RESEARCH DOSSIER
Pricing of New Coverages and Sparse Insurance Experience
Abstract. Pricing a new insurance coverage requires an estimate of the future cost of a contract whose own claims record may be absent, short, selected, immature, structurally changing, or dominated by rare losses. The scientific problem is therefore to identify a prospective insured-loss distribution under a precisely defined contract, rather than to infer a stable price mechanically from an unrepresentative historical average. This monograph investigates the definition of novel coverages; the admissibility and adaptation of analogous insurance experience; exposure measurement, claims development, credibility and hierarchical Bayesian estimation; structured expert evidence; catastrophe and systemic accumulation; expenses, reinsurance and capital; and retrospective and prospective validation. Its analysis distinguishes epistemic uncertainty about the risk mechanism, sampling uncertainty from limited observations, process variability inherent in future claims and uncertainty about the contract and insured portfolio. It derives a conjugate Poisson–gamma example and a fully reconciled synthetic premium indication, then subjects both to deliberately specified sensitivity and stress tests. The sources include adopted U.S. actuarial standards, the original credibility literature, peer-reviewed forecasting and emerging-risk research, and European insurance supervisory provisions used within their respective legal scope. The synthesis establishes that sparse-data pricing is a disciplined exercise in transforming, combining and challenging evidence while retaining explicit provenance for every consequential adjustment. The reliability of a new-coverage price depends on coverage identity, exposure comparability, dependence and tail representation, as well as statistical estimation. The work introduces no new actuarial doctrine, original stochastic model or empirically calibrated market-rate claim.
Keywords: New insurance coverage; sparse insurance experience; prospective ratemaking; technical premium; analogue transfer; credibility; Bayesian updating; exposure units; frequency–severity; expert elicitation; catastrophe risk; cyber insurance; tail uncertainty; rate adequacy; reinsurance; model governance; actuarial standards; predictive validation.
Introduction
The actuarial price of an insurance contract is an estimate made before the insured period is fully observed. Every insurance price is prospective in that elementary sense. New coverages intensify the challenge because the relationship between the insured event, the claim trigger, the exposure denominator and the ultimate settlement amount may itself be incompletely established. Observations can be sparse even when economic exposure is large; a new cyber policy may generate no severe claim over a short monitoring interval while retaining material exposure to a shared external dependency. Conversely, a new endorsement attached to an old and well-observed product may have a substantial volume of adjacent experience that is useful after contractual adjustments.
Werner and Modlin (2016) organize property and casualty pricing around the fundamental insurance equation: prospective premium finances expected claims, claim adjustment expense, underwriting expenses and the selected underwriting profit or risk provision. The Casualty Actuarial Society's ratemaking principles, reinstated in 2021 for reference in U.S.-regulated ratemaking, identify the estimation of future risk-transfer costs as the core subject. The adopted ASOP No. 53 applies to actuaries conducting covered future-cost work in U.S. property and casualty practice; its considerations encompass both cost components and the integrated estimate. These instruments anchor the professional discussion while leaving actual rate permissions and filing obligations to applicable law.
This investigation examines the evidential question preceding rate selection. Its principal quantity is the future expected cost, and where useful the distribution, of a defined coverage per unit of defined exposure. It then examines how this quantity can support a rate or premium indication after other costs are incorporated. Comparisons across legal systems separate U.S. actuarial practice guidance from rules governing Solvency II technical provisions. The latter contain rigorous statements about data quality and expert judgment that illuminate transferable analytical concerns, but their valuation requirements cannot simply be transplanted into product rating as binding law.
The method combines primary-document analysis, the original scholarly credibility literature, explicit statistical derivations, a hypothetical numeric construction and comparisons among plausible loss-generating mechanisms. All numerical inputs are invented for exposition. No insurer claims database, regulatory filing, market premium quote or empirical validation sample was used. Source links identify the document supporting the immediately adjacent proposition; references provide the complete bibliographic identity.
New Peril and Coverage Definition
Defining the prospective estimand
A new coverage is a contractual promise to pay or otherwise perform on specified events during specified periods and under specified conditions, where the relevant historical experience is unavailable or insufficiently comparable. Novelty can enter through the insured peril, the trigger, the insured population, the indemnity definition, exclusions, limits, attachment points, geographic territory or interaction among several of these dimensions. A technical price must therefore start with the contract to be priced. An insurer pricing a new cyber interruption endorsement, for example, must identify whether the trigger requires a malicious event, service outage, interruption threshold, waiting period or qualifying business interruption cost. A count of cyber incidents reported by third parties is only a starting description of an event environment. It is not automatically a count of covered claims.
Let the future policy exposure of interest be E, let N(E) be the number of covered claims arising from that exposure, and let Y_j be the contractual payment including the operation of terms and conditions on covered claim j. Before associated expense provisions, the indicated expected loss cost per exposure unit is
where I denotes information available at the valuation date, C the policy contract, P the anticipated insured portfolio and T the future coverage interval. A pricing estimate is incomplete if any of these objects is not specified. Even the meaning of one exposure must be stated: an insured vehicle-year, building-year, payroll amount, shipment, insured location or cyber-covered organization-year can produce different units of risk. The prospective expectation is conditional on current evidence and explicit assumptions. It can be estimated with finite samples; it does not thereby become a known physical constant.
If claim counts and claim payments satisfy the assumptions of the elementary compound model, E[S|E] = E[N|E] E[Y] provides a frequency–severity decomposition. That identity requires careful interpretation where payment severity is linked to claim multiplicity, shared shocks or exposure characteristics. The universally valid form under dependence is E[S|E] = E[E[S|N,E,other factors]|E], with conditional payment distributions and dependencies represented. A policy limit can couple frequency and payment severity because multiple occurrences erode an aggregate limit.
Data scarcity and nonrepresentativeness
Data scarcity has multiple causes and requires distinct responses. Absence of exposure history yields a missing denominator. A very short history yields a small effective sample size. Long-tail claims yield immature ultimate losses. A recently changed contract yields an apparent historical series whose claims are only partly relevant. Changes in technology, climate, litigation, claims reporting, policyholder behavior or third-party infrastructure can produce a structural break. Selection bias arises when insureds entering a new product differ systematically from the source portfolio, even after known risk variables are adjusted. A database can be large but scientifically sparse if the events defining the proposed coverage occur almost never, or if the relevant conditional subgroup is tiny.
The ASOP No. 23 data-quality guidance addresses identifying relevant data, reviewing material characteristics, considering known limitations and communicating their consequences in applicable U.S. actuarial services. Its focus on appropriate data serves as a useful antidote to counting rows as evidence quality. European rules for the calculation of technical provisions explicitly define completeness and appropriateness of data with attention to homogeneous risk groups, risk trends, methods, external-data origin and comparability. Those European legal requirements concern technical-provision valuation; the same methodological questions are scientifically pertinent to pricing under sparse information.
A practical evidence inventory should record, for each data item, the event definition, population, period, data-generation mechanism, geographic and commercial context, exposure measure, policy wording, reporting delay, settlement maturity, currency, censoring and relevant contractual terms. The inventory also states how many observations truly contain information about the specified estimand. Ten years of incident records from a different cause-of-loss definition may provide less direct evidence than one year of meticulously defined incidents from a closely matched cover.
Contract architecture and insurability
Coverage terms transform physical events into financial obligations. A deductible converts many small losses into zero payments. A sublimit truncates payment severity. A waiting period changes attachment for business interruption. Exclusions may remove a peril while leaving the exposure to correlated covered claims; an aggregate policy limit makes payment depend on the order and number of claims. An appropriate pricing analysis applies the intended terms to the underlying event and damage distribution, rather than multiplying old premium by an arbitrary novelty factor.
In catastrophe-exposed lines, the granularity of the loss event matters. A single storm can generate thousands of claims and multiple reinsurance recoveries, while a trigger written at a portfolio aggregate may operate on the joint event rather than each policy independently. The catastrophe-ratemaking standard ASOP No. 39 provides U.S. practice guidance on including catastrophe loss provisions and considering noninsurance data when insurance experience is not representative. It supports explicit event definition, loss attribution and prospective coverage analysis.
New-coverage underwriting also changes the insured population. Better-protected applicants may select higher deductibles; insureds with hidden elevated risk may disproportionately purchase newly offered protection. A rating analysis must distinguish a causal effect of coverage on behavior, the policy's selection mechanism and a correlation observed in the old source data. For a new product, composition and contract terms may change jointly as the business matures. The price is therefore conditional on an anticipated portfolio mix as well as expected risks of individual applicants.
The temporal and institutional foundation of pricing
Three dates should be recorded: the date on which information is frozen, the period during which coverage is written or earned, and the period to which ultimate claim payments and costs are projected. A late-reported claim from an old accident year may be relevant to future policy costs only after development, trend and coverage adjustments. Historical claim counts and paid amounts should be reconciled to incurred and ultimate amounts where applicable; no blanket development factor applies equally to every event type. When discounting cash flows is material and appropriate to the decision, the revised ASOP No. 20 has been effective for covered work products issued on or after June 1, 2026. Its prospective-cash-flow scope complements ASOP No. 53, while the chosen discount convention remains explicit.
The pricing indication, the premium actually offered, the regulatory filing and the amount recognized for accounting or solvency are four different quantities with different purposes. A new product can have a technical cost estimate and still lack an approved or permitted selling rate. Conversely, a quoted rate can be observed in the market without revealing the assumptions underlying that price. Maintaining these distinctions prevents the financial behavior of market participants from masquerading as an empirical estimate of covered expected losses.
Borrowed and Analogous Experience
A disciplined transfer from other portfolios
When direct experience is sparse, evidence may be borrowed from related classes, jurisdictions, predecessor products, engineering exposure databases, industry studies, reinsurer observations or comparable hazards outside insurance. The transfer question is whether an estimate observed under source conditions can be transformed to approximate the target risk distribution. A principled transfer first identifies the source's measured quantity, next identifies relevant differences between source and target, then quantifies each adjustment with separate data or explicitly justified judgment. Each transformation must preserve units and time bases.
For a source portfolio with annual claim frequency f_s and mean gross loss severity m_s, a simple adjusted attritional loss-cost expression is
where the factors respectively change population mix, underlying hazard, payment under the new coverage, future cost level, and c_t is a separately estimated prospective catastrophe or other non-attritional loss provision per target exposure. This factorization is a bookkeeping illustration, not a general probabilistic theorem. Some terms interact: a deductible adjustment can depend on inflation or hazard type, and increased cyber attack frequency may be correlated with severity. Where interactions are material, recompute payments under an explicit joint loss distribution instead of using products of average factors.
For credible source-to-target transport, the target and source must share an observable causal or predictive relationship. A common exposure label is insufficient: two classes with the same annual revenue can have radically different digital dependencies or concentrations. A transfer can fail because claims reported by a source insurer differ from those a new insurer would accept, because smaller incidents are undercounted, or because source losses exclude a nonstandard endorsement. A comparison should state what remains invariant, why it remains invariant, and how that proposition might be falsified by new data.
Exposure denominators, valuation and development
An exposure denominator is the bridge between source loss and future insured population. It must be aligned to coverage duration, insured units and claim occurrence. Consider a source book with 400 policies held for an average of six months. Its earned exposure is 200 annual policy-years, not 400. If three claims emerge, the observed annualized frequency is 3/200 rather than 3/400. Switching denominators without adjustment changes the indicated frequency by a factor of two even before any modeling. Exposure inflation, partial policy cancellation and midterm endorsements can further alter the denominator.
Severity data must be examined on a comparable value basis. Reported ground-up losses may include uninsured costs and policy exclusions; case-incurred losses may include subjective outstanding estimates; paid losses may omit slow-developing costs. For long-tail liability coverages, the time between occurrence, reporting and final settlement creates a material gap between observed and ultimate loss. In borrowed experience, the source's reporting and reserving practice can differ from that anticipated under the target contract. An appropriate transfer states the valuation basis and the procedure used to project ultimate payments. Development factors borrowed from another class are themselves uncertain and may be less transferable than the original accident event frequency.
Observed monetary losses should be expressed in a consistent currency and price level. Index selection depends on the insured loss mechanism: consumer prices, wages, medical care, construction costs, repair labor, legal settlements and cyber restoration expenses need not develop at the same pace. Currency conversion at a spot date may be insufficient when claim payments span years. Inflation assumptions can also alter deductible exhaustion and policy-limit utilization nonlinearly.
The educational source Basic Ratemaking treats exposure, premium, loss development, trend and risk classifications as distinct construction tasks. Its relevance for new coverages lies in the order of operations: a standardized rate calculation does not repair a poorly defined contract or a mismatched denominator. The actuarial problem is first to make the underlying exposure and loss measures comparable.
Contract and risk-condition adjustments
A useful analogue ledger separates four dimensions. First is event translation: does the observed peril correspond to the new insured trigger? Second is benefit translation: how would the target deductible, attachment point, waiting period, coinsurance or aggregate limit change payment? Third is portfolio translation: what adjustments reflect geography, industry, engineering controls or applicant selection? Fourth is time translation: how would hazard frequency, severity and reporting behavior differ in the prospective period? Each entry documents the source measure, factor or simulation, confidence basis and sensitivity range.
For a severity random variable X and an occurrence limit U above deductible d, covered payment is min(max(X-d,0),U), assuming one independent per-occurrence limit and no additional clauses. Its expectation can be written as the integral of the survival function over the covered layer:
This identity makes the tail requirement transparent. If the source reports only the mean claim size, the target layer expectation generally cannot be computed. One needs the relevant survival function, a defensible range of tail shapes, or additional data. Two severity distributions can share identical means while producing substantially different costs for high attachment layers. Borrowed mean severities should not be treated as adequate information for entirely new limit and deductible structures.
When external incident evidence is a collection of losses above a reporting threshold, adjustment for left truncation must be explicit. Suppose a database observes only losses exceeding $100,000 and no independent information about events below that threshold. It can support estimation of a conditional tail distribution under additional assumptions, but it cannot alone identify the unconditional annual claim frequency. Treating the observation rate as the total claim frequency creates systematic understatement of low-value events; scaling the published tail to all claims without a detection model may create an equally serious error.
Statistical transport and the risk of borrowed certainty
Borrowing observations may reduce sampling variance while adding transport bias. Let θ_t be the target expected cost and θ_s a source estimate. The use of θ_s as an estimator of θ_t introduces error from the source's measurement uncertainty and from δ=θ_t−θ_s, the unknown transfer difference. An internally narrow confidence interval around θ_s says little about δ unless comparability is independently justified. For a fixed but incorrectly transferred source expectation, adding more source observations cannot eliminate the target bias. Thus a million observations from a different causal environment can yield a highly precise estimate of the wrong insurance risk.
This distinction also governs calibration. Standardizing source experience by risk factors can only adjust differences that are observed, measured comparably and adequately represented. Unobserved vulnerabilities, hidden policy exclusions or differences in claim reporting remain residual transport risks. An analysis should compare at least three alternatives: a direct modified analogue, a pooled statistical estimate and an expert- or scenario-based benchmark. Material disagreement among them is substantive information about uncertainty, not automatically an invitation to average the figures into a deceptively precise result.
Credibility under Sparse Data
Classical credibility and related experience
Credibility expresses how much predictive weight should be attached to a limited subject experience relative to an external or collective benchmark. Hans Bühlmann's experience-rating and credibility analysis connected risk heterogeneity, estimation and insurance experience rating. Jewell's later exact Bayesian credibility result established formal relationships between linear credibility and specified exponential-family likelihoods with conjugate priors. These historical results explain why blending experience with a collective estimate has a statistical foundation under explicit distributional assumptions.
The adopted ASOP No. 25 discusses the selection and application of credibility procedures, including blending subject and relevant experience and assessing predictive value. A familiar expression is θ_hat = Z θ_subject + (1−Z) θ_complement, with Z between zero and one. The complement should be relevant to the same quantity and prospective conditions. The numerical value Z is a property of the selected method, variance structure and purpose; it is not a universal function of the number of claims, and a general statement such as “ten claims have forty percent credibility” is unjustified without a model.
A critical distinction concerns the object receiving credibility. Applying credibility to claim frequency need not justify identical credibility for conditional severity, large-loss frequency, reporting delays or the cost of reinsurance. A short observation period can contain meaningful frequency information for frequent small claims while providing almost no information about a million-dollar layer. Blending an overall loss ratio may conceal a change in policy limits or revenue exposures. It is often better to decompose the estimate into components with different evidence bases and then assess their correlations when recombining them.
A Poisson–gamma derivation
Consider covered claim counts N observed over e standardized exposure-years. Assume N conditional on a constant frequency λ follows Poisson(eλ), and model the unknown λ with a gamma prior of shape α and rate β. The prior mean is α/β, and β has units of exposure. The likelihood is proportional to λ^N exp(−eλ). Multiplying it by the gamma density yields a posterior gamma distribution with shape α+N and rate β+e. Its mean is
This is exactly a credibility-weighted blend of the observed frequency N/e and the prior mean α/β, with Z=e/(β+e). The derivation supplies an explicit interpretation of prior exposure weight; it does not make any substantive claim that a chosen prior accurately represents a particular emerging peril. In a genuine application, α and β must come from independently appraised comparable evidence, with documented uncertainty about transport from the source to the target product. A conjugate prior is a mathematical convenience, while the prior's empirical provenance is an actuarial obligation.
For a future exposure e*, the predictive count distribution has mean e*(α+N)/(β+e). Its variance follows the law of total variance:
The second term is epistemic parameter uncertainty under the model. A plug-in Poisson forecast suppresses this term and therefore understates predictive dispersion whenever posterior frequency uncertainty is material. The derivation is exact for the stated model but cannot account for exposure correlation, unobserved structural changes or model-class error without further assumptions.
A reproducible frequency illustration
Take a purely hypothetical analogue with a prior annual covered frequency of 0.040 claims per exposure-year. Express the prior as α=4 and β=100 equivalent exposure-years. Suppose a new portfolio produces three covered claims across 50 earned exposure-years, so its observed frequency is 3/50=0.060. The credibility factor is Z=50/(100+50)=1/3. The posterior frequency is (4+3)/(100+50)=0.0466667, which also equals (1/3)(0.060)+(2/3)(0.040). Under the stated Poisson–gamma model, the new data increase the estimated frequency by 16.67% relative to the prior benchmark.
Now suppose there had been no claims in the same 50 exposure-years. The posterior mean would be 4/150=0.0266667. Yet the probability of observing zero claims over 50 exposure-years at the original frequency 0.040 is exp(−2)=0.135335. A zero count is therefore not proof of zero prospective risk. At a rare peril frequency of 0.001 per year, even 1,000 exposure-years would yield a zero count with probability about exp(−1)=0.367879. These calculations show why the length of exposure history must be interpreted relative to the event rate of interest.
For a prospective 100 exposure-years after the three-claim observation, the predictive claim count mean is 100×7/150=4.6667. Its variance is 4.6667+100²×7/150²=7.7778, whereas a frequency plug-in Poisson forecast gives variance only 4.6667. The extra 3.1111 reflects uncertainty in λ within the assumed family. A calibration exercise can test whether this family and its transported prior correspond to observed development; the calculation alone does not establish empirical validity.
Hierarchical pooling, selection and model diagnostics
New products often comprise multiple sectors, regions, limits and coverage forms, many with zero or one claim. Estimating every cell independently produces unstable rates. Complete pooling imposes a single risk frequency and can erase consequential heterogeneity. Partial pooling allows cell-level parameters to vary around a shared distribution while borrowing information. With a log-rate specification, a practical hierarchical model might use log λ_g = x_g'β + u_g, with random effect u_g drawn from a shared distribution whose variance is estimated from the data or assigned a defensible prior. This description uses standard hierarchical modeling; it is not an original new model proposed by this monograph.
A partial-pooling estimate should answer three questions: which risk attributes genuinely predict covered claims; how many independent units support the between-group variance estimate; and how far the new group lies from the training populations. Sparse groups may shrink heavily toward a pooled mean. If the pooled data are generated by obsolete contract terms or selection processes, shrinkage can import collective bias. Model validation must therefore examine both overall prediction and credible subgroup behavior. The CAS GLM rating monograph presents practical construction of insurance rating plans, while ASOP No. 56 on modeling addresses material assumptions, validation and model risk for applicable actuarial services.
The smallest observable unit is not always an independent unit. If one technology provider's outage affects 400 insured organizations, the insurer does not have 400 independent catastrophe events. Similarly, repeated claims from the same policyholder, property or connected enterprise create shared latent exposure. A high record count can conceal low effective information because cluster correlation reduces independent replication. Backtests should leave entire time intervals, event clusters, insureds or jurisdictions out of training when appropriate, rather than splitting correlated claims randomly across training and validation sets. Random splits can give an implausibly favorable estimate of predictive performance in emerging risks.
Severity, dependence and composite uncertainty
Claim payments are often right-skewed, can be left-truncated or right-censored, and may be bounded by contracts. Separate estimation of a mean frequency and a mean severity is useful only after dependence and coverage terms have been considered. For a high-deductible policy, the tail above the deductible—not the mean loss across all incidents—is the relevant object. A defensible severity analysis therefore compares compatible parametric families, empirical layers where available, engineering limits and sensitivity to heavy tails. If the sample provides no credible information about the tail, the conclusion is that tail uncertainty remains material; assigning an arbitrary single-tail parameter creates artificial precision.
A fully specified prospective loss distribution combines event frequency, conditional payment severity, portfolio-level dependence and policy aggregate structure. Sampling variation, parameter uncertainty and structural uncertainty affect this distribution differently. Sampling uncertainty can diminish as relevant exposure accumulates. Structural uncertainty about an unprecedented hazard may remain until the underlying environment stabilizes, and that environment may itself continue evolving. Distinct provisions for expected loss, capital or required return, and management stress analysis should be reported with enough clarity to prevent a stochastic risk margin from being counted twice.
Expert and External Evidence
Sources of external evidence
External evidence may include engineering reliability studies, hazard maps, publicly reported security incidents, laboratory tests, claims from analogous lines, reinsurer loss events, national statistics, judicial and administrative records, and sector-specific operating data. Each source measures a different stage in the chain from hazardous event to covered claim. An operational incident database measures events according to its own reporting definition. A settlement database measures claims whose legal, insurance and commercial filters have already operated. An engineering model may measure vulnerability conditional on specified loading. Combining these sources requires a transparent transformation between their estimands.
The literature on emerging cyber risks provides an instructive case. Eling and Schnell (2016), in their review of cyber insurance research, identify limited actuarial loss data, changing threat characteristics and accumulation as central obstacles. The conclusion is directly relevant to new coverages: underrepresentation of future cyber incidents in historical claims is a threat to validity even when insurers have rich operational data. More years of old technology exposures may not reproduce a new software concentration or an adversary's altered incentives.
The basic provenance questions for external evidence are who observed the event; what was observable; which incidents were not recorded; how event and financial impact were defined; and which transformations are needed to obtain a covered loss distribution. Licensing and confidentiality can also limit reproduction and auditability. A publisher-provided summary or bibliographic abstract may establish a work's research scope but is not evidence for detailed results absent from that summary. The present monograph accordingly confines attributed claims to source statements inspected in original institutional, publisher or institutional-repository records and independently derives its numerical examples.
Structured expert elicitation
Expert judgment is often unavoidable when claim triggers are novel or tail observations do not exist. A structured process asks qualified experts to define quantities in advance, provide distributional assessments and identify causal mechanisms and uncertainty. Experts can assess annual event probabilities, conditional insured loss, severity quantiles, reporting changes, recovery durations and correlations. The assessment should record definitions, units, conditioning information, dates, experts' roles and the rationale for disagreement. Experts should not be encouraged to conceal uncertainty by reporting only a single central estimate.
Cooke's study of expert uncertainty examines the representation, assessment and combination of subjective probabilities. Its methodological importance is to treat elicitation as a reasoned and documented inference process, with attention to experts' actual knowledge and the quality of probabilistic judgments. Expert opinions may be compared, calibrated where genuine seed questions exist, or represented as a range of competing distributions. The choice of elicitation design must reflect the situation; an apparently numerical aggregation is not automatically more objective than documented disagreement.
Under the European expert-judgment rule in Article 2, assumptions relating to listed Solvency II valuation, capital and investment matters must be informed by personnel with relevant expertise, and internal users must be informed about reliability and limitations. That legal requirement concerns specified Solvency II areas. Its evidentiary lesson for international new-product pricing is that assumptions require knowledgeable ownership, intelligible communication and explicit limitation statements. External experts may be especially valuable for newly emerging hazards, but the insurer remains responsible for the assumptions it adopts within the governing regime.
An integrated synthetic technical rate
Consider an entirely hypothetical new annual coverage written on one standardized exposure unit. Assume an independently selected analogous annual covered event frequency of 0.040 and analogue mean ground-up claim severity of $12,000. An analysis of hypothetical deductible, exclusion and policy-limit terms produces a covered-loss factor of 0.70; a forward cost-level factor is 1.10; a hypothetical underwriting-control factor is 0.90. These factors are fixed illustrative assumptions rather than estimated causal effects. The resulting attritional loss cost per exposure-year is 0.040×12,000×0.70×1.10×0.90=$332.64.
Assume separately modeled expected catastrophe loss of $45.00 per exposure, giving an expected loss component of $377.64. Suppose prospective claims adjustment expense is 12% of expected loss, giving $45.3168. Let fixed acquisition and administration expense be $35.00 per exposure. Suppose all remaining premium-proportional expenses and the selected target underwriting profit and risk provision together consume 20% of written premium, with these categories expressly defined to avoid double counting. Solving the illustrative insurance equation gives
This is a synthetic technical indication per annual standardized exposure, not a quoted commercial rate, regulator-approved rate, empirical estimate or claim about the cost of a particular named insurance product. The assumed factor 0.70 is meaningful only if contract payment terms, source claim distributions and the factor's estimation basis can be justified; the arithmetic cannot supply that justification. A different portfolio mix, catastrophe concentration, terms factor, reinsurance cost allocation or capital charge could change the result substantially. If the target 20% provision includes an amount already placed in fixed expense, the expression would overcharge that component. A reproducible pricing workbook must preserve the exact accounting definitions.
An alternative construction uses the previously derived posterior frequency 0.0466667 in place of 0.040, while keeping all other illustrative assumptions unchanged. Attritional loss cost becomes 0.0466667×12,000×0.70×1.10×0.90=$388.08. Expected losses become $433.08, claim adjustment expense $51.9696, and the corresponding premium becomes ($433.08+$51.9696+$35)/0.80=$650.062. This shows that an apparently modest frequency change can produce a material change in the technical indication. It also demonstrates why the credibility model should be integrated with the same contract and cost assumptions used by the primary rating calculation.
Regulatory, underwriting and business boundaries
Technical rate estimation is one input to a lawful pricing decision. Other constraints include rate regulation, risk classification, distribution agreements, coverage availability, market response, portfolio strategy and insurer solvency. The adopted U.S. ASOP No. 12 on risk classification provides professional considerations for risk classification in covered work; it does not itself grant permission to use every potentially predictive variable in every jurisdiction. Rate approval or filing requirements must be established from the governing jurisdiction and product class rather than inferred from statistical accuracy.
A new coverage may be better managed initially through narrower limits, explicit exclusions, aggregate caps, geographic controls, risk engineering or reinsurance when the tail cannot be priced credibly. Those contract and portfolio choices alter the insured risk and therefore trigger a new prospective analysis. A high premium is not a substitute for understanding an uncontrolled aggregation pathway. In an emerging peril, underwriting can be a form of information production: clear policyholder reporting, claims coding and consistent contract wording improve subsequent actuarial learning.
Scenario and Catastrophe Inputs
Tail events and accumulation
The absence of observed catastrophes is consistent with material catastrophic expected cost. Suppose one annual exposure unit carries a purely illustrative probability 0.001 of a $1,000,000 covered event, with no smaller losses. Its expected insured cost is $1,000 per exposure-year. Across 100 independent exposure-years, the probability of observing zero events is (0.999)^100≈0.9048. Thus 100 exposure-years with no recorded claim can coexist with a substantial prospective cost under the assumed model. This example describes sampling alone; real catastrophe losses can be highly correlated, making a portfolio of 100 insured exposures informationally weaker than 100 independent years of observation.
Event dependence is central for natural catastrophe, shared-service cyber interruption, product recall, terrorism and some liability accumulation. For portfolio losses S=Σ_i X_i, the variance is Σ_i Var(X_i)+2Σ_{i<j}Cov(X_i,X_j). Assuming every covariance equals zero can markedly understate portfolio risk even when each individual policy's expected loss is reasonable. Expected portfolio loss remains the sum of individual expected losses under finite expectations, but the distribution of aggregate claims, required capital, reinsurance attachment probability and liquidity need depends on the joint law. A catastrophe model may require explicit hazard, vulnerability, financial and aggregation modules. The adopted ASOP No. 38 catastrophe-model guidance addresses selection, use and review of such models in applicable actuarial practice.
For a new contract, an event set must be mapped to policy clauses. A storm hazard footprint may be well modeled, yet a new business-interruption endorsement might depend on supplier outages, waiting periods and contingent loss multipliers outside the ordinary property model. A cyber aggregation may arise from a common authentication provider or software library; the exposure measure should include such concentrations where available. Mapping an annualized catastrophe model output to a new policy requires assumptions about hours clauses, limits, multiple occurrences and reinsurance recoveries. Model outputs cannot be presumed to equal the cost of newly specified benefits.
Physical, economic and behavioral scenarios
Scenario analysis changes the mechanisms that govern event frequency, severity, dependence or insured behavior, rather than merely adding an arbitrary percentage to total premium. Useful scenarios include a stable baseline, adverse technology concentration, higher restoration cost, widening coverage interpretation, litigation and claim inflation, weakening cyber defenses, an altered hazard climate, supplier concentration or a higher reporting rate. Every scenario should state which variables change and why, and should distinguish a forecast probability from a deliberately severe stress event.
Using the synthetic premium illustration, suppose future event frequency rises 30%, covered average severity rises 25%, and the separately modeled catastrophe expected loss doubles from $45 to $90, while the 12% claim-adjustment ratio, $35 fixed expense and 20% premium-relative provision remain unchanged. Attritional cost rises from $332.64 to $332.64×1.30×1.25=$540.54. Total expected loss becomes $630.54. The revised technical premium is ($630.54×1.12+$35)/0.80=$926.506, approximately $926.51. The increase from $572.45 is about 61.85%. This calculation is deterministic conditional on the stated stressed parameters; the scenario does not assert the probability of simultaneous changes.
A central estimate may be accompanied by quantified sensitivity to plausible parameter ranges. The analyst should also ask whether the ranges encompass genuinely distinct structural models. A thin-tailed severity distribution, a heavy-tailed distribution and a mixture driven by systemic outage can produce different high-layer premiums even when their fitted attritional loss ratios are similar. Scenario tests should therefore include alternative event structures, not only parameter perturbations within one chosen model.
Stress tests, ambiguity and model competition
Parameter uncertainty describes uncertainty about values inside a specified model. Model uncertainty concerns which mechanisms or distributional family applies. Structural uncertainty includes the possibility of unmodeled pathways such as coordinated attacks, correlated legal interpretations or a new dependency across insureds. These uncertainties can be described by sensitivity grids, posterior predictive distributions, multiple model families and explicit stress outcomes. They must not be compressed into a single confidence interval presented as if it covered every possible source of error.
Compare three assessments of one hypothetical layer: a frequency–severity analogue based on incident records, an event-simulation estimate from engineering risk models and a structured expert forecast. Their disagreement should be decomposed into different event definitions, contract treatment, insured population, horizon and tail structure. If all differences can be traced to those elements, the models become more comparable. If material unexplained differences remain, decision makers should receive the range and the unresolved dependency rather than a false consensus. The appropriate price may then be accompanied by exposure caps or contract limitations, or a decision to defer expansion of the coverage until evidence improves.
A risk margin or required return may compensate for bearing uncertainty and capital consumption, subject to local economic and regulatory conventions. That economic decision must be distinguished from the expected covered claim cost. Adding a safety percentage to expected loss does not make the estimate of expected loss itself more accurate. Similarly, the cost of reinsurance includes terms, limits, commissions and market pricing; it cannot automatically be appended as an additive charge without accounting for recoveries and the risk already retained by the insurer.
Validation and the first years of experience
The most useful post-launch data program begins before the first policy is sold. Preserve the wording version, underwriting characteristics, declared exposure, policy modifications, risk-control verification, reported events, claim payments, outstanding case estimates, closure status and loss-adjustment costs. Define reporting lags and identify which exposures have reached sufficient maturity. A stable event taxonomy allows subsequent observations to revise explicit model components rather than leaving analysts to reconstruct why early-year data appear different.
Validation evaluates the prospective distribution, not merely whether the total premium happened to exceed paid claims in a year. Applicable tools include exposure-weighted residuals, claim-count calibration, large-loss event frequency checks, likelihood-based or other proper scoring rules, held-out period validation, and predictive interval coverage. Gneiting and Raftery (2007), in their proper scoring rules study, explain why appropriate scores incentivize honest distributional probability forecasts. Small validation sets may have low power, especially for tails; an apparently successful absence-of-loss year should be reported with its predictive probability under the pre-launch model.
The validation plan must specify what counts as a failure. Examples include a systematic mismatch between predicted and observed frequency after exposure normalization, a run of large claims inconsistent with the documented tail model, unexpected dependency among insureds, or substantial changes in legal or operational claim definitions. Recalibration may require revising the exposure base, contract interpretation or source analogue, not merely adjusting a coefficient. New observations should update the full provenance ledger, including what was learned and what remains unknown.
Assumption Disclosures
A documented decision chain
A scientifically defensible new-coverage indication requires a reviewable connection from contract to evidence to assumptions to expected loss to premium. The first record is contractual: coverage scope, temporal basis, trigger, exclusions, financial terms, limits and assumed claims handling. The second is empirical: internal and external data origin, completeness, claims maturity, denominator, observation selection and adjustments. The third is inferential: chosen statistical model, external-data transport, credibility weight, expert assessment, catastrophe dependence and estimation uncertainty. The fourth is financial: expected covered losses, claim adjustment expense, underwriting expenses, reinsurance economics, capital or return provisions, taxes where applicable and premium convention. The fifth is governance: decision owner, validation arrangements, revision triggers and relevant legal constraints.
The ASOP No. 41 communications standard concerns communication of actuarial findings in applicable U.S. practice. Combined with the disclosure expectations of ASOP No. 53, it reinforces the need to identify intended use, material assumptions, limitations and uncertainty. Source-claim traceability should be granular: an evidence record showing only that an institution issued an actuarial standard is not sufficient to support an unrelated factual estimate of an emerging peril's annual cost.
A useful assumption register identifies which assumptions are based on direct insured claims, outside insurance experience, engineering measurements, statistical extrapolation, expert elicitation or business judgment. It records the prospective period, relevant population and limitations for each. A numerical scenario can be marked as hypothetical rather than observed. Distinct cost margins should be reconciled so that a catastrophe expectation appearing in both attritional loss data and explicit catastrophe loading is not counted twice. Rate indications should be versioned as the loss environment or policy wording changes.
International supervisory interpretation
U.S. ASOPs guide covered actuarial practice, while insurance law and permissible rating variables depend on jurisdiction. Solvency II provisions govern valuation and solvency subjects in relevant European legal settings. Article 19 of Delegated Regulation (EU) 2015/35 establishes data quality criteria for provisions, including conditions relevant to external data. Article 20 requires documentation of data limitations for those provisions. Article 21 permits appropriate approximations for best estimates when specified data-insufficiency conditions are all met. This is a conditional permission within technical-provision valuation, not a blanket regulatory license to set unsupported new insurance rates.
The European Insurance and Occupational Pensions Authority's revised valuation guidelines, applicable from January 2023, address matters including expert judgment and best-estimate valuation. Article 265 also prescribes technical-provisions documentation, covering data, assumptions, statistical methods and validation. Those instruments illustrate how provenance and proportionality are institutionalized in a specific supervisory framework. Their conceptual relevance to sparse-data pricing is substantial, while the legal purpose remains valuation of insurance and reinsurance obligations rather than commercial product ratemaking.
An internationally useful monograph must preserve both connections and legal boundaries. An insurer considering a new product in several countries should first establish the applicable pricing, underwriting, anti-discrimination, capital, conduct and reporting rules in each market. It should then apply actuarial inference to its defined estimand and document local departures. Transfer of an analytical technique across countries is legitimate when its assumptions hold; transfer of legal authority without jurisdictional analysis is not.
Fairness, communications and governance
Sparse claims experience can amplify classification instability. A newly observed claim may cause a very large indicated relativity in a tiny class, while an adjacent class with no claims may appear implausibly cheap. Partial pooling may moderate sampling noise, but proxy variables may still encode protected characteristics or commercial disadvantage, subject to jurisdiction-specific rules. A documented classification review should examine actuarial relevance, permitted use, data collection bias, differential impacts, transparency and the stability of proposed rate factors under small changes to assumptions. Scientific predictive ability alone does not settle fairness or legal acceptability.
Communication to management should preserve the distinction among a central expected-cost estimate, the uncertainty around it, the cost of the assumed contract, the impact of aggregation and a commercial premium decision. The result should show how it changes under material sensitivities, including dependence assumptions. It should also identify information that would most reduce decision uncertainty. For an emerging hazard, improved event taxonomy or exposure data may be more valuable than a more complicated regression; for a high-layer product, credible severity tail information may be more valuable than thousands of low-value incidents.
A model owner should monitor changes in the covered peril, contract wording, portfolio mix and claim settlement environment. Escalation triggers should include emerging concentration, policyholder selection, unusually rapid claim development, new litigation doctrines or insurer/reinsurer disputes about interpretation. The governance objective is to retain the ability to explain each indication and revise it with evidence, rather than merely preserve a software model that produces a price.
Integrated analytical synthesis
The source comparison and calculations support five linked conclusions. First, contract identity precedes statistical estimation: without a stable covered-event and payment definition there is no unambiguous loss-cost estimand. Second, borrowed evidence carries two independent questions, the precision of the original estimate and the validity of transferring it. Third, credibility can rationally allocate weight under specified assumptions but cannot correct a structurally inappropriate source analogue by increasing its statistical weight. Fourth, the predictive distribution must retain parameter, tail and aggregation uncertainty where they influence future payments, reinsurance and capital. Fifth, a credible technical price requires governance, legal interpretation and continuing predictive review after inception.
These conclusions are an original analytical synthesis of established actuarial, statistical and supervisory materials. They are not presented as newly discovered statistical identities, a newly named pricing framework, a proprietary empirical model or a general solution to insurability. Their practical value lies in the auditable chain of reasoning: each numerical result states its data-generating assumptions, each external source is linked to the supported proposition, and each jurisdictional rule is applied to the purpose for which it was issued.
Conclusion
Pricing of new coverages under sparse insurance experience is the prospective estimation of a specifically defined contractual risk using incomplete but assessable evidence. The strongest analysis identifies exactly what the insurer promises, how its exposure denominator is measured, which observations are transferable and which uncertainty components remain. Credibility blends information under declared statistical conditions. Analogous data need explicit contractual, temporal and portfolio translation. Expert judgment supplies structured evidence where experience is absent, with reliability and limitations made visible. Catastrophe and systemic exposures require event-level dependence and tail-sensitive analysis. The premium indication finally reconciles expected claims, expenses, reinsurance and the chosen financial provisions without conflating expected loss with risk compensation.
The mathematical examples demonstrate both the usefulness and the limits of standard actuarial tools. A Poisson–gamma estimate changes rationally as claims emerge and predicts more variability than a plug-in Poisson forecast. A synthetic frequency–severity price is fully recalculable from assumptions, and a severe scenario reveals meaningful price sensitivity to simultaneous frequency, severity and catastrophe changes. None of those calculations supplies missing empirical support for its assumptions. They instead expose the precise information required for a defensible decision. The principal scientific result is a source-grounded interpretation of new-coverage pricing as a connected contractual, actuarial, inferential and governance problem whose validity must be established at each link and maintained as the insured experience develops.
Methodological Appendix
Definitions and scope of the worked examples
Every monetary example uses U.S. dollars for arithmetic convenience, a hypothetical standardized annual policy exposure and an invented loss-generating process. The first frequency example assumes independent Poisson events conditional on a time-invariant annual rate, with a gamma shape–rate prior. The formulas use an exposure denominator e, observed covered count N, future exposure e*, shape α and rate β. A credible collective prior requires real relevant evidence in application. The second premium illustration treats loss components as undiscounted prospective amounts per exposure and holds expense conventions fixed. It is not an IFRS 17 fulfillment-cash-flow estimate, a Solvency II best estimate, a statutory rate filing or a capital model.
Independent arithmetic verification targets
| Quantity | Formula | Expected value |
|---|---|---|
| Prior frequency | 4 / 100 | 0.0400000 |
| Observed frequency | 3 / 50 | 0.0600000 |
| Credibility Z | 50 / 150 | 0.3333333 |
| Posterior frequency | 7 / 150 | 0.0466667 |
| Posterior frequency with zero claims | 4 / 150 | 0.0266667 |
| Zero-claim likelihood at expected count 2 | exp(-2) | 0.1353353 |
| Future count expectation | 100 × 7 / 150 | 4.6666667 |
| Future predictive count variance | 100 × 7 / 150 + 10000 × 7 / 22500 | 7.7777778 |
| Base attritional loss | .04 × 12000 × .70 × 1.10 × .90 | $332.64 |
| Base total expected loss | 332.64 + 45 | $377.64 |
| Base claim adjustment expense | .12 × 377.64 | $45.3168 |
| Base premium | (377.64 × 1.12 + 35) / .80 | $572.4460 |
| Updated attritional loss | (7 / 150) × 12000 × .70 × 1.10 × .90 | $388.08 |
| Updated premium | ((388.08 + 45) × 1.12 + 35) / .80 | $650.0620 |
| Stress attritional loss | 332.64 × 1.30 × 1.25 | $540.54 |
| Stress total loss | 540.54 + 90 | $630.54 |
| Stress premium | (630.54 × 1.12 + 35) / .80 | $926.5060 |
Evidentiary limitations and reproducibility
The examples are analytical demonstrations and use no observed insurance sample. Source research verifies the documentary identities and propositions recorded in EVIDENCE.json; it does not assert that paywalled books and articles were read in their entirety. A valid production application would require independently validated covered-claim and exposure data, actuarial judgment signed by qualified professionals, applicable legal examination, sensitivity assessment and monitoring. The referenced standards apply only within their respective scopes and effective dates. The manuscript is an international scholarly research work, not an insurer-specific premium recommendation.
References
Actuarial Standards Board. (2000; revised May 2011). Actuarial Standard of Practice No. 39: Treatment of Catastrophe Losses in Property/Casualty Insurance Ratemaking. https://www.actuarialstandardsboard.org/asops/treatment-catastrophe-losses-propertycasualty-insurance-ratemaking/
Actuarial Standards Board. (2005; revised May 2011). Actuarial Standard of Practice No. 12: Risk Classification (for All Practice Areas). https://www.actuarialstandardsboard.org/asops/risk-classification-practice-areas/
Actuarial Standards Board. (2010). Actuarial Standard of Practice No. 41: Actuarial Communications. https://www.actuarialstandardsboard.org/asops/actuarial-communications/
Actuarial Standards Board. (2013). Actuarial Standard of Practice No. 25: Credibility Procedures. https://www.actuarialstandardsboard.org/asops/credibility-procedures/
Actuarial Standards Board. (2016). Actuarial Standard of Practice No. 23: Data Quality. https://www.actuarialstandardsboard.org/asops/data-quality/
Actuarial Standards Board. (2017). Actuarial Standard of Practice No. 53: Estimating Future Costs for Prospective Property/Casualty Risk Transfer and Risk Retention. https://www.actuarialstandardsboard.org/asops/estimating-future-costs-prospective-propertycasualty-risk-transfer-risk-retention/
Actuarial Standards Board. (2019). Actuarial Standard of Practice No. 56: Modeling. https://www.actuarialstandardsboard.org/asops/modeling-3/
Actuarial Standards Board. (2021). Actuarial Standard of Practice No. 38: Catastrophe Modeling (for All Practice Areas). https://www.actuarialstandardsboard.org/asops/actuarial-standard-of-practice-no-38-revised-edition/
Actuarial Standards Board. (2025). Actuarial Standard of Practice No. 20: Analysis of Property/Casualty Cash Flows, Including Discounting. Effective June 1, 2026. https://www.actuarialstandardsboard.org/asops/asop-no-20-analysis-of-property-casualty-cash-flows-including-discounting/
Bühlmann, H. (1967). Experience rating and credibility. ASTIN Bulletin, 4(3), 199–207. https://doi.org/10.1017/S0515036100008989
Casualty Actuarial Society. (1988; reinstated 2021). Statement of Principles Regarding Property and Casualty Insurance Ratemaking. https://www.casact.org/statement-principles-regarding-property-and-casualty-insurance-ratemaking
Cooke, R. M. (1991). Experts in Uncertainty: Opinion and Subjective Probability in Science. Oxford University Press. https://doi.org/10.1093/oso/9780195064650.001.0001
Eling, M., & Schnell, W. (2016). What do we know about cyber risk and cyber insurance? Journal of Risk Finance, 17(5), 474–491. https://alexandria.unisg.ch/entities/publication/17442364-7c66-49c2-bce1-c95292b5305f
European Commission. (2015). Commission Delegated Regulation (EU) 2015/35, Article 2: Expert Judgement. EIOPA Solvency II Single Rulebook. https://www.eiopa.europa.eu/rulebook/solvency-ii-single-rulebook/article-2415_en
European Commission. (2015). Commission Delegated Regulation (EU) 2015/35, Article 19: Data Used in the Calculation of Technical Provisions. EIOPA Solvency II Single Rulebook. https://www.eiopa.europa.eu/rulebook/solvency-ii-single-rulebook/article-2432_en
European Commission. (2015). Commission Delegated Regulation (EU) 2015/35, Article 20: Limitations of Data. EIOPA Solvency II Single Rulebook. https://www.eiopa.europa.eu/rulebook/solvency-ii-single-rulebook/article-2433_en
European Commission. (2015). Commission Delegated Regulation (EU) 2015/35, Article 21: Appropriate Use of Approximations to Calculate the Best Estimate. EIOPA Solvency II Single Rulebook. https://www.eiopa.europa.eu/rulebook/solvency-ii-single-rulebook/article-2434_en
European Commission. (2015). Commission Delegated Regulation (EU) 2015/35, Article 265: Valuation of Technical Provisions—Documentation. EIOPA Solvency II Single Rulebook. https://www.eiopa.europa.eu/rulebook/solvency-ii-single-rulebook/article-5928_en
European Insurance and Occupational Pensions Authority. (2022). Revised Guidelines on Valuation of Technical Provisions. Applicable January 1, 2023. https://www.eiopa.europa.eu/publications/revised-guidelines-valuation-technical-provisions_en
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
Goldburd, M., Khare, A., Tevet, D., & Guller, D. (2025 revision). Generalized Linear Models for Insurance Rating (2nd ed.; CAS Monograph No. 5). Casualty Actuarial Society. https://www.casact.org/monograph/cas-monograph-no-5
Jewell, W. S. (1974). Credible means are exact Bayesian for exponential families. ASTIN Bulletin, 8(1), 77–90. https://www.cambridge.org/core/journals/astin-bulletin-journal-of-the-iaa/article/credible-means-are-exact-bayesian-for-exponential-families/692C8AF44B832D2DE1C20713B5BF90DE
Werner, G., & Modlin, C. (2016). Basic Ratemaking (5th ed.). Casualty Actuarial Society. https://www.casact.org/sites/default/files/2021-03/5_Werner_Modlin.pdf