Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Nonstationary Risk and Forward-Looking Insurance Pricing

Abstract. Insurance prices refer to future obligations arising from contracts written today. The statistical sample used to estimate those obligations is historical; its relevance depends on how hazards, exposure, vulnerability, contractual coverage, judicial practice, technology, claims behavior, economic prices, and risk dependence evolve between the experience and coverage periods. This monograph investigates the actuarial identification, estimation, validation, and governance of prospective insurance costs under nonstationarity. It distinguishes distributional nonstationarity from ordinary trending, data deficiencies, shifts in insured portfolio mix, claims development, structural breaks, and uncertainty about the scenario itself. It analyzes frequency–severity models, exposure-adjusted loss costs, state-space and change-point approaches, physical catastrophe modeling, correlated cyber losses, social inflation, external information, predictive distributions, and reinsurance and capital effects. Transparent hypothetical calculations decompose an apparent 65% rise in loss cost, separate changing risk from changing portfolio composition, examine scenario weighting and gross premium adequacy, and demonstrate how delayed claim cash flows interact with discount-rate changes. An original source-critical synthesis connects an auditable forecast to an explicit estimand, time horizon, distribution of future exposure, scenario evidence, model uncertainty, regulatory constraints, and ex post recalibration. The central finding is that forward-looking pricing is the estimation of future insured cost conditional on an identified contractual and portfolio state, supported by tested forecasts and disclosures of structural uncertainty. Historical averages remain informative only to the extent that the mechanisms generating their observations remain transferable. No new named theory, doctrine, original empirical dataset, or claim of experimentally established predictive superiority is advanced.

Keywords: Nonstationary risk; prospective ratemaking; insurance pricing; structural change; distribution shift; hazard; exposure; vulnerability; loss trend; claims inflation; social inflation; catastrophe models; climate change; cyber accumulation; legal change; frequency–severity models; dynamic generalized linear models; scenario analysis; probabilistic calibration; model risk; credibility; insurer capital; reinsurance; insurance affordability; prudential regulation; actuarial governance.

Introduction

A premium quoted for tomorrow is an obligation valued today. For an insurer underwriting a policy covering the next twelve months, the observed claims of the previous twelve months are evidence about future cost, but their nominal sum is not the object to be priced. The scientific target is the distribution of future covered losses and the financing necessary to carry them under the contract expected to be in force. A stable rate indication requires a defensible bridge from experience to that future distribution.

In the U.S. professional setting, the prospective future-cost standard (ASB 2017, ASOP No. 53) explicitly addresses changes in historical conditions, expected frequency and severity trends, policy coverage, legal environment, and infrequent claims. This is the actuarial foundation for using history conditionally. The trending standard (ASB 2009, ASOP No. 13) recognizes both economic and social influences on projected values. Neither statement licenses mechanical extrapolation after a material shift in the process generating claims.

The question is historically broader than insurance. Milly and coauthors' critique of hydrological stationarity (2008) challenged treating past natural variability as an unchanging envelope. For insurance, that observation becomes relevant only after translating changing natural hazards into covered losses, including insured-location selection, insured values, vulnerability and contractual limits. The IPCC synthesis assessment (2023) establishes global evidence for human-caused warming and many changing extremes, while retaining variation in attribution and confidence by hazard and region. It does not supply a universal percentage increase to apply to insurance premiums.

The scholarly task is therefore to determine what aspect of the risk process has changed, how evidence identifies that change, and how the change should appear in a prospective cost distribution. It is not sufficient to label an old estimate inaccurate. A credible replacement must state the exposure it prices, the event period, the information set, the price level and the scenario assumptions. This dossier develops those requirements as an integrated analytical account of existing actuarial and statistical methods.

Scientific Object, Definitions, and Research Method

The scientific object is the conditional future loss-cost distribution for a specified portfolio and coverage period under evolving underlying states. Let t denote calendar time, i a policy or risk unit, X_i(t) its exposure, H_t a hazard process, V_i(t) its susceptibility to the hazard, C_i(t) contractual coverage, and M_t the economic and institutional claims environment. The intended prospective estimand is the expectation, or an explicitly selected alternative functional, of covered claim costs for a future period given the information available at pricing time. Time may appear both through covariates and through changes in their effects.

A process is strictly stationary if its joint distributions are unchanged under shifts in time; weak stationarity normally concerns a constant mean and covariance dependent only on lag, given finite second moments. Insurance pricing typically does not require either mathematical property of every variable. It requires that the relevant predictive relationship be transferable, after justified adjustments, from observed periods to the prospective risk. Conditional stability of claim severity for comparable risks can coexist with changing population composition. Conversely, apparent stability of aggregate losses may conceal offsetting increases in hazard and improvements in protection.

Nonstationarity may result from gradual trend, discontinuous regime change, changing volatility, new dependence, shifts in the conditional relation between covariates and losses, or selective entry and exit from the insured sample. Covariate shift means the distribution of risk characteristics changes; concept drift means the conditional outcome relationship changes. These terms are useful descriptions, not independent evidence of a cause. Exposure drift can be measured with geographic and coverage data, while an unobserved shift in susceptibility may be only partially identifiable.

Four accounting distinctions govern the analysis. Loss development is the evolution in recognized value of claims attached to an existing accident or underwriting period; trend is the change in corresponding expected costs across exposure periods. Inflation changes monetary severity and may have different general, construction, medical, wage and legal components. Risk mix changes the composition of insured contracts even if the risk cost of each homogeneous class is unchanged. ASOP No. 13 specifically distinguishes trend from development. A nonstationary claim estimate may require all four adjustments, but they must enter at their respective causal and accounting positions.

Methodologically, this study combines a close reading of adopted actuarial standards, official regulatory and climate-risk statements, primary scholarly forecast-evaluation papers, and one scholarly cyber-insurance literature review. It develops mathematical identities, hypothetical examples and scenario sensitivity tests rather than fitting a proprietary insurance dataset. The cited sources establish professional scope, observed institutional findings and specific statistical concepts. Derived calculations are independently reproducible from assumptions displayed below. Comparisons across the United States, European Union and global supervisory guidance concern the nature of their instruments rather than an assertion of uniform legal force. The research cutoff is November 15, 2026.

Stationarity Assumptions in Ratemaking

Historical experience as conditional evidence

Suppose a portfolio has n past exposure units and total developed incurred losses L. The naive loss cost is L/n. Even after claims are developed to their expected ultimate value, L/n is an estimate for the historical sample's coverage, risk mix, hazard distribution, claim settlement system and monetary level. It is a forecast for a new population only under a transportability assumption. That assumption may be reasonable for relatively stable short-tail risks after modest adjustments; it is weaker for geographically concentrated catastrophe coverages, rapidly changing cyber exposures and long-tail liability claims subject to legal developments.

The ASOP No. 53 historical-data provisions (ASB 2017) require attention to exposure consistency, trends and changes in law, contract, business mix, claims operations and reinsurance. This creates a practical map of possible departures from transportability. The forecaster should identify which historical data remain informative for each component and when external analogues or explicitly hypothetical scenarios replace the unsupported historical assumption.

A useful generic decomposition is expected covered claim cost per policy-period = expected claim count per unit of exposure × expected paid severity per claim, when severity is conditioned consistently on the event definition and applicable coverage. More generally, with segments j and future exposure weights w_j, prospective loss cost equals the sum over j of w_j times the segment-specific future frequency–severity product. This identity can be expanded into catastrophe and attritional components, limits and deductibles, claim expenses and ceded recoveries.

The decomposition is an accounting and probability identity, not a new predictive model. Its value is diagnostic: claims counts can rise because the underlying peril occurs more often, because more properties are located in exposed areas, because claims reporting rises, or because the insurer sells more policies. Each interpretation implies different evidence and a different response.

Historical windows and structural breaks

Long historical series reduce sampling variance when their observations are comparable. Under structural change, older observations can increase estimation bias. Short windows adapt more quickly but may be dominated by randomness, especially for severe events. Neither recency weighting nor a single detected break resolves the problem automatically. A decision about the estimation window should be evaluated on prospective predictive performance and source-level plausibility.

Consider a portfolio whose claim probability rises permanently after a safety regulation is relaxed. A ten-year average spanning the former regime systematically understates the new probability if the regimes are pooled without adjustment. Now consider a single unusually severe storm under an unchanged hazard regime: discarding all earlier years because one year was extreme exaggerates the evidence for a structural break. The appropriate inference depends on a model of the event-generating process, independent explanatory information and the uncertainty of the change-point estimate.

A simple formalization specifies a latent risk state z_t and observations Y_t. State-space methods can posit z_t = z_(t-1) + eta_t and a conditional loss model Y_t | z_t, X_t. Sudden jumps can be represented with regime indicators or alternative change-point mechanisms. Their innovation distributions, priors, initializations and identification requirements matter. A stochastic state that responds rapidly to noise will overfit catastrophe occurrences, while one that barely moves may conceal a regime change. Model selection should reflect the decision horizon rather than retrospective graphical smoothness.

Exposure, claims and economic timing

An exposure unit must correspond to the insured obligation: vehicle-years, property-years, payroll, revenue, insured value or another economically relevant denominator. When deductibles, policy limits or covered activities change, a constant count of policies does not mean equivalent risk. Inflation-indexing insured sums can increase a nominal insurer payout even when physical damage is unchanged. Conversely, increasing a deductible can reduce paid severity while leaving total economic losses intact.

The distinction between accident year, report year, calendar year and policy year is fundamental. Inflation in the payment year may affect settlement of claims from earlier accident years. Calendar-period triangle effects can coexist with distinct accident-year changes. Failure to separate them produces double counting: a modeller can add a severity trend already implicit in claim development and then treat the combined increase as new risk.

ASOP No. 23 provides data-quality responsibilities (ASB 2016). Useful checks include stable event definitions, geocoding confidence, loss and expense coding, duplicate events, coverage transformations, policy-limit censoring, missing deductibles, subrogation, changes in claim closure and reporting lags. Small geographic cells with zero losses require exposure and tail information; zero is an observation, not proof of zero risk.

Credibility and relevance

A credibility blend between an individual insurer's experience and an external benchmark is useful when the latter is predictive. The adopted credibility procedures standard (ASB 2013, ASOP No. 25) emphasizes the choice of relevant experience. In a hypothetical blend Z × 1,300 + (1 − Z) × 1,100 with Z = 0.4, the resulting indicated cost is 1,180 per exposure. This arithmetic is correct, but its forecast is only as defensible as the externally supplied 1,100. If that complement reflects an obsolete climate, legal or technological state, increasing credibility in the complement can aggravate rather than cure forecast bias.

Credibility is thus conditional on comparability. A heavily observed but structurally obsolete portfolio may have low prospective relevance. A new external dataset with few years but accurate current hazard and exposure characteristics can sometimes be more informative, although it may have wide estimation intervals. Statistical sample size cannot substitute for validation of the relationship between the sample and the future risk.

Changing Hazard and Exposure

A prospective frequency–severity representation

Let N_i be the claim count for exposure i over the priced period and S_ik its positive covered severity. The paid loss is L_i = sum from k = 1 to N_i of S_ik. Under conditional independence of count and identically distributed severity given the future state, E[L_i | state] = E[N_i | state] × E[S_i | claim, state]. When count and severity share a random driver, the unconditional expectation instead integrates their conditional product over that driver; multiplication of unconditional means may be wrong.

For example, a frequency GLM could use log E[N_i] = log e_i + alpha + beta x_i + gamma(t), with log exposure e_i as offset. A separate positive-severity model estimates E[S_i | claim, x_i, t] subject to limits and deductibles. Time-varying coefficients or latent states can express more complex changes. Any improvement in fit should be tested on future-origin holdouts because calendar time can serve as a proxy for unrecorded claims-processing changes.

A physical peril model may instead simulate annual event frequency, event footprints, local intensity, asset exposure, vulnerability and financial coverage, generating a distribution of insurer payouts. This becomes prospective when hazard rates, coastlines, land use, defenses and exposure valuation correspond to the coverage horizon. A climate-model output is not itself an insured loss projection. Physical damage, policy conditions, claims behavior and insurance penetration complete the translation.

Independently reproducible loss-cost calculation

Consider one homogeneous property class with historical annual claim probability 0.08 and mean covered claim severity 12,000 monetary units. The historical expected loss cost is 0.08 × 12,000 = 960 per exposure-year. Suppose a clearly labeled hypothetical forward scenario changes the probability to 0.11 and covered mean severity to 14,400, while definitions and limits are held consistent. The prospective expected loss cost is 0.11 × 14,400 = 1,584. The increase is 624, or 65% relative to 960. The frequency factor is 0.11/0.08 = 1.375, and the severity factor is 14,400/12,000 = 1.2; their product is 1.65.

These are invented assumptions, not observed global catastrophe rates or an actuarial indication for a real insurer. The example isolates a crucial inferential problem. A 65% observed cost increase could arise through many different frequency–severity combinations, and the observed loss total alone cannot identify the decomposition. It may further hide changing risk mix, insurance limits, deductibles and reporting. Accurate decomposition requires event, exposure and claim records under comparable coverage.

Now assume two fixed classes, A with an expected covered cost of 900 and B with 1,800. A portfolio initially weighted 80% A and 20% B has an expected loss cost of 1,080; if its risk mix shifts to 50% each, the expected cost becomes 1,350, a 25% increase. The class-specific risk has not changed. Treating the increase as an across-the-board 25% hazard trend would misprice both classes and obscure the role of underwriting and sales strategy.

Variance, dependence and catastrophe tails

If counts are conditionally Poisson with intensity lambda and independent severities have mean mu and variance sigma², compound-Poisson variance is lambda × (sigma² + mu²). Cross-policy independence would then permit aggregation of individual variances. Shared weather, infrastructure, legal, supply-chain or cyber drivers introduce covariance terms. For a portfolio sum T = sum L_i, Var(T) = sum Var(L_i) + 2 × sum over i<j of Cov(L_i,L_j). This identity shows why the same average loss cost can coexist with radically different capital and reinsurance needs.

Nonstationarity may affect the tail more rapidly than the mean. A future event footprint can intersect newly constructed assets, while the policy average changes only modestly. Dependence can strengthen under shared infrastructure, compound flooding and wildfire-weather conditions, or weaken through adaptation. In catastrophe ratemaking, ASOP No. 39's treatment of catastrophe losses (ASB 2000, revised 2011) reinforces the need for specialized methods rather than raw short-period event averaging.

An annual average loss and a selected exceedance probability describe different distributional characteristics. A rate that recovers expected insured loss is not necessarily sufficient for ruin protection or reinsurance capacity. Additional capital, reinsurance, risk margins and profit requirements depend on a defined financial objective, capital structure and risk appetite; they are not mathematically identical to an arbitrary percentage loading.

A forecast may deteriorate despite better historical fit

Suppose a static GLM is trained on a period with stable hazard and achieves low in-sample error. A new model introduces a climate covariate with little historical variation. The latter may fit worse historically while being scientifically more plausible for future conditions. That plausibility must still be tested: its coefficient could be weakly identified, its scenario input could be biased, and the catastrophe tail could be dominated by assumptions. No principle says that adding a forward-looking covariate automatically improves predictive accuracy.

Conversely, a complex machine-learning model can exploit information leaked from claim settlement into a supposedly prebinding policy record and appear predictive. When the variable is unavailable at quotation time, the apparent skill disappears. The prospective information set must therefore be specified at the exact pricing decision date, with data cutoffs and known reporting delays.

Climate and Social Inflation

Climate signals and insured cost translation

Physical climate science constrains plausible hazard scenarios; it does not assign a market rate. The IPCC climate synthesis (2023) supports the conclusion that human-caused climate change influences many extremes. The strength, direction and geographic scale of conclusions differ by peril. A frequency adjustment for heat-related insured events cannot be automatically copied into winter storms, floods or tropical cyclone counts. Inadequate spatial or peril disaggregation makes a superficially scientific price adjustment unidentifiable.

For a geographic property portfolio, risk can be represented through an event catalogue, hazard footprint, localized insured values, vulnerability curves and policy-financial terms. Changes in precipitation intensity or sea level enter the hazard component; movement of property development into exposed areas enters exposure; flood barriers and resilient construction enter vulnerability; retentions and limits enter coverage. Attribution to climate for an observed loss-cost movement requires separating these contributions and their interactions.

EIOPA's non-life climate underwriting report (2021) describes an important economic feedback: if premiums reflect heightened risks, insurance can become less affordable; prevention actions can improve resilience and insurability. Such prevention works through vulnerability and sometimes exposure. It is analytically wrong to model a resilient building as if it changed the occurrence rate of a regional storm. It reduces the damage distribution conditional on that storm.

The IAIS climate supervision application paper (2025) frames climate risk in existing supervisory principles. EIOPA's 2025 monitoring findings also report growing scenario analysis in insurers' own risk assessments, with remaining data granularity and horizon challenges. These documents support the need for decision-useful climate scenarios, while their own descriptive findings do not establish the predictive superiority of any particular peril model.

Hazard horizon and policy horizon

A ten-year climate scenario may be decision-relevant for asset allocation and long-lived insurance availability but should not automatically be loaded in full into a twelve-month insurance contract. The policy horizon, renewal flexibility and plausible path of hazard changes matter. Conversely, a one-year property policy written in an area of accelerating hazard can face material change within the period, and a long-term guaranteed product can lock assumptions beyond ordinary annual renewal cycles.

Actuarial pricing should therefore identify the contract inception date, coverage duration, expiration and expected claim settlement timing. A multi-year financial plan can use scenario trajectories across future underwriting years, while each binding rate indication corresponds to the liabilities and pricing rules of its own relevant term. Cross-period averages are informative only when their time weighting matches the intended cash flows and insurance exposure.

Economic inflation and legal-social change

Inflation has several layers: reconstruction inputs, vehicle parts and labor, health care, wages, legal defense and claim settlement. A single general price index may fail when insurance severity responds to sector-specific constraints. Claims inflation can also be affected by repair technology, scarcity of contractors and claims handling. The ASOP No. 13 social-influences definition (ASB 2009) explicitly includes shifts in court practice, legal precedents and claim consciousness within actuarial trend considerations.

The Swiss Re Institute's liability excess inflation analysis (2026) uses excess claim-severity growth beyond economic factors to discuss non-economic drivers of liability costs, and finds variation across countries. That is an insurer research institute's analytical measure; a claim that one variable caused a specific jurisdiction's verdict trend would require court-level evidence and a causal design beyond the material inspected here. In ratemaking, a change in tort doctrine affects the probability and magnitude of legal liability, coverage interpretation, the cost of defense and settlement incentives, sometimes with substantial delay.

Suppose an ordinary severity index implies 6% annual change for two years. A 1,000-unit base rises to 1,000 × 1.06² = 1,123.60. A separately substantiated permanent 20% shift in the claim-cost relation would imply 1,200 on the same base before any additional trend. If both mechanisms act independently at the stated dates, the combined figure is 1,348.32. If the 20% measurement already includes those two years of inflation, multiplying again double counts it. Numerical transparency cannot replace evidence about which effects are additive, multiplicative, overlapping or displaced in time.

Claims development and late settlement

Long-tail liability lines expose the distinction between occurrence timing and cash payment. A claim from a past accident year can be paid under future wage, legal and medical prices, while policies written today generate claims under an evolving legal system. An accident-year ultimate estimate cannot be inserted into a present-value analysis without a consistent payment pattern and economic basis. Development assumptions and future cost trends can be jointly sensitive to legal changes.

As of the research cutoff, revised ASOP No. 20 is in effect for covered U.S. actuarial work products issued on or after June 1, 2026. Its current cash-flow standard (ASB 2025) covers discounted and undiscounted property/casualty cash-flow analyses, including prospective underwriting studies. It does not require discounting or a risk margin in all analyses. The act of discounting and the choice of risk margin must remain tied to the objective and relevant requirements.

For a hypothetical payment of 1,000 due exactly three years hence, present value at a flat annual effective 5% is 1,000/1.05³ = 863.84, while at 3% it is 1,000/1.03³ = 915.14. These computations isolate discount-rate sensitivity; a real claim portfolio has a distribution of payment amounts, dates, inflation paths, recovery, investment risk and other financial assumptions. The expected amount, its tail and the asset-liability financial effect must be evaluated together.

Availability and affordability as empirical consequences

The NAIC's 2026 homeowners market analysis reports state-level variation in U.S. homeowners premiums, claims and nonrenewals from 2018 through 2024. Those are historical market observations; the report does not identify a single causal share of climate change in all changes, and it is not evidence for 2025 or 2026 claim outcomes absent additional data. The 2026 nationwide data call seeks more granular information on costs, coverage, availability and mitigation. A request for data should not be cited as though its eventual findings were already known.

EIOPA's natural-catastrophe protection-gap dashboard provides a separate historical perspective on insured and economic catastrophe losses. A rising insured-loss total may reflect improved take-up, higher property values, adverse hazards or a mix of these effects. An insurance protection gap is a social and economic outcome shaped by both insurance supply and demand; even actuarially adequate prospective prices can coexist with an expanding gap.

Emerging Technologies and Cyber

New exposures with obsolete analogues

Technological change may alter the number, severity, dependence and observability of insured events. A vehicle safety feature can reduce certain collision types while increasing repair complexity and unit replacement costs. Connected building systems can aid prevention while introducing shared software vulnerabilities. A new cyber coverage may appear to have few past claims because prior contracts excluded an event class rather than because the peril was rare.

The research literature summarized by Eling and Schnell's cyber insurance review (2016) identifies sparse loss data, changing threats and accumulation risk as central obstacles. The paper is used here at the level of its verified abstract and bibliographic description, not to imply replication of its literature search. The correct response to missing data is to expose assumption sensitivity and consider analogous information whose relevance is articulated, rather than converting missing experience into a zero-loss estimate.

Dependency and accumulation

A common provider outage can produce claims at many independently underwritten firms. Portfolio-wide frequency and severity can change through increasing dependence even when each firm's expected individual loss changes little. Conventional diversification breaks down if the same cloud platform, authentication service, software vulnerability or payment network connects nominally distinct insureds. Scenario design must examine common-cause events, contractual aggregation language and policy sublimits.

A useful conceptual distinction is between within-insured correlation and across-insured correlation. Cyber controls may improve resilience for one insured while the same software standard makes many insureds exposed to a shared exploit. The first affects expected local loss and perhaps attachment; the second affects insurer tail aggregation and reinsurance. A single portfolio-average cyber frequency fails to express both.

Automated models under distribution change

Automated underwriting can fit highly nonlinear interactions and extract weak signals from granular data, but training performance depends on the environment in which examples were generated. Model drift can arise because users change behavior in response to pricing, because reporting or data suppliers change, because the claim process changes, or because a previously predictive proxy ceases to be predictive. External data may improve timeliness and simultaneously create legal, accuracy and fairness risks.

The IAIS 2025 AI supervision paper addresses accountability, robustness, explainability, ethics and redress within the global supervisory framework. For an actuarial forecast, governance includes verifying that features exist at quote time, that their definitions remain stable, that sensitive or protected attributes are treated according to applicable law, and that automated recalibration has human accountability.

Technology and contract interaction

Cyber pricing is particularly sensitive to how an event is defined. A breach event, a ransomware interruption, a third-party notification expense and a business-interruption loss have different coverage triggers, deductibles and limits. A change in the proportion of policies including contingent interruption can change the insurer's aggregate exposure even if external cyberattack activity remains unchanged. It is therefore necessary to rebuild the historical loss record under current contract terms or construct an auditable bridge.

The central inference is that technology is both a risk driver and a measurement instrument. A new reporting system can increase observed claim frequency without altering underlying incident frequency. Rapid adoption of a preventive device can reduce physical events while increasing detected near misses. A data-generating intervention should be documented before observed changes are interpreted as peril changes.

Scenario and External Risk Data

The evidence hierarchy of prospective assumptions

A forward-looking model may use claims and earned exposures, external catastrophe records, weather and climate scenarios, building-stock data, macroeconomic projections, legal developments, engineering tests and expert elicitation. Each dataset measures different objects. Hazard science supplies information about event intensity and occurrence; claims records supply conditional covered payouts; policy records supply the mapping from physical and legal events to insured obligations.

The ASOP No. 56 modeling standard (ASB 2019) calls for scrutiny of model structure, risks of overfitting and inputs. The practical consequence is that external science should enter only through a documented interface. If a hazard vendor supplies annual flood depths but the insurer has no matching geocoding, floor-level information or vulnerability assumptions, a precise loss number would conceal data gaps rather than resolve them.

A source record should include provenance, last update, geographic scope, measurement unit, collection method, revisions, uncertainty, permitted use, validation history and applicability to the insured population. When multiple datasets disagree, the discrepancy deserves explicit reconciliation. One source may measure total economic losses and another insured payouts; one may use local currency and another inflation-adjusted international dollars. These differences are not forecast errors in the same variable.

Probabilistic scenarios versus imposed weights

A scenario is a coherent description of a possible future, with disclosed assumptions and consequences. Its weight is a probability assertion if treated as one. Experts can compare plausible futures without claiming well-calibrated probabilities for each. Weighted averages should therefore be named probabilistic only when the weighting has a reasoned predictive interpretation. Otherwise the outputs are conditional scenario indications.

For an entirely hypothetical three-state cost distribution, let the cost per exposure be 1,000 with weight 0.5, 1,400 with weight 0.3 and 2,200 with weight 0.2. Its mean is 0.5 × 1,000 + 0.3 × 1,400 + 0.2 × 2,200 = 1,360. The 90th percentile, using the usual smallest-value-with-CDF-at-least-0.9 definition, is 2,200. The scenario mean, a 90th-percentile requirement and a worst-case 2,200 cost answer different questions. None is automatically a legally permitted filed premium.

If the weights reflect expert judgment rather than historical frequencies, the output must state that fact. More extreme alternatives may need separate stress testing instead of being assigned invented probabilities. A scenario set bounded by a narrow choice of hazards can be falsely reassuring even when it generates a wide statistical confidence interval conditional on those hazards.

Robustness and model uncertainty

Ordinary parameter intervals condition on a particular model form and dataset. Model uncertainty considers plausible changes to the structure, covariate pathways, link function, regime assumptions and dependencies. Scenario uncertainty concerns the external states themselves. These concepts overlap but should be documented separately. For example, a flood severity parameter may be uncertain conditional on a climate scenario, while future flood defenses and population shifts belong to external scenarios and model structure.

A practical sensitivity matrix changes one dimension at a time and then tests meaningful interactions: hazard frequency, event intensity, asset growth, vulnerability, deductible, inflation, limits, reinsurance attachment and coverage take-up. One-factor sensitivities alone fail when the effects interact, such as an increase in both hazard intensity and property value within an excess layer. The outcome should include the estimated mean, tail metric and financial impact for each scenario; a single composite score obscures the mechanism.

Validation under temporal dependence

Randomly shuffling policy records across all years produces leakage when future claims share catastrophe events, trends or claims administration with training observations. A rolling-origin design trains using data available by time t and evaluates later outcomes without using information that had not yet existed at the earlier decision. For slow-developing claims, valid evaluation must respect reporting and settlement maturity. An apparently low-error forecast that has observed more of a claim's future than the real pricing process would have observed is not an ex ante test.

Gneiting, Balabdaoui and Raftery's probabilistic forecasting research (2007) distinguishes calibration and sharpness and discusses proper scoring. A forecast can be narrow yet systematically wrong; a broad forecast can appear reliably covering outcomes while being uninformative. Evaluation therefore requires both calibrated predictive statements and useful discrimination. Insurance tests should include class-conditional and geographic calibration, not only portfolio averages.

Diebold and Mariano's forecast-accuracy comparison (1995) provides a statistical framework for comparing competing predictions under chosen loss functions. Its validity depends on how loss differentials, dependence and evaluation samples are treated. This dossier does not conduct that test; citing it identifies an appropriate family of methods. For rare catastrophes, a few years may be inadequate for a decisive head-to-head forecast ranking, and stress tests of physical consistency can provide complementary evidence.

Loss functions and decision relevance

A model that minimizes mean absolute error of claim costs need not optimize the risk of insolvency, rate stability or protection of vulnerable policyholders. A proper scoring rule assesses distribution forecasts; a business decision requires a loss function reflecting relevant economic consequences. Mean cost is central to expected-loss adequacy, while tail risks inform reinsurance and capital. Predictive quantiles should not be confused with statistical confidence bounds around an estimated mean.

EIOPA's October 2026 risk dashboard explicitly separates current risk levels, backward-looking risk trends and forward-looking outlooks as distinct information products. The broader analytical lesson is to document which period and kind of inference each signal supports. An inflation outlook is not an observed claims development factor, and a backward-looking premium increase is not an independent forecast of tomorrow's catastrophe costs.

Reporting irreducible limits

When external climate, litigation or technology assumptions cannot be empirically validated within the exposure horizon, the scientific output should report this limitation as part of the forecast. A wide scenario interval is not a substitute for missing drivers or contract-mapping errors. The highest integrity is achieved when the decision maker can see exactly which conclusions depend on unverifiable assumptions, rather than a claim that the model has absorbed all uncertainty.

Rate Revision and Governance

From conditional loss cost to gross premium

A pure expected loss cost omits loss adjustment expense, fixed administration costs, premium-based expenses, capital remuneration and reinsurance-related costs. The adopted expense-provision standard (ASB 2023, ASOP No. 29) covers expense provision judgments. The profit and capital standard (ASB 1997, revised 2011, ASOP No. 30) makes clear that capital cost and contingency provision have distinct functions. The precise appropriate measure depends on the assignment and regulatory environment.

For an illustrative premium per exposure, assume expected insured loss 1,360, loss adjustment costs 140, fixed costs 50, and explicitly allocated reinsurance cost 100. Total per-exposure non-premium-proportional cost is 1,650. Assume premium-based expenses equal 20% of premium and a target underwriting margin equals 5% of premium, with no other items. Solve P = 1,650 + 0.20P + 0.05P; hence P = 1,650/0.75 = 2,200. This is an arithmetic illustration, not an empirical commercial premium, an approved regulatory rate or a universally valid pricing formula. It presumes that the reinsurance cost is incremental and not already captured elsewhere, and the margin assumption is supplied rather than derived from a capital model.

Where contracts provide multi-period coverage, the equation requires payment and premium timing, discounting choices, capital requirements and possibly policyholder options. U.S. ASOP No. 20's 2026-effective cash-flow provisions apply within its scope to both discounted and undiscounted P/C cash-flow analyses. Their existence does not turn an actuarial gross premium formula into a financial reporting measurement under IFRS 17.

Regulatory boundaries and financial reporting

Insurance ratemaking is regulated at jurisdictional level, and permissible classification, rate filings, solvency treatment and policyholder communications vary. U.S. ASOPs are professional standards for covered actuaries; IAIS application papers provide global supervisory guidance; EIOPA materials address European supervision; and national laws determine legally binding rate and conduct requirements. An insurer cannot infer a universal right to charge a modeled cost from a global standard.

The IFRS Foundation's IFRS 17 requirements concern recognition and measurement of insurance contracts for financial reporting, including fulfillment cash flows and a contractual service margin, subject to applicable adoption. A filed price, an economic capital requirement, an IFRS 17 liability and a regulatory solvency amount are different estimands even when they share projected cash flows. Each requires its own measurement date, assumptions and accountability.

The IAIS climate supervisory guidance encourages climate-related risk management in the existing insurance regulatory framework. Its use in this dossier is comparative and analytical. The report does not create binding premium-setting rules in all IAIS member jurisdictions.

Model change control and decision accountability

A legitimate rate revision has a recorded reason: changed exposure, observed claim costs, validated trend, scenario assumptions, contract terms or capital need. Governance needs a versioned data snapshot, model implementation, author and review signoffs, comparison to the previous indication, assessment of distributional consequences, regulatory filing status where applicable, and a trigger for recalibration.

A change log should distinguish updated parameters within a stable model from a redesign of the model's structure. Major redesign requires independent validation of data lineage, conceptual plausibility, implementation, sensitivity and monitoring. Validation teams should challenge not only coefficient estimates but also the definition of claim, the forecast horizon, the role of external models, uncertainty and the scope of management overrides. ASOP No. 56's model governance guidance provides relevant U.S. professional context.

Monitoring metrics should be defined in advance: realized-to-expected loss frequency and severity by cohort, mature accident-year loss ratios, geographic calibration, renewal-selection effects, claim development residuals, dispersion, exceedance calibration and rate adequacy. The actual claims-to-premium comparison needs to reflect written versus earned timing; otherwise inflation and cohort mix can create false alerts.

The price–selection feedback

Higher premiums can change who buys insurance. High-risk policyholders may be unable to afford coverage, lower-risk policyholders may self-insure, and insureds may change deductibles or preventive behavior. Conversely, more granular risk prices can incentivize mitigation. Future insured exposures therefore respond endogenously to the rate and insurance availability. The next period's loss portfolio cannot always be projected with historical customer weights held fixed.

This feedback can be described as a sequence: risk-informed price, policyholder response, changed risk mix, realized losses, renewed price. The sequence is a qualitative economic mechanism, not an independent named model. Quantifying it would require elasticity estimates and a causal design measuring take-up, behavioral responses, coverage and income constraints. Historical premium changes alone do not identify demand elasticity because prices often rise where loss costs also rise.

EIOPA's climate impact-underwriting assessment proposes incentives for adaptation through underwriting and pricing. The NAIC homeowner data program likewise illustrates the need to observe affordability, insurance access and mitigation in granular terms. Governance must evaluate both expected insurance cost and the market consequences of how that cost is allocated.

An integrated analytical assessment

The source-based synthesis of this dossier establishes an ordered set of substantive questions for every forward-looking indication. First: which insured future loss and financial quantity is being estimated? Second: which historical exposures and claims are comparable, after development and monetary adjustment? Third: which drivers alter conditional frequency, severity, dependence and coverage? Fourth: which future scenarios and probabilities are supported, and which remain judgmental? Fifth: how do model uncertainty and heterogeneity affect the decision's downside? Sixth: which rate, capital, contractual and conduct constraints govern implementation? Finally: what observation and governance process will determine whether the forecast remains defensible?

This ordering is a method of evaluating existing evidence, rather than a new authorial system or theoretical framework. Its central contribution is clarification of the distinct inferential links. A model can satisfy one link and fail another: a physically sound hazard model may have inaccurate insured-value data; a well-fitted loss development model may assume an obsolete legal environment; a correctly computed gross rate may ignore adverse selection; and a well-documented model may use a probability-weighted scenario mixture whose weights have no empirical foundation. The remedy is traceable qualification and revision of the affected component.

Conclusion

Nonstationary risk requires prospective insurance pricing to make the transfer from historical evidence to future contractual cost explicit. The mathematical starting point is a specified insured population, coverage period and expected loss distribution, decomposed where possible into exposure, event frequency, covered severity, claim handling, dependence and financial terms. Ordinary loss development, monetary trend, changes in portfolio composition and shifts in the physical or institutional risk-generating process can occur simultaneously, and their effects must be identified separately before any combined indication is credible.

The strongest forecasts connect actuarial records to the mechanisms behind future claims, use external physical and economic evidence at the appropriate scale, state assumptions behind scenarios and probabilities, compare predictions through temporal holdouts and distributional calibration, and examine tail, capital and availability consequences. Model sophistication increases the need for traceable assumptions, particularly when data are sparse or the past environment has ceased to be predictive. Case studies here establish the numerical consequences of specified hypothetical assumptions, not empirical premium changes or new universal laws.

The professional, financial and supervisory documents reviewed provide complementary responsibilities. U.S. actuarial standards address prospective future costs, trends, modeling, catastrophe losses, expenses and current cash-flow analysis; IPCC and other scientific sources inform hazard assessments; IAIS and EIOPA address relevant risk governance; and jurisdictional law determines rate and conduct constraints. IFRS 17 governs financial-reporting objectives distinct from commercial rate indications. These differences are substantive and must remain visible in international scientific writing.

The final analytical conclusion is that a forward-looking rate is a claim about a future distribution conditional on a dated information set, a defined insurance contract and an identifiable population of exposures. Its scientific quality rests on the strength of the evidence supporting that conditional claim, the clarity of its uncertainty, and the possibility of revising it against subsequent observations. That conclusion is a critical synthesis of established risk science and actuarial practice. No novel named theory, proprietary empirical finding or verified causal impact is claimed.

References

  • Actuarial Standards Board (ASB). 2017. Actuarial Standard of Practice No. 53: Estimating Future Costs for Prospective Property/Casualty Risk Transfer and Risk Retention. Adopted December 2017, effective August 1, 2018. Official source or scholarly record.

  • Actuarial Standards Board (ASB). 2009, revised 2011. Actuarial Standard of Practice No. 13: Trending Procedures in Property/Casualty Insurance. Official source or scholarly record.

  • Actuarial Standards Board (ASB). 2016. Actuarial Standard of Practice No. 23: Data Quality. Effective April 30, 2017. Official source or scholarly record.

  • Actuarial Standards Board (ASB). 2013. Actuarial Standard of Practice No. 25: Credibility Procedures. Effective May 1, 2014. Official source or scholarly record.

  • Actuarial Standards Board (ASB). 2000, revised 2011. Actuarial Standard of Practice No. 39: Treatment of Catastrophe Losses in Property/Casualty Insurance Ratemaking. Official source or scholarly record.

  • Actuarial Standards Board (ASB). 2019. Actuarial Standard of Practice No. 56: Modeling. Effective October 1, 2020. Official source or scholarly record.

  • Actuarial Standards Board (ASB). 2025. Actuarial Standard of Practice No. 20: Analysis of Property/Casualty Cash Flows, Including Discounting. Effective June 1, 2026. Official source or scholarly record.

  • Actuarial Standards Board (ASB). 2023. Actuarial Standard of Practice No. 29: Expense Provisions for Prospective Property/Casualty Risk Transfer and Risk Retention. Effective July 1, 2024. Official source or scholarly record.

  • Actuarial Standards Board (ASB). 1997, revised 2011. Actuarial Standard of Practice No. 30: Treatment of Profit and Contingency Provisions and the Cost of Capital in Property/Casualty Insurance Ratemaking. Official source or scholarly record.

  • Intergovernmental Panel on Climate Change (IPCC). 2023. Climate Change 2023: Synthesis Report. Summary for Policymakers. Geneva: IPCC. Official source or scholarly record.

  • Milly, P. C. D., J. Betancourt, M. Falkenmark, R. M. Hirsch, Z. W. Kundzewicz, D. P. Lettenmaier, and R. J. Stouffer. 2008. “Stationarity Is Dead: Whither Water Management?” Science 319 (5863): 573–574. doi:10.1126/science.1151915. Official source or scholarly record.

  • European Insurance and Occupational Pensions Authority (EIOPA). 2021. Report on Non-Life Underwriting and Pricing in Light of Climate Change. Luxembourg: EIOPA. Official source or scholarly record.

  • International Association of Insurance Supervisors (IAIS). 2025. Application Paper on the Supervision of Climate-Related Risks in the Insurance Sector. Basel: IAIS. Issued April 16, 2025. Official source or scholarly record.

  • European Insurance and Occupational Pensions Authority (EIOPA). 2025. “EIOPA Monitoring Exercise Marks Progress in the Integration of Climate Change Considerations into Insurers’ Risk Assessments.” July 23, 2025. Official source or scholarly record.

  • European Insurance and Occupational Pensions Authority (EIOPA). 2025. Dashboard on Insurance Protection Gap for Natural Catastrophes. Updated December 5, 2025. Official source or scholarly record.

  • National Association of Insurance Commissioners (NAIC). 2026. “NAIC Releases First-of-Its-Kind National Analysis of Homeowners Insurance Market Trends.” August 5, 2026. Official source or scholarly record.

  • Gneiting, Tilmann, Fadoua Balabdaoui, and Adrian E. Raftery. 2007. “Probabilistic Forecasts, Calibration and Sharpness.” Journal of the Royal Statistical Society: Series B 69 (2): 243–268. doi:10.1111/j.1467-9868.2007.00587.x. Official source or scholarly record.

  • Diebold, Francis X., and Roberto S. Mariano. 1995. “Comparing Predictive Accuracy.” Journal of Business & Economic Statistics 13 (3): 253–263. doi:10.1080/07350015.1995.10524599. Official source or scholarly record.

  • International Association of Insurance Supervisors (IAIS). 2025. Application Paper on the Supervision of Artificial Intelligence. Basel: IAIS. Released July 2, 2025. Official source or scholarly record.

  • International Accounting Standards Board (IASB). 2017, with subsequent amendments. IFRS 17 Insurance Contracts. Effective for reporting periods beginning January 1, 2023. Official source or scholarly record.

  • Swiss Re Institute. 2026. “Liability Excess Inflation: Still a Key Challenge for Insurers and Their Clients.” January 8, 2026. Official source or scholarly record.

  • Eling, Martin, and Werner Schnell. 2016. “What Do We Know About Cyber Risk and Cyber Risk Insurance?” Journal of Risk Finance 17 (5): 474–491. doi:10.1108/JRF-09-2016-0122. Official source or scholarly record.

  • National Association of Insurance Commissioners (NAIC). 2026. “State Insurance Regulators Issue Nationwide Homeowners Market Data Call.” March 26, 2026. Official source or scholarly record.

  • European Insurance and Occupational Pensions Authority (EIOPA). 2026. Insurance Risk Dashboard. October 2026 edition. Official source or scholarly record.