breaks, trends, latent regimes, and parameter drift
A structural break is a change in a parameter, relationship, or stochastic mechanism at a particular point or within an identifiable transition interval. A piecewise constant mean μ_t = μ_1 for t < τ and μ_t = μ_2 for t ≥ τ represents a one-break model. A gradual trend μ_t = a+bt, a random walk μ_t = μ_{t-1}+η_t, a mean-reverting latent state, and a transition among discrete regimes give other mechanisms. They can fit similar short sequences while implying sharply different long-horizon predictive distributions.
This nonuniqueness is a central identification problem. A high recent claims period may be generated by a permanent structural change, an unusually adverse random realization under a stable distribution, a temporary regime, a reporting change, or a shift in portfolio composition. Fitting a complicated dynamic model does not resolve these alternatives without comparative predictive evidence or externally credible covariates. Conversely, discarding recent high observations merely because they are exceptional presupposes that the old risk law remains valid.
The mathematical consequences of a break were established in Perron's 1989 unit-root investigation, which shows how a single break in a deterministic trend can compromise conventional testing. Bai and Perron (1998) develop inference for multiple changes in linear regression models, including estimation of break dates and tests on the number of changes under stated assumptions. Their contribution concerns a defined model class; it does not make every sudden visual change statistically significant, nor does it establish the causal mechanism behind a change date.
A latent Markov regime is another model class. Let R_t take values in a finite set, with transition probabilities P(R_{t+1}=j | R_t=i) = p_ij, and suppose loss distributions vary by R_t. Inference then averages or filters over uncertain states rather than assigning an observed period to a known state with certainty. Hamilton's 1989 switching-regime analysis is a seminal development of probabilistic state inference and forecasting. A homogeneous Markov chain started in its invariant distribution may yield a stationary unconditional observed process, even though conditional forecasts evolve with filtered state probabilities. If transition probabilities themselves vary over calendar time, stationarity of the same specification requires further assumptions and is not inherited from the constant-transition formulation.
Change-point inference and testing
Retrospective segmentation identifies candidate changes after a set of observations is available. In a regression or generalized linear model, break dates may be chosen to minimize a penalized measure of within-segment lack of fit. Increasing the number of segments improves in-sample fit, so a penalty or explicit statistical test is necessary to avoid explaining random noise as changing parameters. The multiple-break procedures of Bai and Perron offer a classical route; the PELT change-point algorithm of Killick, Fearnhead, and Eckley (2012) addresses efficient exact penalized segmentation under its cost and pruning conditions. Computational efficiency does not prove that the selected breaks describe a stable causal process.
A prospectively monitored insurer faces an additional problem: there may be daily, weekly, or monthly opportunities to announce a break, and checking continuously raises the probability of a false alarm relative to a one-time test at a fixed significance level. The principles behind Page's continuous inspection method show why a detection rule needs explicit monitoring thresholds and an understood false-alarm or average-run-length interpretation. Retrospective break confidence intervals and prospective alarm thresholds should be reported separately; their nominal coverage or error rates are not interchangeable.
Claim series can also be sparse and overdispersed. A large retrospective change in the estimated rate may not be significant if only a few events were observed. Break tests derived for Gaussian linear regression errors should not be imposed mechanically on rare-event counts or serially clustered severities. Suitable likelihoods, robust resampling under the proposed null, or simulated finite-sample critical values may be more defensible. The existence of a statistically detected break may remain uncertain even when management has rational reasons to stress a materially plausible alternative.
Structural identification under limited data
A nonstationary account is vulnerable to overinterpretation when the number of potential parameters increases faster than the number of informative time periods. Ten thousand policies exposed to one flood are not ten thousand independent realizations of the annual regional flood regime. Likewise, a 30-year catastrophe series cannot credibly estimate unrestricted calendar-year-specific 99.5th-percentile losses. Pooling related regions or peril types may improve estimation only if the relationship can be justified and sensitivity to differences is examined.
Three layers of uncertainty should be distinguished in source and decision records. Process variability persists even if every parameter is known. Estimation uncertainty arises because a finite, incomplete sample is used to infer a specified law. Structural uncertainty arises because several different mechanisms can explain the observations and yield different future trajectories. Increasing sample size under one unchanged regime chiefly addresses estimation uncertainty for that regime. It may not resolve what happens following a legal, technological, or climatic transition.
Explanatory covariates can make certain changes more intelligible. A claim frequency trend that follows an increasing share of elderly insureds has a different policy implication from an age-adjusted deterioration in underlying health risk. A severity increase after a coverage expansion should be mapped to the contract change. An apparent catastrophe break that coincides with a switch in event definitions requires data reconciliation before a hazard conclusion. Such analyses can reduce structural ambiguity, but a plausible mechanism does not by itself establish causal identification from observational data.
Forecast design across breaks
A break-aware forecast must define whether it treats a change as permanent, transient, or subject to reversal. For a recent break judged permanent, older data may inform dispersion and severity shape while the post-break observations receive more weight in estimating the new mean. For a temporary regime, the forecast should mix plausible future states with transition probabilities. For smooth drift, a calendar-time or covariate model may be more parsimonious than repeated breaks. When no model dominates, a finite set of auditable scenarios can expose decision sensitivity without pretending to know a unique future distribution.
There is a fundamental asymmetry between explaining history and forecasting a new transition. A segment fitted after seeing the complete historical record benefits from information unavailable when an earlier forecast would have been issued. It cannot be scored as if its historical fitted errors were genuine real-time predictions. Every out-of-sample experiment must reconstruct the information set at each forecast origin, including the release times of claims data and any economic or environmental covariates. A method that reliably detects a break several periods after it occurs may still offer more useful future performance than a highly sensitive detector with frequent false alarms; the tradeoff is decision-specific.
Worked distributional-tail counterexample
Consider a synthetic annual loss L taking only two values: 100 million with probability p. Define Value at Risk at level α by VaR_α(L) = inf{x: P(L≤x)≥α}. At α=0.995, if p=0.004, P(L=0)=0.996; hence VaR_0.995(L)=100 million. A difference of two-tenths of a percentage point in annual loss-event probability moves the specified quantile from zero to the full $100 million severity.
The expected loss changes more smoothly, from 600,000. Expected shortfall defined as the average loss in the worst upper 0.5% of outcomes, with the usual fractional treatment of probability mass at the threshold, equals 100 million = 100 million in the second. These quantities are model implications, not estimates of a real insurer's capital need. They demonstrate that a minor change in exceedance probability relative to a regulatory quantile cutoff can create a very large jump in a discrete-loss VaR measure. Discontinuities at thresholds make both tail modeling and model uncertainty important, and an exact quantile computed from an incorrectly extrapolated distribution remains an incorrect forecast.
This example also illustrates why a model’s performance should be assessed with its economic loss function rather than with a single mean absolute prediction error. The relative seriousness of a 0.004-to-0.006 probability shift depends on contract size, catastrophe accumulation, reinsurance structure, liquidity, capital, and the decision horizon. It cannot be reduced to the statistical detectability of a two-period frequency difference.
Regime-Dependent Dependence
Dependence conditional on a state
Risk accumulation depends on the joint law of individual losses, not solely on their individual means and variances. Let X_1,...,X_n be insured losses and let R be an unobserved environmental, market, or operating regime. Conditional on R the losses may be independent. Unconditionally they can be strongly dependent because they share the same random regime. The law of total covariance gives
Cov(X_i,X_j) = E[Cov(X_i,X_j | R)] + Cov(E[X_i | R],E[X_j | R]).
If distinct risks are independent conditional on R, the first term is zero, but the second need not be. A portfolio that appears diversified within each conditional environment can therefore have material common-state loss variability across environments. This is a standard probability identity applied here to insurance loss, rather than a new model or dependence doctrine.
Time-varying dependence may take several forms. Marginal frequency or severity may change while the dependence structure remains stable; pairwise correlation may change even with stable individual marginal distributions; extreme co-occurrence may change while linear correlation is nearly unchanged; and the mix of risks within a portfolio may change while every within-class law stays fixed. Identifying the mechanism matters because a portfolio's diversification benefit, capital need, and reinsurance demand can respond differently under each possibility.
A model with regime-conditioned claims can be consistent with conditional stationarity within a state and with unconditional stationarity when the latent state process is itself stationary. The operational concern is that at a particular forecast origin the filtered state probability and expected persistence may differ from their long-run mixture values. If the state-transition mechanism itself changes, using a long-run mixture derived from old transitions compounds the error. Statistical labels must refer to precisely which process and conditioning convention are being discussed.
Common shocks and aggregation variance
For S_n = Σ_i X_i with finite second moments, Var(S_n) = Σ_i Var(X_i) + 2Σ_{i<j}Cov(X_i,X_j). Averaging losses per exposure divides this expression by n². Under independent homogeneous risks, the variance of the average shrinks at rate 1/n. Under a persistent common factor contributing a positive average covariance across pairs, a nonzero component may remain as n grows. A large policy count therefore provides substantial protection against independent idiosyncratic loss variation but not necessarily against systemic or catastrophe accumulation.
A general decomposition is Var(S_n) = E[Var(S_n | R)] + Var(E[S_n | R]). The first term concerns random losses conditional on the known state; the second concerns variation in the aggregate expected loss across states. A portfolio-level capital model that includes only the first term, or that replaces the random state by its average risk probability, omits the second. A common-factor model calibrated from only calm years may additionally underestimate how the state distribution and conditional dependence behave during stress.
Two risks can display different joint tail probabilities despite the same unconditional correlation. For a severe catastrophe portfolio, the practical risk object may be P(S_n > K), P(X_i > a, X_j > b), or the distribution of reinsurance layer recoveries. Correlation is a summary of second moments, not a full description of extreme co-occurrence. The high-threshold analysis of Davison and Smith provides a setting for examining exceedances and dependence, while the choice of joint-tail model requires separate empirical justification.
Worked illustration: latent-state accumulation
Consider 1,000 hypothetical insured units. At the start of a year one shared state R is drawn: calm with probability 0.9, stressed with probability 0.1. Conditional on a calm year, every unit independently produces a specified loss with probability 0.01; conditional on stress, the loss probability for each unit is 0.12. Independence is conditional on the one common state and does not imply unconditional independence. Let N be the total number of units producing a loss.
Conditional expected counts are E[N | calm]=10 and E[N | stress]=120. The unconditional mean is 0.9×10+0.1×120=21. The marginal loss probability of each unit is q=0.021. An incorrect unconditional independence model would assign a binomial count variance 1,000×0.021×0.979=20.559, with standard deviation approximately 4.534.
Under the stated shared-state construction, the expectation of conditional variance is 1,000[0.9×0.01×0.99+0.1×0.12×0.88]=19.47. The variance of the conditional expected count is 0.9×0.1×(120−10)²=1,089. Therefore Var(N)=1,108.47 and SD(N)≈33.294. Unconditional covariance between two distinct unit-loss indicators equals Var[q_R]=0.9×0.1×(0.12−0.01)²=0.001089; the implied pairwise indicator correlation is approximately 0.05297. Although this correlation appears modest, the aggregate count standard deviation is more than seven times the independent benchmark.
This calculation is fully reproducible from elementary probability. It demonstrates that a comparatively small average pairwise correlation can materially change portfolio dispersion because the number of cross-pairs increases quadratically. The illustration assumes fixed membership, homogeneous conditional probabilities, one regime draw per year, and no within-state event contagion. If regimes last multiple years, serial dependence must additionally be modeled. If the marginal loss amounts also rise in the stress state, joint frequency–severity effects can amplify the aggregate difference. No real-world correlation or claim rate is inferred from these assumed numbers.
Changes in correlation and portfolio diversification
For two portfolio components A and B with fixed unit variances and correlation ρ_t, the variance of their sum is 2+2ρ_t. If ρ_t changes from 0.2 to 0.8 while the unit marginal variances remain one, aggregate variance increases from 2.4 to 3.6 and aggregate standard deviation rises from approximately 1.549 to 1.897, a 22.47% increase. This simple identity isolates changing dependence without requiring that either component's marginal variance increase. In monetary applications the components must have consistent scales, horizons, and exposure units; correlation alone does not specify an actual dollar capital charge.
Correlation can change with the selected time window because of shifting risk mechanisms or because the sampling environment changes. Crisis observations frequently dominate estimated covariance matrices, but the mere existence of a high correlation during an episode does not establish that it will recur with the same probability and duration. Regime-conditioned covariance estimates, stable factor definitions, and sensitivity to stress duration can expose the range of plausible outcomes. They must be evaluated against enough independent stress observations to support their level of specificity.
The relevant connection to financial resilience is conditional: diversification credits based on historical low correlations may be unreliable when a common stress factor affects both sides of a portfolio. Reinsurance can transfer a defined loss layer but does not remove basis risk, counterparty exposures, or the underlying common peril. Geographic diversification can reduce sensitivity to geographically local events while leaving exposure to global inflation, epidemics, or correlated macroeconomic shocks. The appropriate unit of diversification is the joint distribution of the covered loss under the decision-relevant state, horizon, and contract conditions.
Extreme dependence and multi-year persistence
Annual risk models sometimes replace a multi-year sequence by independent draws from a single annual distribution. If the stress regime has persistence, the probability of several consecutive adverse years can be much larger than the product of independent unconditional annual probabilities. For a two-state Markov chain, P(R_{t+1}=stress | R_t=stress)=p_SS governs continuation, while the initial filtered probability determines the chance of being stressed at the decision date. The number of adverse years over a planning horizon depends jointly on these probabilities and on the loss distribution within each state.
A multi-year insurance contract, capital recovery plan, or reinsurance renewal strategy must therefore ask whether losses cluster temporally as well as across insureds. When several bad years occur consecutively, capital and liquidity can be depleted before replenishment, and reinsurance pricing or availability can also change. Annual marginal tail probabilities alone do not describe this path dependency. Calculating multi-year exhaustion probabilities requires assumptions about premium revisions, investment returns, claim settlement, risk transfer and state transition; none follows solely from an annual VaR statistic.
A more realistic dependence model is not automatically preferable to a parsimonious model. Its additional parameters need credible evidence and stable prediction performance. Where extreme events are rare, scenario boundaries and conservative stress testing may be more transparent than unstable estimates of high-dimensional tail parameters. The scientific responsibility is to show how loss consequences vary under plausible dependence structures and to identify what observations would distinguish them.
Out-of-Sample Risk Validation
Chronological validation and leakage prevention
The decisive empirical test of a time-varying risk model is whether it improves predictions made with information actually available at each forecast origin. Suppose an insurer predicts losses for year t using data through year t−1. To evaluate models fairly, an experiment should imitate that decision: fit the model on information available at the end of t−1, issue the forecast for t, record the realized result, and repeat for subsequent origins. Expanding or rolling windows are both legitimate design choices if their horizon, selection rule, and revision policy are fixed and disclosed.
Randomly dividing adjacent claims across training and test partitions can create leakage when multiple observations share a catastrophe, claim-development event, risk regime, or exposure history. Data leakage also occurs if a covariate is included at its revised historical value even though it was published only after the forecast origin. An apparent accuracy improvement from such procedures is an artifact of access to future information. Event grouping, calendar-time partitioning, publication-lag adjustment, and locked forecast timestamps preserve the integrity of the comparison.
The benchmark should reflect the actual decision. Useful alternatives may include a fixed historical-rate model, an exposure-standardized seasonal model, a rolling recent-window estimator, a prespecified break model, a switching-regime model, and a structured scenario mixture. Each has different sensitivity to old information and noise. Including a simple stable benchmark prevents the unexamined assumption that additional model complexity improves prediction. A model may rank best for one-year mean claims but fail for five-year aggregate tail loss; neither result substitutes for the other.
Proper scores and tail diagnostics
A complete predictive distribution deserves a score that evaluates the distribution itself. The foundational review by Gneiting and Raftery (2007) explains proper scoring rules, including the logarithmic score and the continuous ranked probability score, and how they encourage probabilistically faithful reporting. For a loss distribution forecast F and realization y, the continuous ranked probability score can be written as CRPS(F,y)=∫_{−∞}^{∞}[F(x)−1{y≤x}]²dx when the integral is finite. Lower realized CRPS is better under the loss-score convention. It combines predictive location and distributional spread without reducing a density forecast to its expected value.
For event probabilities, the Brier score (p−o)² evaluates a specified binary event, where o∈{0,1}. A single observed catastrophe does not permit strong inference about a one-in-200-year risk probability; the number of effectively independent trials and the regime context matter. Tail-specific tests can inspect exceedance frequencies, exceedance severity, conditional coverage following consecutive exceedances, and sensitivity to chosen thresholds. Exceedance indicators should be examined for clustering; a correct unconditional exceedance proportion can coexist with systematically mistimed risk forecasts.
The expected tail loss and the frequency of ruin are different targets from point forecasting. An insurer can have an excellent average annual-loss score yet miss the extreme events that dominate solvency planning. Tail-focused scores must themselves be selected carefully to preserve meaningful incentives. Evaluation should identify whether the model forecasts the full predictive distribution, a quantile, an exceedance probability, or a conditional mean beyond a threshold. A test of one target does not validate all others.
Model comparison and calibration
The Diebold–Mariano forecast comparison of 1995 develops methods for evaluating differences in forecast accuracy across competing predictions under stated properties of loss differentials. Giacomini and White (2006) examine conditional predictive ability, providing a complementary perspective on whether one forecast procedure performs better given observed information. These procedures require attention to dependence between forecast errors, overlapping horizons, model re-estimation, and the number of usable forecast origins. A nominally small statistical p-value computed using an inappropriate independence assumption may be misleading.
Calibration assesses whether events occurring under forecasts assigned a particular probability happen at compatible rates. Under nonstationarity, calibration must be examined by forecast vintage, regime, exposure group, and relevant time horizon. A forecast can be calibrated on average across a long historical mixture yet systematically understate risk in the recent high-risk regime. Equally, a stable model may appear inferior during a brief anomalous period despite better long-run performance. Rolling diagnostics are informative when their window size and associated sampling noise are shown.
Sharpness describes the concentration of forecasts; it is useful only alongside calibration. A very wide forecast interval may attain high coverage without offering a discriminating risk estimate. A very narrow interval may appear efficient until an adverse state arrives. The proper-scoring-rule framework supports comparing both aspects through a defensible predictive objective rather than rewarding width alone or interpreting every missed rare event as model failure.
Comparisons must also be decision-weighted. The direct economic costs of underpricing expected claims, overcapitalizing a low-risk book, purchasing unnecessary reinsurance, and experiencing unpaid claims are different. These costs depend on capital constraints, policyholder protection, regulation and business strategy. A formally proper score supplies statistical discipline, while a decision analysis explains why a particular forecast feature matters. Scientific assessment should report both where feasible and avoid presenting a business loss function as a universal statistical truth.
Structural-break monitoring and intervention
An operating model needs a monitoring process that specifies the statistic tracked, the observation lag, the alarm threshold, the expected false-alarm burden, the responsible decision maker, and the action following an alert. A cumulative-sum alarm is one possibility; segmented parameter stability tests and predictive-score deterioration are others. The choice should match event frequency and severity, data latency and how quickly an adverse structural shift can cause material financial harm.
Intervention deserves separate validation. Updating rates too often may cause premium volatility, adverse selection, customer confusion, or operational cost; updating too slowly allows material drift. A statistically credible break can require review rather than an immediate tariff change. Where regulation restricts rate revisions or contracts are multi-year, the institution may instead adjust underwriting limits, reinsurance, capital or risk monitoring. The mathematical model informs these actions; it does not create an automatic management command.
The monitoring record should preserve pre-alert forecasts, post-alert revisions, parameter snapshots and contemporaneous assumptions. Otherwise the organization may inadvertently judge earlier decisions with hindsight. After an actual structural event, the revised model should be tested with subsequent observations rather than awarded approval solely because it explains the break retrospectively. Such records also allow a later reviewer to distinguish a justified discretionary override from ex post storytelling.
Reproducibility, falsification, and evidentiary boundaries
A reproducible nonstationarity study discloses time origin, observation date, legal coverage version, exposure definition, included and excluded records, inflation adjustment, frequency denominator, severity censoring, development treatment, known covariate-release lag, candidate models, fitted parameter ranges, diagnostics, evaluation horizon, and decision threshold. Inference should separate observations used to choose a model from observations used to validate it. If selection is repeated after viewing the test set, the test set ceases to be a genuinely independent check and its evidence should be labeled accordingly.
The mathematical exercises in this monograph can be recalculated independently from their full assumptions. Their role is to falsify overly broad propositions, such as “more insured units always guarantee a small average-loss variance” or “a small difference in tail-event probability cannot materially affect a quantile measure.” They do not identify how often these situations occur in actual markets. The cited journal pages, abstracts, and official institutional records provide documentary support for their respective methods or policy statements. Full article methods were not universally independently re-estimated; no actuarial administrative dataset, field causal design, or new primary catastrophe catalog was analyzed.
For a genuine insurer deployment, scientific validation additionally requires the institution's actual observation system and portfolio data, restrictions on permissible pricing or capital assumptions in the relevant jurisdiction, governance approval, and a monitoring audit. A publication about theoretical models cannot stand in for those institution-specific controls. The boundary is especially important for a model trained on third-party event data, where coverage of nonreported events and historical changes in data-provider collection practices may be hard to verify.
Adaptation and Uncertainty
Forward-looking risk decisions
Nonstationarity changes the information needed for a decision. The relevant probability is prospective and conditional on the current and anticipated risk environment. A risk committee that asks whether a policy can be written at a particular premium, whether capital will meet obligations, or whether a reinsurance layer will respond adequately needs a forecast about the covered future interval. The forecast must incorporate a defensible account of current insured assets, legal definitions, event intensity, severity, dependence, and the path of future exposure. Historical frequencies retain information about mechanisms, variation, and conditional responses, but the decision problem determines how that information is used.
A practical risk assessment can distinguish three forecast objects. The first is the present-state forecast, which assumes the currently assessed regime persists over a near horizon. The second is the transition forecast, which assigns probabilities to plausible future changes and their durations. The third is a conditional stress forecast, which specifies consequences under deliberately adverse states without assigning them unsupported occurrence probabilities. Each has a distinct function. A present-state forecast addresses ordinary near-term operations; a transition forecast addresses probability-weighted planning; a conditional stress forecast examines survivability and concentration even when the probability of a novel event is weakly identified.
An institution may employ several such views simultaneously. The decision memorandum should state their relation and avoid adding stress losses to a probability-weighted expected loss as though both were disjoint random events. When probability weights are unavailable or scientifically fragile, scenario analysis can present bounded consequences rather than numerical expected values falsely suggesting precision. Model uncertainty can be addressed through parameter intervals, predictive mixtures or a documented range of plausible structural specifications; the legitimacy of each range depends on its evidentiary construction.
Insurance pricing, claims reserves, and reinsurance
An annual gross premium typically reflects expected insured claims, expenses, risk loading, capital costs, and business or market considerations, subject to applicable insurance regulation. Under changing conditions, the expected claims component must correspond to the contract's future coverage period. A historical loss ratio mixes premium rates, portfolio composition, claims emergence, inflation, underwriting changes, and the loss mechanism. It can signal deterioration without identifying which component changed. The time-varying frequency example above illustrates how a stable-looking full-history average can understate the current expected loss under a stipulated persistent rate increase.
Prospective pricing should separate the estimation of current per-unit risk from the projection of future insured exposure. A risk-based tariff fitted on historical cohorts may require updated offsets for earned units and prospective adjustments for age, geography, insured value, policy terms, or other permitted variables. As more features are introduced, the effective sample size within a subgroup may become small, and fairness, comparability, and local legal restrictions remain substantive. A technically accurate trend estimate is not an authorization to use any variable for classification.
Reserve estimation has a different time index. A claim portfolio can incur obligations in one year, report claims in later periods, and settle over several additional years. If inflation, legal outcomes, claims handling, or reporting delays change, a historical development factor may extrapolate obsolete patterns. Reserving therefore requires separating occurrence experience, report development and payment escalation, with model assumptions reviewed against newer cohorts. A prospective best estimate of existing liabilities must reflect the relevant measurement framework and should not be confused with the premium for future, unearned risk. Under some methods, external covariates and explicit inflation scenarios can help distinguish a changing severity environment from slower settlement.
Reinsurance responds to both aggregate expected loss and the shape of the tail. The attachment and exhaustion of a layer determine which parts of the changing distribution the contract covers. A greater event frequency can increase the expected number of layer penetrations even if event severity is unchanged; increased severity may shift expected losses between the insurer and a reinsurer; and higher dependence can concentrate numerous risks into one event. Dynamic risk assessment should evaluate net and gross loss distributions, basis and counterparty exposures, reinstatement terms, aggregate limits, collateral, and the timing of recoveries. A reinsurance program designed for an old regime cannot be evaluated solely by comparing historic nominal catastrophe losses with a future nominal retention.
Prudential governance and capital
Insurance supervision connects quantitative modeling to continuing governance. The International Association of Insurance Supervisors' Insurance Core Principles online framework describes enterprise risk management, identification of risk interrelationships, and own risk and solvency assessment under ICP 16. The global supervisory framework is influential across jurisdictions, while its implementation and legal force depend on national and regional rules. The framework motivates a continuing view of material risk and its evolution, including risks that may not be represented adequately by a static measure.
In the European Union, Solvency II Article 45 establishes the own risk and solvency assessment as a part of the insurance undertaking's risk-management system and addresses overall solvency needs, continuing capital compliance, and possible differences between the undertaking's risk profile and assumptions in the solvency capital requirement. The precise legal text and its consolidated amendments should be checked for the relevant undertaking and assessment date. The Directive is an EU legal source, not a worldwide numerical rule, and a statistical counterexample in this monograph does not recalculate a regulated insurer's capital requirement.
EIOPA's 2022 guidance on climate scenarios in ORSA provides an official example of forward-looking materiality evaluation and longer-horizon scenario use. It was developed in relation to EIOPA's 2021 supervisory opinion and concerns EU insurance supervisory practice. Its methodological relevance reaches beyond one statistical fitting technique: different horizons and changing risk environments may require an integrated assessment rather than sole reliance on annual historical loss averages. The document does not establish a single universally correct climate probability trajectory or a legally identical obligation in all jurisdictions.
Capital planning under time-varying risk has at least three horizons. Immediate liquidity concerns the ability to pay claims when due; near-horizon solvency concerns resources relative to a specified capital-risk metric; and multi-year resilience concerns the sequence of losses, premium adjustments, replenishment and further shocks. One-year distribution calibration does not automatically validate the latter two. A dynamic stress program should trace how repeated losses and changing correlations affect assets, liabilities, contractual recoveries, liquidity, and future capital availability, documenting the limitations of assumed transitions.
Climate, longevity, and cross-sector illustrations
Climate provides a prominent physical instance of time-dependent hazard probability, exposure, and vulnerability. In their 2008 stationarity and water-management argument, Milly and coauthors challenged the continued default reliance on unchanging hydrologic statistics under climate change. The paper is a perspective on water management, and its conceptual relevance to insurance lies in the limits of extrapolating former hydrologic distributions to future insured hazards. It does not imply that every flood series everywhere has the same trend or that a particular insurer can dispense with quantitative probability models.
The IPCC Working Group II 2022 Summary for Policymakers explicitly evaluates interacting hazards, exposure, vulnerability, adaptation, and compound risks. For insurance, that distinction cautions against attributing an observed increase in insured economic loss to a single hazard trend. Population movement, higher property values, changing vulnerability and coverage expansion may contribute alongside changing physical extremes. The scientific question is whether the selected model identifies and projects each material component at the geographical and temporal scale needed for the decision. An insurer may experience increasing losses because exposed insured values grow even where a particular physical hazard frequency remains stationary.
Longevity illustrates a different form of change. The research by Brouhns, Denuit, and Vermunt (2002) places age-specific mortality in a Poisson log-bilinear regression framework and uses a time-varying mortality index for projected life tables and annuity valuations. Their source concerns specified demographic data and mortality modeling, not a universal projection of future mortality. It shows why an insurer valuing long-duration liabilities cannot rely automatically on one period's age-specific death rates. The age composition of an annuity portfolio and changes in the mortality law must be analyzed as distinct inputs.
Cyber claims can also respond to time-varying vulnerability, threat techniques, reporting practices, and coverage terms. The frequency of incidents recorded by an insurer may change as its detection service improves or as a new type of attack spreads; historical severity can also be censored by evolving policy limits. These mechanisms establish plausible pathways for structural uncertainty, not an empirical trend claim in the absence of a cited and comparable incident dataset. Cyber-specific assertions require contemporaneous incident, underwriting, and security evidence and are therefore left to a separate domain research dossier.
Several cross-sector lessons emerge. A physical risk can evolve slowly while financial exposure grows quickly. A human health process can change gradually while administrative reporting changes abruptly. A technology hazard can experience a sharp transition with very little representative post-transition data. An annual insurer risk model must respect each relevant time scale; transferring a trend coefficient from one sector or jurisdiction to another without measuring comparable mechanisms is not a valid synthesis.
A practical evidence-and-decision workflow
For a time-varying risk assessment, the institution can proceed through a sequence of documented tasks. First, define the insured event and the covered future horizon, including the date at which available information is frozen. Second, reconstruct historical exposure and administrative comparability. Third, inspect frequency, severity, development and dependence separately, noting observations removed or reclassified. Fourth, formulate a small set of plausible stationary, covariate-driven, break-aware, and regime-dependent candidates. Fifth, test them chronologically against an unchanged benchmark using proper probabilistic scoring and the decision-relevant tail metrics. Sixth, record remaining parameter and structural uncertainty and examine targeted scenarios. Seventh, connect the resulting forecast range to premium, reserves, capital, reinsurance, liquidity, and management decisions under applicable rules.
This sequence is an operational synthesis of established techniques, not a novel doctrine or exclusive methodology. Its purpose is evidentiary traceability. When a model is changed, a reviewer should be able to identify which observed fact, assumption, data correction, law, or new scenario caused the change, and whether the update improved subsequent out-of-sample performance. A model's label is secondary to whether its assumptions and consequences can be inspected and compared.
Decision reports can record both the most supported conditional forecast and a range of structural alternatives. A range can be quantitatively useful only when its construction is explicit: bounds obtained by varying an assumed intensity growth rate differ from confidence intervals based on a valid sampling model; scenario losses conditioned on a hypothetical high-risk state differ from unconditional quantiles. The terms “reasonable,” “prudent,” and “extreme” have different meanings across scientific and institutional settings. Each risk metric needs an unambiguous definition and horizon before interpretation.
Open scientific and measurement questions
Several issues remain genuinely open for future empirical work. Rare catastrophes provide little power to distinguish a permanent frequency change from a brief cluster when the event catalog and insured exposure have also evolved. Joint-tail models must balance physical knowledge against the instability of estimating many dependence parameters from sparse extremes. An insurer's claims series is generated by policy selection and claim reporting as well as by the underlying hazard, which complicates comparisons over time. Dependence between exposure growth and hazard evolution may be especially important for long-term adaptation, yet regional and contract-specific data can be incomplete.
Forward-looking evaluations face verification delays. A twenty-year climate or longevity forecast cannot be fully validated after one annual observation, although individual mechanisms and near-horizon implications can be checked sooner. A successful one-year forecast method may remain untested for transitions not yet observed. Bayesian updating and ensemble methods can represent alternative trajectories, but they cannot manufacture evidence about a completely new mechanism. Research quality therefore depends on explicit limits as well as formal precision.
Finally, it is necessary to distinguish model development from decision authority. The technical work can show that historical stationarity is an unsupported premise for a specified portfolio and horizon, or that a conditional distribution has changed within a documented model. It can also quantify consequences of plausible alternative states. The choice of insurance protection, prudential capital, legal classification, customer treatment, and public risk response requires institutional decisions grounded in relevant law, contracts, values, and governance. Transparent scientific inference improves these decisions by making their quantitative premises testable.
Conclusion
Nonstationarity is a property of the probabilistic object under study, relative to its time index, transformation, and information structure. Changes in conditional forecasts, unconditional loss distributions, latent regime probabilities, and the measurement of insured exposure are analytically different events. A statement that historical experience has become unreliable is scientifically meaningful when it identifies the particular loss variable, broken assumption, supported alternative explanation, and decision-relevant prediction horizon.
The formal results establish several practical consequences. A time-varying event intensity makes expected future counts depend on the integrated prospective intensity rather than on an undated historic average. A shared latent regime can leave large common-state variance in a portfolio whose individual risks are independent within each regime. Changes in joint dependence can materially alter aggregate loss distributions without changes in marginal variance. A small shift in an extreme-event probability can create a discontinuous change in a specified high-level VaR for discrete losses even when expected loss changes smoothly. Each of these conclusions follows from explicit standard probability identities or fully declared illustrative assumptions.
Statistical diagnostics provide evidence about specified hypotheses; they do not guarantee future invariance. Structural-break estimators, state-space and switching-regime models, extreme-value methods and conditionally heteroskedastic time-series models each answer bounded questions. The appropriate standard of evaluation is a chronological, information-consistent prediction experiment, supported by proper scoring, risk-specific tail assessment, and disclosure of estimation and structural uncertainty. Model improvement must be demonstrated for the relevant horizon and loss metric rather than asserted from in-sample complexity.
The insurance implications extend through premium adequacy, claims development, risk accumulation, reinsurance, capital, liquidity, and oversight. The IAIS supervisory framework and the applicable European sources recognize continuing risk management and forward-looking assessment, while jurisdiction-specific implementation governs actual obligations. Climate and mortality research supply distinct empirical motivations for projecting changing mechanisms; neither supplies an automatic universal calibration for every insurance exposure.
The original contribution of this research dossier is a documented analytical synthesis across stochastic processes, forecast evaluation, dependence, actuarial loss measurement and prudential decision horizons, together with explicit synthetic counterexamples and source-to-claim mapping. It introduces no new theory or proprietary model and makes no claim of empirical discovery. The durable scientific conclusion is that a risk estimate earns predictive meaning only by stating what loss is measured, for whom, over which period, conditional on what information, under which contemplated future probability law, and with what available evidence for that law.
References
Bai, J., & Perron, P. (1998). Estimating and testing linear models with multiple structural changes. Econometrica, 66(1), 47–78. https://doi.org/10.2307/2998540
Bollerslev, T. (1986). Generalized autoregressive conditional heteroskedasticity. Journal of Econometrics, 31(3), 307–327. https://doi.org/10.1016/0304-4076%2886%2990063-1
Brouhns, N., Denuit, M., & Vermunt, J. K. (2002). A Poisson log-bilinear regression approach to the construction of projected lifetables. Insurance: Mathematics and Economics, 31(3), 373–393. https://doi.org/10.1016/S0167-6687%2802%2900185-3
Cox, D. R. (1955). Some statistical methods connected with series of events. Journal of the Royal Statistical Society: Series B (Methodological), 17(2), 129–157. https://doi.org/10.1111/j.2517-6161.1955.tb00188.x
Davison, A. C., & Smith, R. L. (1990). Models for exceedances over high thresholds. Journal of the Royal Statistical Society: Series B (Methodological), 52(3), 393–425. https://doi.org/10.1111/j.2517-6161.1990.tb01796.x
Dickey, D. A., & Fuller, W. A. (1979). Distribution of the estimators for autoregressive time series with a unit root. Journal of the American Statistical Association, 74(366a), 427–431. https://doi.org/10.1080/01621459.1979.10482531
Diebold, F. X., & Mariano, R. S. (1995). Comparing predictive accuracy. Journal of Business & Economic Statistics, 13(3), 253–263. https://doi.org/10.1080/07350015.1995.10524599
Engle, R. F. (1982). Autoregressive conditional heteroscedasticity with estimates of the variance of United Kingdom inflation. Econometrica, 50(4), 987–1007. https://doi.org/10.2307/1912773
European Insurance and Occupational Pensions Authority. (2022, August 2). Application guidance on climate change materiality assessments and climate change scenarios in ORSA (EIOPA-BoS-22/329). https://www.eiopa.europa.eu/publications/application-guidance-climate-change-materiality-assessments-and-climate-change-scenarios-orsa_en
European Parliament and Council. (2009). Directive 2009/138/EC on the taking-up and pursuit of the business of insurance and reinsurance (Solvency II), Article 45, as subsequently amended. Official Journal of the European Union. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A32009L0138
Giacomini, R., & White, H. (2006). Tests of conditional predictive ability. Econometrica, 74(6), 1545–1578. https://doi.org/10.1111/j.1468-0262.2006.00718.x
Gneiting, T., & Raftery, A. E. (2007). Strictly proper scoring rules, prediction, and estimation. Journal of the American Statistical Association, 102(477), 359–378. https://doi.org/10.1198/016214506000001437
Hamilton, J. D. (1989). A new approach to the economic analysis of nonstationary time series and the business cycle. Econometrica, 57(2), 357–384. https://doi.org/10.2307/1912559
Intergovernmental Panel on Climate Change. (2022). Climate Change 2022: Impacts, Adaptation and Vulnerability. Working Group II contribution to the Sixth Assessment Report: Summary for Policymakers. https://www.ipcc.ch/report/ar6/wg2/chapter/summary-for-policymakers/
International Association of Insurance Supervisors. (2024, December update; online edition). Insurance Core Principles and ComFrame: ICP and ComFrame Online Tool, including ICP 16, Enterprise Risk Management for Solvency Purposes. https://www.iais.org/icp-online-tool/
Killick, R., Fearnhead, P., & Eckley, I. A. (2012). Optimal detection of changepoints with a linear computational cost. Journal of the American Statistical Association, 107(500), 1590–1598. https://doi.org/10.1080/01621459.2012.737745
Kwiatkowski, D., Phillips, P. C. B., Schmidt, P., & Shin, Y. (1992). Testing the null hypothesis of stationarity against the alternative of a unit root: How sure are we that economic time series have a unit root? Journal of Econometrics, 54(1–3), 159–178. https://doi.org/10.1016/0304-4076%2892%2990104-Y
Milly, P. C. D., Betancourt, J., Falkenmark, M., Hirsch, R. M., Kundzewicz, Z. W., Lettenmaier, D. P., & Stouffer, R. J. (2008). Stationarity is dead: Whither water management? Science, 319(5863), 573–574. https://doi.org/10.1126/science.1151915
Page, E. S. (1954). Continuous inspection schemes. Biometrika, 41(1–2), 100–115. https://doi.org/10.1093/biomet/41.1-2.100
Perron, P. (1989). The great crash, the oil price shock, and the unit root hypothesis. Econometrica, 57(6), 1361–1401. https://doi.org/10.2307/1913712