Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Prospective Loss Cost Estimation and Pure Premium Ratemaking

Abstract. Prospective loss cost estimation is the quantitative conversion of observed and otherwise relevant risk experience into an expected future claim cost per specified unit of insurance exposure. In property and casualty insurance this task is foundational to pure premium ratemaking, because the premium associated with a future risk-transfer period must be established while a substantial portion of the ultimate insured losses remains unknown. The scientific object of this monograph is the prospective unit cost of insured claims, with claim adjustment expenses identified explicitly according to the selected cost convention. The investigation develops the mathematics of loss cost, frequency and severity, exposure definition, loss development, claim cost trend, coverage changes, credibility, catastrophe provisions, predictive models, uncertainty and model validation. It explains how the cost estimate is translated into a rate indication and how that indication relates to, while remaining analytically distinct from, expenses, profit, capital, contract terms and an insurer's implemented market price. Primary professional references include the Casualty Actuarial Society's ratemaking literature, the Actuarial Standards Board's standards on future costs and related methods, and relevant official supervisory texts. Peer-reviewed statistical literature supplies alternative modeling approaches and sources of parameter uncertainty. An independently recalculated synthetic portfolio illustrates a full deterministic pure premium calculation and a separate frequency–severity projection; no synthetic figure is represented as observed market evidence. Comparative analysis identifies the circumstances in which apparent differences between rating methods are differences in estimand, data maturity, exposure mix or contract scope. The central conclusion is that a defensible prospective loss cost is a clearly dated, exposure-denominated and contract-specific expectation supported by explicitly reconcilable historical adjustments, forward-looking assumptions and quantified sensitivity. Its reliability depends as much on the identification of the future insured portfolio as on the choice of estimator.

Keywords: Prospective loss cost; pure premium; property and casualty insurance; actuarial ratemaking; expected claim cost; exposure base; claim frequency; claim severity; loss development; incurred but not reported losses; trend; rate indication; credibility; generalized linear models; catastrophe loading; validation; actuarial standards; uncertainty.

Introduction

Insurance contracts transfer specified contingencies in exchange for a price fixed before their complete claims cost is known. This temporal asymmetry creates the central problem of actuarial pricing: the estimate must describe the insured events, exposure mix, coverage terms and cost level expected during the future period, even though the most directly observable data concern earlier periods and incompletely settled claims. Werner and Modlin (2016) organize the analysis around the fundamental insurance equation, which reconciles prospective premium, loss, expense and profit components. The present research concentrates on the loss component, including the documentation needed to connect it to the complete equation.

The Casualty Actuarial Society's ratemaking statement describes the estimation of the costs of risk transfer; its unusual institutional history is relevant to citation accuracy. First issued in 1988, it was rescinded in 2020 and reinstated in 2021 for reference in U.S.-regulated ratemaking. Its current reference status should not be confused with a statute, an internationally binding standard or a substitute for jurisdiction-specific requirements. Actuarial Standard of Practice No. 53, Estimating Future Costs, applies to covered prospective property/casualty risk-transfer and risk-retention actuarial work and explicitly reaches components of a future cost estimate as well as their aggregation. The standard is an especially useful organizing authority because loss costs are often estimated separately from claims administration costs, other expenses, reinsurance and capital costs.

Pure premium is sometimes presented as a mechanical quotient of observed losses and exposures. That quotient is a historical average with a particular valuation date. It becomes a prospective actuarial estimate only after the relevant claims have been brought to ultimate, adjusted to comparable contractual and exposure bases, and projected to future conditions using sufficiently justified assumptions. An average can be computed precisely and nevertheless answer the wrong question. For instance, total paid claims on a young liability accident year may dramatically understate eventual claims because only the earliest payments are represented. Conversely, the inclusion of a single severe event in a short experience period may overstate normal expected costs if its exceptional nature and the associated catastrophe exposure are ignored.

This monograph adopts a defined estimand. Let LTL_T denote the aggregate ultimate claim costs, on a stated gross or net basis and for a specified coverage definition, attributable to a future portfolio during target period TT. Let ET>0E_T>0 be the matched exposure units for that period. The aggregate prospective loss cost is E[LT∣I]/ET\mathbb{E}[L_T\mid \mathcal{I}]/E_T, where I\mathcal{I} is the pricing-date information set and where the exposure quantity is treated as a known planned denominator. If ETE_T is random, the exposure-weighted ratio E[LT]/E[ET]\mathbb{E}[L_T]/\mathbb{E}[E_T], the conditional loss cost for one unit, and E[LT/ET]\mathbb{E}[L_T/E_T] may differ; the selected measure must be identified. This apparently technical distinction matters in portfolios with material growth, cancellation, seasonality and changing mix.

The scope is international. U.S. actuarial standards are interpreted in their proper professional and jurisdictional setting, whereas conceptual and statistical methods are analyzed for their broader applicability. The U.S. NAIC model rating law illustrates one regulatory architecture; it must not be read as a uniform law enacted identically in all U.S. jurisdictions, much less as global law. The European Union's Solvency II Article 77 concerns technical provisions on a probability-weighted present-value basis and is introduced to clarify a neighboring measurement purpose, rather than to make technical provisions interchangeable with a prospective pricing loss cost.

The method is source-critical comparative synthesis, mathematical analysis and transparent calculation. First, source claims are tied to primary standards, original research or authoritative publication records. Second, mathematical identities are separated from empirical assumptions. Third, numerical examples are wholly synthetic and independently recalculated. Fourth, distinctions between systematic, parameter, data and judgmental uncertainty are retained through the analysis. Fifth, no new authorial theory, doctrine or proprietary rating system is asserted. The resulting conclusions constitute an original analytical integration of established actuarial and statistical knowledge.

Pure Premium and Risk Cost

The prospective quantity and the insurance equation

The pure premium, often called the loss cost or burning cost, expresses expected insured claim cost per unit of relevant exposure. Werner and Modlin's Basic Ratemaking states the historical identity as total losses divided by total exposures and, with consistent definitions, as claim frequency multiplied by claim severity. The future-oriented construction retains the same dimension but requires projected values. If expected claim count is E[NT]\mathbb{E}[N_T] and expected severity per count is E[ST]\mathbb{E}[S_T] under a suitable conditional loss model, a commonly used representation is

πT=fTsT,fT=E[NT]ET,sT=E[LT]E[NT].\pi_T = f_T s_T,\qquad f_T=\frac{\mathbb{E}[N_T]}{E_T},\qquad s_T=\frac{\mathbb{E}[L_T]}{\mathbb{E}[N_T]}.

The equations are identities when the severity is defined by the aggregate expectation ratio and the same loss population and valuation convention apply. The equality should not be replaced by E[NT]E[XT]\mathbb{E}[N_T]\mathbb{E}[X_T] without considering dependence between frequency and severity or the assumptions of the model. Covariate-conditioned frequency–severity estimation is often preferable because exposure characteristics affect both dimensions. With exposure class xx, the portfolio expectation is a weighted sum ∑xET(x)fT(x)sT(x)\sum_x E_T(x) f_T(x)s_T(x), divided by ∑xET(x)\sum_x E_T(x).

The term pure premium is used differently across documents and jurisdictions. Some calculations include all allocated and unallocated loss adjustment expenses; others first estimate the indemnity or benefit loss cost and then add separately defined claims-handling amounts. The authoritative premium foundations chapter gives a formulation incorporating claims and claims adjustment expenses and distinguishes the resulting quantity from non-claim insurer expenses and the profit provision. A scholarly analysis must record the chosen convention and reconcile it to the intended premium equation, rather than assume a universal accounting identity.

The target cost may be gross of reinsurance, net of an agreed deductible, net of salvage and subrogation, limited to an occurrence or policy limit, or restricted to a stated peril. Each transformation changes the random variable under examination. Accordingly, the rate indication should carry a complete cost definition: coverage, trigger, reporting basis, inflation basis, valuation date, expense boundary, participation, contractual limits, currency and tax treatment. Two prices may differ because they measure two different bundles of rights and losses.

Pure premium, frequency, severity and expenses

Frequency, measured as expected claims per exposure, is sensitive to how a claim and an exposure are counted. A single event producing multiple claimants may be recorded as one occurrence but several claim files. Reopened claims can affect the count under report-year or paid-year conventions. Severity can mean indemnity per reported claim, per paid claim, per occurrence or per claim closed with payment; these denominators yield different levels. The CAS educational text stresses the importance of consistent claim-count and loss-cost definitions before frequency and severity are calculated.

For a simplified rate per exposure unit, let π\pi include the prospective costs assigned to losses and claims adjustment expenses, FF denote fixed non-claim expenses per exposure, vv premium-proportional expense ratio and qq the assumed underwriting profit provision as a proportion of premium. An internally consistent simplified relationship is

P=π+F1−v−q,P=\frac{\pi+F}{1-v-q},

provided v+q<1v+q<1 and provided the cost and investment-income conventions are those explicitly chosen for this simplification. This is an accounting reconciliation, not a universal statutory tariff formula. Changes in payment timing, investment income, commissions, assessments, capital return or premium collection can alter it. The revised ASOP No. 29 on expenses has been effective since July 1, 2024, while ASOP No. 30 treats profit, contingencies and the cost of capital. Their subject matter helps prevent the pure-premium estimate from silently absorbing unmeasured expense or return assumptions.

A rate and a market price are also decision-distinct. An actuarial cost-based indication may be modified for legally permissible transition, distribution and competitive reasons. Such modifications should be visible in the reconciliation: the historical experience, prospective risk estimate, permissible filed rate, applied premium and actual earned premium are different records. The science of pure premium therefore provides an auditable foundation for pricing decisions without converting every business adjustment into an actuarial loss estimate.

Historical foundations and contemporary professional authority

Traditional loss cost methods were developed to transform past insured outcomes into prospective tariff levels and relativities. Their established computational components remain development to ultimate, adjustments for current coverage and benefit levels, economic and non-economic trend, catastrophe consideration, experience weighting and loading for other costs. The modern practice extends these components with structured statistical inference, predictive segmentation and scenario analysis. Werner and Modlin (2016) provide an integrated account of this sequence; the second edition of Goldburd, Khare, Tevet and Guller (2025) supplies a contemporary treatment of generalized linear models in insurance rating.

The regulatory and professional status of a text is as important as its mathematical usefulness. A professional standard binding on members of designated actuarial organizations in the United States, a CAS educational monograph, an advisory model act and a peer-reviewed empirical paper have different evidentiary functions. Professional material provides definitions and processes within its stated scope; research papers support particular model properties or empirical findings; model law supplies a legislative template rather than direct proof of local law. A country-specific tariff may be governed by mandatory rates, file-and-use arrangements, prior approval, actuarial fairness rules, consumer protection or market competition; the corresponding legal conclusions require the law actually in force for the selected product and jurisdiction.

The December 2025 revision of ASOP No. 20 is effective for covered work products issued on or after June 1, 2026. It broadens the framework for property/casualty cash flow analysis, including applications to ratemaking. The date matters when a cost estimate depends upon discounting or cash flow timing. In contrast, ASOP No. 53's last revision and effective date are December 2017 and August 1, 2018, respectively. A source ledger should record these distinctions rather than reference standards generically as if their texts and dates never changed.

Ratemaking, reserving and decision use

Claim reserving estimates remaining costs of past insured events as of a valuation date, while prospective ratemaking estimates costs attributable to future exposures and policy conditions. Development factors and stochastic models can be used in both processes, but their target periods and cost definitions differ. Mack's 1993 chain-ladder uncertainty analysis addresses the standard error of reserve estimates; it provides essential insight into uncertainty in projected ultimate losses, without directly establishing the future exposure distribution or the applicable premium rate. The actuarial textbook by Wüthrich and Merz (2008), Stochastic Claims Reserving Methods, develops reserving methods and prediction uncertainty pertinent to the historical-to-ultimate step.

Under ASOP No. 43, property/casualty unpaid claim estimates have a specific prospective payment objective for incurred events. The pricing actuary may use an ultimate loss estimate developed from reserving techniques, then apply further trend, coverage and exposure adjustments to fit a future insurance portfolio. Reusing the mathematical development tool does not erase the distinction between ultimate cost of past cohorts and expected cost of new cohorts.

A third use is financial-reporting measurement. Solvency II's technical provisions framework, for example, requires a best estimate of obligations plus a risk margin, with specified present-value treatment. An expected future pure premium may be an input to a valuation calculation, yet a reserve includes the stock of obligations and depends on its own boundary and discounting rules. The interpretation of one estimated number must always accompany its intended decision.

Prospective Experience and Exposure

Exposure measurement and temporal alignment

The denominator determines the meaning of the loss cost. A car-year counts automobile insurance exposure over an annual equivalent; a building-year has a different distribution of value, use and peril characteristics; payroll, receipts, insured values, turnover or vessel-days can be more suitable for other coverages. Exposure should represent the relevant opportunity for insured loss and should be proportional to risk to the extent justified by the chosen rating structure. A count of policies alone can be misleading when their terms, limits, insured values or durations differ.

Written exposure relates to insurance contracts issued in a period, earned exposure to portions of coverage provided in a period, and in-force exposure to a valuation snapshot. Accident-year losses generally align most naturally with earned exposure attached to the occurrence period, whereas policy-year losses can be matched to the exposure of policies incepting in the cohort. When the data span rapid growth, declining limits, unusual retention or business transfer, using written exposure against calendar-year paid losses can create an artificial pure premium trend. This is the same temporal consistency issue addressed in the NAIC's product filing guidance, which distinguishes common loss aggregation bases.

An exposure audit should confirm units, earned-period fractions, cancelled policies, endorsements, underlying sums insured, reinstatements and the treatment of zero-exposure records. A payroll-class liability policy, for instance, may produce premium audits after the insurance period, meaning that final rated payroll can be known only with a lag. A projected average severity per claim can also shift when deductibles or insured amounts change. Exposure matching consequently requires both a record-level time key and a contract-terms key.

If a portfolio contains multiple units ii, a prospective class cost can be written

πT=∑iei,Tμi,T∑iei,T,\pi_T=\frac{\sum_i e_{i,T}\mu_{i,T}}{\sum_i e_{i,T}},

where ei,Te_{i,T} is expected exposure and μi,T\mu_{i,T} is cost per comparable exposure. The formula shows why the future mix must be forecast: changes in the proportions of classes alter the portfolio pure premium even when every within-class risk parameter stays constant. A flat aggregate indication can conceal opposing trends among classes, territories and coverage layers.

Claim valuation bases and data reconciliation

Paid losses, reported incurred losses, case outstanding amounts, incurred-but-not-reported claims, incurred-but-not-enough-reported development and ultimate losses should be reconciled through the same claim population. At any valuation date, incurred accounting amounts may equal cumulative paid plus outstanding case estimates, subject to the entity's case and recovery conventions. Ultimate losses are an estimate of the total eventual cost of the cohort. A paid development triangle and an incurred development triangle provide complementary information, because settlement speed and reserving practices can change independently of actual risk.

Source-level reconciliation should identify whether gross payments include defense costs, whether claim adjustment expenses are allocated to claims, whether salvage and subrogation are netted, and how deductibles or self-insured retentions are represented. Recoveries should be associated with the correct accident cohort or prospective contract layer. Changes in claims-system coding, audit standards, litigation strategy and settlement practice can make a historical ratio internally consistent while externally non-comparable.

The Data Quality standard, ASOP No. 23, guides selection, review, use and reliance on data, and distinguishes a review from a formal audit. It does not require a conventional data audit in every actuarial assignment, nor does it apply to a wholly hypothetical data set. This monograph's numerical illustrations are explicitly hypothetical and are independently checked as arithmetic; that check is not presented as validation of an insurer's actual records.

Claim inflation requires an appropriately granular valuation. A closed claim settled in a high-inflation period may have a different cost distribution from a newly reported claim that will settle years later. Payment date, accident date, report date and contract inception date capture different stages of the claim process. Long-tail liability portfolios may require separation of payment inflation, defense expenses, claimant behavior and legal changes rather than a single annual multiplier.

Portfolio composition and comparability

Historical experience is representative only to the extent that its insured risk, underwriting rules, distribution channels, retention, territory, coverages and exposures map to the prospective business. The actuary must distinguish a real change in within-risk loss cost from a composition shift. Suppose two property classes each retain a fixed true cost of 300 and 900 per year. A past portfolio equally split between the classes has a 600 average; a future portfolio with 80% of the low-risk class has a 420 average. Nothing about the within-class risk has improved. The aggregate change is entirely the change in exposure weights.

A comparison may therefore use standardization: apply historical class costs to a common reference exposure distribution, then separately estimate the future distribution. This isolates class-specific cost movements from changing volume. Similar logic applies to exposure-weighted average policy limits, deductibles, attachment points and peril geography. The risk classification logic should remain consistent with ASOP No. 12 in the contexts where that standard applies.

Changes in underwriting appetite can also create selection effects. If an insurer exits high-risk regions after adverse experience, using its previous aggregate losses and simply dividing by a smaller count of policies may misstate the current expected cost. Conversely, newly acquired segments with sparse history can be underrepresented. Related insurer, industry or market data can supply a partial basis, with explicitly documented similarity and credibility weights. Pooling dissimilar populations solely to raise the observed claim count can create spurious precision.

Contracts, attachment points and loss limitations

The insured loss is created by both the hazard process and the contract's payment function. With ground-up event loss XX, per-event deductible dd, and insurer payment limit ℓ\ell above the deductible, a simplified indemnity is

Y=min⁡{(X−d)+,ℓ},(z)+=max⁡(z,0).Y=\min\{(X-d)_+,\ell\},\qquad (z)_+=\max(z,0).

Then the expected cost is E[Y]\mathbb{E}[Y], not E[X]−d\mathbb{E}[X]-d or the raw mean of the observed ground-up losses. The deductible, limit and occurrence definition determine the part of the severity distribution borne by the insurer. A capped loss cost may underestimate excess-layer costs when improperly extrapolated, and a reported-loss database with truncation cannot automatically identify the unobserved lower or upper tail.

An increased-limits factor, excess loss factor or deductible credit can be derived from a fitted severity distribution or compatible external experience, but its applicability depends on exposure and claim-type composition. Policy aggregate limits introduce dependence across claims. Reinstatements and annual aggregates require simulations or joint distribution calculations where the simple per-event payment formula is insufficient. These contract mechanics are part of the prospective risk cost, not cosmetic adjustments to the final premium.

Coverage evolution is particularly important in liability insurance, where the scope of covered damages, defense obligations and retroactive dates can vary materially. A property product covering replacement cost at current construction prices differs from one covering stated actual cash value. Prospective ratemaking therefore requires a documented map from historical insurance obligations to the proposed obligations.

Data architecture, quality and provenance

A reproducible loss cost study can be represented as linked records for policy, exposure, risk attributes, claim, payment, reserve valuation, recoveries and rating factors. Each record should carry a stable identifier, event and accounting dates, source-system version, currency and coverage basis. An effective audit trail preserves both the original input and the transformations, including development-factor selections, data exclusions, trend assumptions and manual corrections.

The scientific record should identify the valuation date and the projected earning or occurrence dates. A rate study often combines multiple data vintages: exposures may be updated monthly, case estimates daily and industry trend data quarterly. The study's cut-off freezes the information set for a particular indication. Subsequent adjustments belong to a new analysis version, with explicit reasons. This prevents hindsight from being misrepresented as information available at the original pricing date.

The Modeling standard, ASOP No. 56, identifies considerations for designing, selecting, using, reviewing and evaluating models, including model purpose, inputs and output validation. It supports a governance approach in which a predictive algorithm is documented together with the surrounding claims and underwriting system. Model version control, reproducible data lineage, access controls and analytical review matter because a model with correct coefficients can still be applied to a wrongly coded exposure file.

For cross-border analysis, additional metadata are necessary: currency conversion dates, local inflation measures, claim settlement rules, taxes and levies, privacy and permissible data use. Source-data protection constrains what can be disclosed, but it does not prevent a transparent specification of definitions, methods, assumptions, ranges and tests. Where claims records cannot be public, independently reproducible synthetic demonstrations can illuminate the mathematics without implying empirical access.

Claim Development and Trend Adjustments

Development from observed to ultimate loss

A loss cohort is incompletely observed until every relevant claim and payment has been resolved. Development is the estimate of how the currently observed amount will change as the same cohort matures. For accident year ii valued at development age aia_i, a selected cumulative development factor DaiD_{a_i} produces a conventional estimate U^i=Li,aiDai\widehat U_i=L_{i,a_i}D_{a_i}, where Li,aiL_{i,a_i} is a consistently defined paid or incurred loss measure. This simple multiplication combines a claim observation with an inference about the future. Its legitimacy rests on the relationship between earlier and later valuations of comparable claims, rather than on the visual appearance of a triangle.

In an age-to-age chain-ladder calculation, development ratios can be estimated from cumulative triangular observations. If Ci,aC_{i,a} denotes cumulative amount for origin period ii at age aa, a volume-weighted factor is

d^a=∑i∈IaCi,a+1∑i∈IaCi,a,D^a=∏k≥ad^k×dtail,\widehat d_a=\frac{\sum_{i\in I_a}C_{i,a+1}}{\sum_{i\in I_a}C_{i,a}}, \qquad \widehat D_a=\prod_{k\geq a}\widehat d_k \times d_{\mathrm{tail}},

where IaI_a contains only origin cohorts with both ages observed and dtaild_{\mathrm{tail}} is a separately supported allowance, if needed. The appropriate triangle may be based on paid losses, incurred losses, claim counts or average claim sizes. A selected factor can differ from the unadjusted volume-weighted ratio when there are material changes in settlement practice, outliers or data quality.

The reserve literature provides rigorous cautions. Mack (1993) derives a distribution-free standard error under specific chain-ladder assumptions. Such an error measure characterizes uncertainty within the selected structural framework; it is not an assurance that those assumptions fit a rapidly changing portfolio. The more comprehensive treatment by Wüthrich and Merz (2008) examines stochastic reserving methods and prediction error. In ratemaking, the resulting ultimate loss estimates are still historical-cohort costs and therefore require separate forward-looking adjustments.

Development can be performed on a claim-count basis, a severity basis or an aggregate loss basis. For a homogeneous class, estimating the number of claims eventually reported and their ultimate average size may reveal different trends from multiplying a cumulative incurred triangle by one aggregate factor. An abrupt change in case reserve adequacy can distort incurred development but leave paid development less immediately affected. An acceleration of payment can have the opposite pattern. The difference is analytically useful: it should trigger investigation into claims operations, cohort maturity and source definitions.

A defensible development review compares payment and incurred patterns, prior-period back-tests, company and relevant industry experience, case-reserve adequacy, the maturity of large claims and the influence of unusual events. It should include an explicit tail study where claims remain open beyond available ages. For sparse long-tail portfolios, judgment and externally informed expected loss ratios may need greater weight. For short-tail motor physical damage, very long development tails may have little relevance, although late reporting and subrogation can remain material.

The commonly used expected-loss or Bornhuetter–Ferguson family combines an independently developed expected ultimate cost with the portion of losses anticipated to emerge after the valuation date. Its conceptual attraction is limited responsiveness to immature observed losses, but its output is only as representative as the prior expected loss level and emergence pattern. It should be understood as a different conditional expectation structure from pure chain ladder. Neither method creates information that is absent from historical claims, and neither obviates checks for contract changes and prospective cost level.

Time trend and future cost levels

Development and trend answer different temporal questions. Development estimates additional emergence within an already incurred cohort; trend estimates changes between exposure periods attributable to evolving conditions. The distinction is explicit in the ASOP No. 13 definition of trend, which excludes the process commonly called development. Conflating them can double-count or omit cost change.

Let tit_i be the representative accident or exposure cost date for historical cohort ii, and tTt_T the representative cost date for the future business. With a constant annual effective trend gg, one simplified projection is

U^i(T)=U^i(1+g)tT−ti.\widehat U_i^{(T)}=\widehat U_i(1+g)^{t_T-t_i}.

This formula makes two assumptions visible: the annual rate gg adequately represents relevant changes over the full projection interval, and the chosen dates are meaningful average cost dates. Non-integer projection periods are legitimate, but the claim cost date should not be confused with the date when premiums are written or when ultimate claims are paid. When inflation or portfolio change is nonlinear, direct index ratios, piecewise trends or component-specific projections may be preferable.

The prospective claim cost level reflects multiple forces: repair labor prices, construction materials, medical costs, litigation outcomes, settlement practice, loss frequency, fraud patterns, technology, replacement costs, climate exposure, supply-chain conditions and applicable law. A general consumer price index can serve as contextual information but cannot automatically replace the inflation of the insured loss category. For medical liability, for example, changes in claim severity may reflect court awards and claims management as well as health-service price indices. For property loss, material and labor scarcity after a catastrophe can produce a temporary demand-surge component not captured by a national aggregate index.

Statistical trend analysis may compare accident-year frequency, severity, pure premium, external indices and prospective exposure changes. Sensitivity to end dates, trend windows, exponential versus linear functional forms and exceptional periods must be examined. In a rapidly nonstationary environment, historical extrapolation is best treated as a conditional scenario rather than a timeless property of the risk. The ASOP No. 53 future-cost guidance makes the future period's conditions central to the actuarial task.

Frequency and severity decomposition

An aggregate pure premium trend can be decomposed into frequency and severity trend when the components are defined compatibly. Suppose the claim frequency changes at annual rate gfg_f and claim severity at gsg_s. Under a multiplicative approximation,

1+gπ=(1+gf)(1+gs).1+g_{\pi}=(1+g_f)(1+g_s).

Thus a 2% annual decrease in frequency combined with a 7% annual increase in severity produces an overall annual pure-premium change of 0.98×1.07−1=4.86%0.98\times1.07-1=4.86\%. Over two years, a starting frequency of 0.20 claims per exposure and a starting severity of $3,000 produce a projected frequency 0.20(0.98)2=0.192080.20(0.98)^2=0.19208, projected severity 3000(1.07)2=$3,434.703000(1.07)^2=\$3,434.70, and projected loss cost 0.19208×3434.70=$659.740.19208\times3434.70=\$659.74. The starting loss cost was $600. The increase in aggregate cost is approximately 9.96% over two years. Every numerical input is synthetic.

Decomposition is useful because the drivers have different implications for pricing and validation. A rise in reported claim frequency may follow a change in reporting thresholds, product design or covered exposures rather than an increase in the underlying frequency of accidents. Rising average severity may reflect truncation under higher deductibles, increasing policy limits or an increasing share of injury claims. A frequency–severity model should ideally forecast the cost conditional on the same contract layer and claim definition. The 2016 peer-reviewed study by Frees, Lee and Yang demonstrates how multivariate frequency–severity regression can address correlated insurance outcomes, although its estimated relationships are specific to its modeled data.

Exposure can affect the two components differently. A fleet growing through smaller vehicles might have more insured trips but lower average claim severity. A home portfolio changing from low-rise to high-value coastal properties might experience a shift in catastrophe loss distributions that cannot be explained by multiplying last year's aggregate average by an inflation factor. Even when aggregate claims are forecast through a single compound model, separate diagnostic frequency and severity analyses can uncover definition changes and assumption conflicts.

Shifts in law, technology and claims handling

Future losses are conditional on institutions and behavior. Legal reforms can alter liability standards, available remedies, limitation periods, defense obligations and settlement timing. Technology can change claim incidence, detection, repair cost and the severity distribution simultaneously. A new vehicle safety device may reduce collisions while making repairs more expensive; changes in building codes may reduce certain disaster losses but apply unevenly across an existing building stock. Estimates should specify whether an identified influence is supported by observed data, expert analysis or a scenario assumption.

Trends in a historical insurance portfolio sometimes reflect internal changes rather than external risk. A claims department may accelerate closure, change case-reserving policy or revise fraud detection. If reported severity increases because low-value claims are screened out before they become recorded claims, the change can be a denominator effect. Similarly, higher deductibles may reduce insured claim counts while leaving the underlying accident process unchanged. A future estimate must incorporate the actual insured payment function, not mechanically carry forward an observational artifact.

Claims judicial inflation is particularly relevant for long-tail liability. A single exponential trend over many years can imply implausible costs when statutes or courts create step changes. Alternative methods can use separate calendar-period level shifts, explicit legal scenarios or carefully selected comparable portfolios. Their use should be disclosed along with the sensitivity of the result to unresolved litigation or regulatory developments. The fundamental criterion remains fitness to the future contractual risk cost.

Large losses and catastrophe experience

A portfolio may contain routine losses, individual large losses and catastrophe losses with substantially different frequency, severity and dependence structures. One large liability claim can dominate an otherwise stable series; a natural catastrophe can affect many policies in the same location. Such events need explicit treatment because frequency and claims dependence determine both the expected cost and the capital required to absorb exceptional outcomes.

The established ASOP No. 39 catastrophe framework asks actuaries to identify relevant catastrophe exposures, evaluate historical and noninsurance information, consider future conditions and determine an appropriate provision. The standard explains how a short historical experience window with no disaster can omit real exposure, while a window containing a major event can overstate a normalized expected cost if that event is not adjusted consistently. It also recognizes demand surge, contagion and differences in claims adjustment expense behavior.

One method separates identified catastrophe losses from the attritional data, trends the remaining claims to the prospective cost level, then adds a catastrophe allowance based on a defined expected annual loss under future geography, coverage and insured values. The method is only coherent when the historical catastrophe component has actually been removed and the allowance is expressed per corresponding exposure. A separate large-loss limitation may cap historical individual claims at a selected threshold, develop the limited amounts and restore an above-threshold load supported by a severity model. Overlapping catastrophe and large-loss definitions must be reconciled to prevent duplicate loading.

Catastrophe models can use event frequency distributions, hazard intensity, exposure geolocation, vulnerability, insurance financial terms and simulated damage outcomes. The ASOP No. 38 catastrophe modeling standard provides professional context for examining catastrophe models, including model assumptions, input suitability and result interpretation. A modeled annual average loss is an expectation under the chosen model and event set, not the most likely one-year claim payment. Its uncertainty depends on hazard and climate assumptions, engineering vulnerability, geocoding, exposure concentration and the tail behavior of insured events.

The distinction between expected loss and risk loading is critical. A modeled catastrophe expected annual loss belongs within an expected claim-cost estimate; a separate charge for capital, risk bearing or adverse deviations has a different economic purpose. Combining them without label produces a rate that cannot be reconstructed. The same principle applies to reinsurance: expected gross loss, ceded recoveries, reinsurance premium and reinsurer capital charge represent different cash flows and contract layers.

Expected Loss Cost Construction

The mathematical estimator

For experience periods i=1,…,ni=1,\ldots,n, a conventional aggregate estimator begins with consistent observed losses LiL_i, exposure EiE_i, selected development factor DiD_i, cost-level factor TiT_i, coverage adjustment KiK_i and any other justified exposure-adjustment factor AiA_i. A transparent basic form is

π^Tattr=∑i=1nLiDiTiKiAi∑i=1nEi∗,\widehat\pi_T^{\mathrm{attr}} = \frac{\sum_{i=1}^{n}L_iD_iT_iK_iA_i} {\sum_{i=1}^{n}E_i^{*}},

where Ei∗E_i^{*} indicates exposure expressed on the common prospective rating basis. If adjustments are properly expressed per-exposure rather than as total-loss multipliers, the numerator and denominator must be transformed coherently. The particular placement of a factor in the algebra must follow its underlying meaning; a ratio applied blindly to losses alone may be dimensionally inappropriate.

A credibility-weighted estimate can replace the simple total-loss weighting. Write πiT\pi_i^T for the prospective adjusted loss cost of experience period ii and choose nonnegative weights wiw_i summing to one, based on exposure, relevance, temporal stability or predictive performance. Then π^T=∑iwiπiT\widehat \pi_T=\sum_i w_i\pi_i^T. Equal weights, exposure weights, inverse-variance weights and recent-period weights answer different variance and bias questions. A systematic choice requires evidence, not merely a preference for a smooth result. Traditional pure premium ratemaking methods and credibility concepts are developed in Werner and Modlin's fifth-edition textbook.

After attritional claims are projected, other expected claim-related components may be added with compatible bases:

πTtotal=πTattr+πTlarge+πTcat+πTclaim expense,\pi_T^{\mathrm{total}} = \pi_T^{\mathrm{attr}} +\pi_T^{\mathrm{large}} +\pi_T^{\mathrm{cat}} +\pi_T^{\mathrm{claim\ expense}},

provided each component is disjoint or any overlap has been explicitly reconciled. An expense ratio applied to attritional claims and an additional catastrophe allowance that already includes claims adjustment expenses should not lead to a further automatic expense multiplier on the catastrophe component. The model must reflect actual cost segmentation.

A reproducible synthetic worked example

Consider a fictional one-year property/casualty portfolio with three accident-year cohorts. The dollar amounts in the observed-loss column are cumulative paid attritional indemnity amounts, with identified catastrophe losses, individual large-loss excess loads and claims adjustment expenses excluded. The development factors project the paid indemnity amounts to ultimate attritional indemnity. Historical and future coverage terms and exposure units are assumed comparable; mix and limit adjustments are therefore one. The chosen annual effective attritional pure-premium trend is 5.5%. A future reference cost date is represented by the explicitly assumed elapsed years shown below. These periods are illustrative, not observations from any insurer or market.

Accident year Earned exposure units Paid attritional indemnity Selected ultimate factor Projection years
2022 1,600 $620,000 1.28 4
2023 1,700 $695,000 1.15 3
2024 1,800 $810,000 1.07 2
Total 5,100 $2,125,000 — —

Developing paid losses gives $793,600 for 2022, $799,250 for 2023 and $866,700 for 2024. The respective trend multipliers are 1.0554=1.2388246506251.055^4=1.238824650625, 1.0553=1.1742413751.055^3=1.174241375, and 1.0552=1.1130251.055^2=1.113025. These yield projected ultimate attritional indemnity totals of approximately $983,131.24, $938,512.42 and $964,658.77. Dividing each by its earned exposure gives $614.46, $552.07 and $535.92 per unit. The three-year exposure-weighted result is

π^Tattr=620,000(1.28)(1.055)4+695,000(1.15)(1.055)3+810,000(1.07)(1.055)21600+1700+1800=$565.94.\widehat\pi_T^{\mathrm{attr}}= \frac{ 620{,}000(1.28)(1.055)^4+ 695{,}000(1.15)(1.055)^3+ 810{,}000(1.07)(1.055)^2 }{1600+1700+1800} =\$565.94.

The numerical precision in the equation is retained for auditability, while reporting is rounded to cents. There is no statistical claim that the input trend or development factors are known to this level of accuracy.

Assume, separately, that expected claim adjustment expenses on the attritional component equal 12% of that component, and that a disjoint prospective catastrophe loss and catastrophe-specific claim expense provision has been independently estimated as $35.00 per exposure. These two assumptions are synthetic and are not inferred from the observed attritional loss data. The all-in claims-related loss cost under this convention is

π^Tclaim cost=565.941652785245(1.12)+35=$668.85 per exposure.\widehat\pi_T^{\mathrm{claim\ cost}}= 565.941652785245(1.12)+35 =\$668.85\ \text{per exposure}.

Suppose the fixed non-claim expense provision is $35.00 per exposure, variable expense ratio is 18% of premium and the simplified underwriting profit provision is 7% of premium. The implied illustrative total rate is

PT=668.8546511194744+351−0.18−0.07=$938.47 per exposure.P_T=\frac{668.8546511194744+35}{1-0.18-0.07} =\$938.47\ \text{per exposure}.

The difference of $269.62 between the all-in claim-related cost and the illustrative rate is the combined effect of fixed expenses and percentage-of-premium expenses and profit under the stipulated equation. It does not represent an additional pure premium. No claim is made about the commercial suitability or legal admissibility of this hypothetical tariff. In a live study the fixed and variable expenses, underwriting profit and investment-income treatment would require separate analyses under applicable standards including ASOP No. 29 and ASOP No. 30.

The formula is reproducible using only the rows and assumptions stated here; it can be entered into a spreadsheet or evaluated in a scientific programming environment. A code-equivalent expression for the initial attritional cost is `sum(paid[i] * development[i] * 1.055 ** years[i] for i in 1..3) / 5100`. That expression exposes the sequence of operations without relying upon proprietary software, confidential insured data or hidden judgment.

Trend sensitivity and the meaning of the estimate

The annual trend matters because it compounds over different projection periods. Holding every other assumption fixed, the indicated attritional pure premium under a 4.5% annual trend is $549.98, under 5.5% it is $565.94, and under 6.5% it is $582.24. The associated illustrative total rates, after the same 12% attritional claims expense factor, $35 catastrophe provision, $35 fixed expense, 18% variable expense and 7% profit assumption, are $914.64, $938.47 and $962.81. A one-percentage-point change on either side of the base annual trend therefore changes this simplified illustrative rate by roughly $24 per exposure.

Annual attritional trend Attritional pure premium Total claims-related cost Illustrative total rate
4.5% $549.98 $650.98 $914.64
5.5% $565.94 $668.85 $938.47
6.5% $582.24 $687.11 $962.81

This sensitivity isolates a single assumption; it is not a confidence interval. Development factors, claim definitions, catastrophe costs, geographic mix and expense ratios are treated as fixed for this table. In a genuine pricing analysis a credible sensitivity range would arise from historical fit, external drivers, scenarios and judgment. The table demonstrates a broader analytical result: an apparently small annual trend disagreement can materially influence a future rate when compounded through several cost years and then magnified by premium-proportional charges.

Credibility and related experience

Insurance classes vary dramatically in their statistical depth. An insurer with only a few hundred exposures or very infrequent claims may not possess sufficient independent experience to estimate its prospective loss cost reliably. Credibility procedures blend subject experience with relevant external or related experience. The ASOP No. 25 credibility guidance recognizes this use in ratemaking and describes selecting procedures suited to the context.

A simple algebraic expression is π^C=Zπ^S+(1−Z)π^R\widehat\pi_C=Z\widehat\pi_S+(1-Z)\widehat\pi_R, with 0≤Z≤10\leq Z\leq1, subject estimate π^S\widehat\pi_S, related estimate π^R\widehat\pi_R, and credibility weight ZZ. For illustrative purposes only, assume the synthetic subject estimate of $565.941652785245 has credibility 0.65 and a comparable prospective external estimate is $530.00. Then π^C=0.65(565.941652785245)+0.35(530)=$553.36\widehat\pi_C=0.65(565.941652785245)+0.35(530)=\$553.36 per exposure. Using the same claims-expense and catastrophe assumptions yields $654.77 in claims-related cost. The weight and comparison level are illustrative assumptions, not fitted outputs.

Credibility is an empirical and methodological choice. A full-credibility threshold based on a desired tolerance and probability level may be appropriate under specified claim assumptions, but it is not universally transferable across types of risk or sizes of claim. A low-frequency high-severity line can require far more data than a homogeneous high-frequency class for a given relative error. Empirical Bayes and Bühlmann-type methods relate pooling weights to variance between and within risks; hierarchical models can achieve a corresponding statistical partial pooling with explicit assumptions.

The relevance of a complement of credibility must be examined. An industry loss cost using different deductibles, coverage terms or territory is not automatically an appropriate comparator. If the insurer's class is distinct in known ways, those differences should be adjusted or the external data down-weighted. Strong statistical precision from a mismatched population can be inferior to a less precise but better defined estimate. ASOP 25 addresses the predictive value of the information, rather than endorsing a mechanical weight without evaluation.

Statistical alternatives and predictive pricing

Pure premium can be estimated through direct aggregate modeling, separate frequency and severity models, compound distributions, generalized linear models, generalized additive models, hierarchical Bayesian methods, credibility models or other predictive techniques. The choice depends on the loss process, availability and representativeness of individual records, contractual loss transformations, required transparency and validation evidence. No single model architecture dominates for every peril, class and jurisdiction.

A common generalized linear model for claim count uses a log link and an exposure offset,

log⁡E[Ni∣xi]=log⁡ei+xi⊤β,\log\mathbb{E}[N_i\mid x_i]=\log e_i+x_i^\top\beta,

while a separate positive claim-severity model, often with a gamma-type mean–variance structure, can use log⁡E[Si∣xi]=xi⊤γ\log\mathbb{E}[S_i\mid x_i]=x_i^\top\gamma. Combined expected cost per exposure follows only after the assumptions about claim-count, severity and conditional dependence have been addressed. The CAS's 2025 revision of Generalized Linear Models for Insurance Rating develops practical model building, validation and rating plan applications. The distinction between statistical fit and regulatory or operational fitness remains essential.

A compound Poisson–gamma or Tweedie-type structure can model an outcome containing a mass at zero and positive continuous total claims without separately estimating observed frequency and severity in every case. Smyth and Jørgensen (2002), in their paper on Tweedie claim-cost dispersion modeling, discuss modeling claim cost and dispersion using double generalized linear models. Such a specification can be useful for aggregate cost, but its assumptions about the occurrence process and positive loss distribution should be examined against actual claims, exclusions, deductibles and large losses.

Frees, Lee and Yang's (2016) multivariate frequency–severity study illustrates another important limitation of simplistic decomposition: multiple claim outcomes and their dependence can matter for estimating the total expected insurance cost. If the insured event creates correlated property damage and bodily injury claims, modeling each marginal separately and combining them without considering relevant dependence can misstate uncertainty, and may also distort expected payments under aggregate caps or nonlinear policy terms.

Predictive modeling needs out-of-time validation. Randomly splitting policies from the same catastrophe or event across training and test samples can leak event information; repeated renewals of a risk across the two sets can produce similarly optimistic assessments. Validation should match the intended future use and respect clustering by policyholder, household, vehicle, location or catastrophe event as appropriate. Model complexity should be weighed against interpretability, stability, data sufficiency, permissible predictors and governance requirements, consistent with the ASOP No. 56 modeling framework.

Rate indication reconciliation

A final indicated rate change requires comparison with a current rate level using consistent exposure, benefit and expense bases. If the prospective indicated premium per exposure is PTP_T and the premium at current rates for the same prospective exposure is PcurrentP_{\mathrm{current}}, a simplified overall indication is I=PT/Pcurrent−1I=P_T/P_{\mathrm{current}}-1. The denominator should be premium on-level to an explicitly chosen current rate schedule, adjusted for rate changes and earning patterns, rather than raw earned premium if the historical rate level has changed. This distinction is central to the NAIC's rate filing discussion and to conventional actuarial ratemaking.

A class indication further requires a consistent class-specific comparison. Suppose the portfolio-level actuarial premium is adequate in aggregate but class A is materially underpriced and class B overpriced relative to expected cost. The class-level adjustments may be economically and actuarially relevant even when the total target premium remains unchanged. However, legal constraints, credibility, customer impact and implementation schedules may restrict the feasible change. Those separate decisions must be documented, not retroactively inserted into a loss development factor.

There is also a circularity risk. If class exposures, realized premiums and underwriting selection are all functions of the prior tariff, a rating algorithm trained on observed insured outcomes estimates conditional behavior under the historical underwriting and sales system. Rate changes can alter who buys and retains coverage. A forward-looking indication should therefore consider plausible portfolio response and selection scenarios, especially where price elasticity and competitive dynamics are material. This is an economic interpretation built upon the expected loss cost, not an invention of a new statistical principle.

Stability and Limitations

Estimation error and process variation

Prospective loss cost is an expectation conditional on the selected information and future portfolio. The realized loss cost will generally differ from that estimate even when the estimator is well calibrated. This distinction is mathematically fundamental. For a homogeneous exposure portfolio with independent claim counts following a Poisson distribution with intensity λE\lambda E, and independently identically distributed severities XX with finite second moment, aggregate loss L=∑j=1NXjL=\sum_{j=1}^{N}X_j has

E[L]=λE E[X],Var⁡(L)=λE E[X2].\mathbb{E}[L]=\lambda E\,\mathbb{E}[X], \qquad \operatorname{Var}(L)=\lambda E\,\mathbb{E}[X^2].

Consequently, for a fixed exposure amount EE, the standard deviation of realized loss cost per unit is proportional to E−1/2E^{-1/2}, when the compound Poisson assumptions hold. This is process variability. Uncertainty in the estimated λ\lambda, severity distribution, future exposure mix or trend parameters adds estimation risk. A large homogeneous portfolio may reduce idiosyncratic process variation, yet remain exposed to correlated catastrophe or social-inflation changes shared by many policies.

The chain-ladder prediction error studied by Mack (1993) illustrates the need to distinguish uncertainty in ultimate historical losses from subsequent forward projection. The mean squared prediction error treatment depends on structural assumptions about triangle development. Its results cannot be transferred mechanically to expected future accident years without additional assumptions about portfolio continuity, inflation and new-risk emergence.

The relationship between prediction intervals and confidence intervals deserves special care. A confidence interval for a model parameter addresses estimation uncertainty under an inferential framework. A prediction interval for aggregate future claims also addresses randomness of future claim outcomes, and potentially parameter uncertainty. For high-severity lines, the predictive distribution is often asymmetric and heavy-tailed, making normal approximation inappropriate for distant quantiles. For some severity models the first moment exists but the variance may be extremely large or undefined; then conventional standard error formulas can be unstable or inapplicable.

Loss cost uncertainty should be reported in a form consistent with its economic use. A best-estimate expectation can be accompanied by ranges for trend, development and catastrophe load, as well as a separate risk distribution for capital planning. A risk percentile can be relevant for solvency and risk appetite but does not, by itself, define the fair or required average pure premium. In a contract with deductibles and aggregate limits, a change in tail shape can influence the mean insurer payment and the quantile in different directions.

Structural uncertainty, scenario analysis and validation

An actuarial model is a simplified representation of an evolving risk process. Structural uncertainty arises when the model family, data generating process or coverage mapping is misspecified. A credible estimate therefore requires diagnostics that could lead to rejecting the proposed method, not merely performance statistics calculated after it is accepted. The ASOP No. 56 requirements emphasize appropriateness for intended purpose, data and assumptions, validation, and model risk considerations.

One validation process back-tests earlier indications against subsequently observed and developed experience. The study must reconstruct the information available at the historical pricing date rather than retrospectively supply later claims data to the original estimator. For example, using a 2026 ultimate-loss estimate to judge what could have been known at a 2023 pricing date improperly benefits from hindsight. The correct back-test freezes each vintage of exposure, paid and incurred claims, factors and external indices. Its outcome can evaluate the combined process, and selected component back-tests can identify the source of error.

Validation metrics should be exposure-weighted and decision-sensitive. Mean error and aggregate actual-to-expected ratios can reveal calibration; deviance or proper scoring rules can compare probabilistic models; class-level diagnostics can detect systematic bias; stability plots can indicate rating-factor drift. A portfolio-wide actual-to-expected ratio near one may conceal persistent overprediction for low-risk classes and underprediction for high-risk classes. In the presence of sparse cells, the uncertainty of those apparent class errors must be estimated. The analyst should not treat random variation as evidence of discrimination or assume that untested segmentation is fair.

Scenario analysis should reflect causal exposure. A property portfolio can be evaluated under alternative frequency trends, storm hazard conditions, repair-price inflation and geographic mixes. A liability portfolio can be evaluated under alternative settlement delays, defense expenses and legal interpretations. Scenarios can reveal where the indicated premium depends on assumptions that are fragile to a small input change. Stress tests may deliberately leave the central forecast's plausible distribution and therefore should be labeled stresses rather than probability-weighted central estimates.

The sensitivity analysis in the earlier worked example shows a 4.5–6.5% annual trend assumption range producing a $914.64–$962.81 illustrative rate. This 48-dollar span is meaningful in the narrow calculation but is not a full pricing uncertainty interval. It excludes correlated shocks, errors in development factors, differences in exposure mix, and uncertainty in fixed expenses and profit provisions. A comprehensive risk disclosure should distinguish the model's tested sensitivity from the broader collection of untested risks.

The interpretation of future cost conditions is also date-dependent. The same model can be appropriate in a stable book and inadequate after major underwriting or coverage changes. Historical goodness of fit does not establish post-change forecast accuracy. Periodic reassessment is particularly important when economic inflation, climate hazards, vehicle technology or legal conditions change the relationship between observed covariates and insured costs.

Dependence, tails, data sparsity and model governance

Dependence among losses reduces the extent to which more policies alone produce stable aggregate costs. Coastal buildings exposed to the same hurricane, vehicles damaged by the same hailstorm and multiple liability claims resulting from one product defect are jointly exposed. Even when each policy is priced using a valid marginal expected loss cost, the aggregate portfolio variance, capital requirements and reinsurance costs depend on accumulation and dependence. Catastrophe models make that dependence explicit by simulating shared events; ordinary individual-policy regressions may require additional event, territorial or temporal structure.

In high-severity casualty lines, loss cost estimation may be strongly influenced by settlement values and relatively few cases. Severe claims can be right-censored by policy limits, open at valuation, affected by changes in legal interpretation and more sensitive to defense costs. Simple parametric fits may extrapolate poorly beyond observed amounts. Threshold selection, diagnostic tail plots, historical large-loss review and sensitivity to alternative heavy-tailed distributions are therefore substantive evidence, not discretionary decoration. The chosen upper tail can materially affect the mean when deductibles, attachment points and limits change.

Sparse segments create a conflict between segmentation and stability. More classes can increase homogeneity within each estimated cell but leave too few claims to estimate that cell's cost reliably. Partial pooling can improve estimation if shared structure is credible, but it may also impose inappropriate similarity. The CAS GLM monograph treats model selection, variable evaluation and ratemaking applications, offering tools to analyze such tradeoffs. The credibility framework of ASOP No. 25 also places the predictive value of data at the center of method selection.

An operational rating system usually contains more than a regression. Raw data are extracted, joined, cleaned and transformed; assumptions and model coefficients are selected; expected costs are normalized; rating factors are rounded or constrained; total-rate reconciliation is performed; approved rates are deployed; and subsequent experience is monitored. A defect at any step can invalidate the prospective estimate. Therefore, governance should include reproducible source-data snapshots, code review, deterministic numerical tests, independent model challenge where available, monitoring, implementation verification and an archive of prior rate versions. The scientific review must cover the entire chain of inference.

A sophisticated method can create the illusion of objectivity through many covariates. A model's outputs remain conditional on historical underwriting decisions, data availability, coding policies and permitted rating variables. Sensitive information may appear through proxy variables. Fairness review consequently examines variable definition, legitimate relation to risk, applicable law, data errors, reasonable alternative classifications, and differential effects. An internally calibrated model can still be inappropriate for a regulated product if it applies impermissible factors or produces inadequately documented changes.

Fairness, legality and communication

The objective of a prospective pure-premium estimate is a defensible prediction of insured cost for a defined risk transfer. Whether and how that prediction may influence a premium is a legal and institutional matter. The NAIC's Property and Casualty Model Rating Law illustrates regulatory principles including rate adequacy and prohibitions on excessive or unfairly discriminatory rates, but states adopt and implement their own laws. It is necessary to verify the current legal provisions of the particular U.S. state rather than present the model text as universally binding.

Across jurisdictions, permitted rating variables and disclosure requirements vary. The underlying statistical relation between a predictor and insured loss does not establish that the predictor is legitimate for price differentiation. Protected characteristics, indirect proxies, use of external data and automated decisions can raise legal and governance issues. An actuarial study should preserve an auditable separation among scientifically estimated costs, legally acceptable relativities, constrained filed rates and actual customer premiums. This supports both scientific accountability and regulatory oversight.

The Actuarial Communications standard addresses the disclosure of material methods, assumptions and limitations within the relevant professional scope. The same basic principle has an international scientific value: a reader should be able to identify what was measured, how a historical number was adjusted, what assumptions project it forward, which decisions rely on it, and where uncertainties remain. A concise result table without these foundations can be harder to scrutinize than a longer transparent methodological explanation.

Communication should avoid an unsupported inference from precision. A pure premium of $565.94 computed from synthetic factors is arithmetically exact to cents under its assumptions; the true prospective expectation for an actual portfolio may be materially different. Presenting more decimal places cannot resolve uncertain claim development or future inflation. When a model produces stable-looking outputs under incomplete evidence, the distinction between reproducibility and empirical accuracy should be explained explicitly.

The scientific ethics of source-based research also require precise attribution. A CAS teaching text can support a general procedural description without establishing local statutory compliance. A reserving study can support an uncertainty method without proving a future frequency trend. A supervised model can support claim-cost prediction without identifying the causal effect of a new tariff on customer composition. Keeping these roles distinct reduces conceptual overreach and improves the transferability of the analysis across jurisdictions.

Ratemaking Applications

Property and motor insurance

Property insurance often combines an attritional expected loss cost with exposure to infrequent correlated catastrophe losses. The exposure base may reflect building count, insured value, insured area, location or another rating unit. The loss estimate needs to reflect construction characteristics, occupancy, hazard zone, protection measures and coverage limits where legally permitted and data-supported. Property cost trend may differ for buildings and contents, and catastrophe vulnerability can change after modifications in building codes or mitigation measures. Aggregate historical losses alone seldom identify all these components.

A property ratemaking study can first project non-catastrophe indemnity to future replacement-cost levels, then evaluate separately a modeled catastrophe expected annual loss for the portfolio's future geocoded exposure and coverage. An appropriate catastrophe loss treatment reconciles historic catastrophe removals with the prospective loading. An insurer changing the percentage of coastal policies must update both expected attritional mix and catastrophe accumulation; a rate indication based only on the old portfolio distribution can be systematically wrong despite correct claim-level factors.

Motor insurance presents a different decomposition. Claim frequency may depend on driving exposure, vehicle usage, road conditions and accident incidence; severity can be driven by vehicle repair costs, medical and legal costs, sensor technology, repairability and policy terms. A collision may produce both property damage and injury claims with different reporting and settlement patterns. The claim-cost estimate should identify whether claims arise from vehicle physical damage, third-party property damage, bodily injury or comprehensive perils. Combining all motor claims into a single trend can conceal offsetting movements and mismatched development tails.

A useful motor application is separate frequency and severity validation by coverage, with partial pooling for sparse rating cells and periodic monitoring of changes in vehicle technology. However, a telematics or external data predictor can be statistically useful and still have legal or governance constraints. The predictive relationship should therefore be evaluated as one input to an allowable rating plan, not as a command to charge a customer an unconstrained individualized price.

Both property and motor ratemaking benefit from decomposition by occurrence and settlement timing. Short-tail physical damage may be largely known within a relatively short development interval, whereas litigated bodily injury may continue developing for years. A uniform age-to-ultimate factor or claim-cost trend for both can distort their relative contributions. The technical solution is coherent source-data segregation followed by transparent aggregation to the selected future rating unit.

Liability and long-tail business

Long-tail liability requires particular attention to reporting delays, settlement delays, legal development, defense costs, changes in limits and persistent inflation. An accident-year observation may contain only a small fraction of the ultimate claim costs at early valuations. The selected development model, loss-trend horizon and severity tail can dominate the resulting prospective indication. A calculation that appears to produce a stable annual pure premium may in fact be supported primarily by a selected expected-loss prior rather than a mature observed loss record.

The pricing actuary should examine reported counts, closed counts, paid severity, case incurred amounts and outstanding claim reserves together. If claim reporting accelerates because of a new notification platform, the observed early claim count may rise without an equal change in lifetime incidence. If settlement policies accelerate payments, paid development patterns may change while ultimate outcomes remain similar. A retrospective adjustment that fails to distinguish these processes can interpret administrative changes as risk trend.

The ASOP No. 43 unpaid-claim framework provides a relevant professional foundation for estimating outstanding losses and expenses, while the future cost assumptions fall within the analysis addressed by ASOP No. 53. Treating the two sets of methods as complements preserves their different temporal targets. An estimate of ultimate costs for old liability cohorts is an input to, rather than a final answer for, the future insured period.

Liability products with claims-made wording require special treatment. A claims-made cohort and an occurrence-based cohort may be associated with different exposure triggers and development pathways. Retroactive dates, extended reporting periods and prior-acts coverage can alter the insured event population. For classes with large defense expense exposure, a severity analysis should distinguish indemnity and defense payment patterns, especially when defense is inside or outside limits. A change in policy wording can change insured cost even with no change in the underlying accident rate.

In estimating the prospective cost, the analyst may combine historical company development with an industry benchmark, legal review, expert scenarios and hierarchical or credibility methods. Each supplement must be linked to a particular uncertainty it addresses. The method should disclose the share of the resulting expectation derived from observed company losses, related information and judgment. This is especially valuable when claims settle over periods longer than the available company history.

Catastrophe-exposed and specialty portfolios

Catastrophe-exposed insurance adds layers of uncertainty from event occurrence, geophysical hazard, asset vulnerability, correlated insured exposure, contingent business interruption and economic conditions following disaster. Expected annual catastrophe loss can be estimated with an event-based model incorporating hazard frequency, exposure location and policy conditions. It should be recognized as a model-conditioned long-run expectation, not the observed loss of an ordinary calendar year. The claims associated with catastrophe events may also require special development and inflation assumptions, including demand surge and settlement complexity.

Where insurance is written in layers, the expected ceded or retained loss differs from a proportional share of ground-up cost. For a layer attaching at aa with limit ℓ\ell, payment can be represented as min⁡{(X−a)+,ℓ}\min\{(X-a)_+,\ell\}. Expected cost depends on the entire relevant severity distribution, especially its tail. Changing attachment points can alter projected pure premium nonlinearly. The price of reinsurance additionally includes terms and the counterparty's expenses and capital requirements; the expected ceded claim cost is not by itself the full reinsurance premium.

Marine, aviation, cyber and other specialty classes may have sparse and heterogeneous experience. A few large events can dominate historical averages, while contract changes or emerging technology can limit the predictive value of distant observations. A useful strategy is to combine clearly comparable exposures with carefully documented expert scenarios, external loss experience and sensitivity tests. However, imported data are credible only to the extent that differences in geography, industry, attachment, legal regime and policy terms have been evaluated. The relative scarcity of data supports more explicit limits on inference, not automatic selection of a more complex model.

Some emerging risks are correlated across insureds and evolve faster than historical claims can mature. Cyber incidents, for example, may produce simultaneous claims through a shared software or infrastructure dependency, while policy language and coverage litigation remain in development. Prospective loss cost estimates should therefore specify a set of covered event definitions and scenarios. Their uncertainty is not adequately characterized by a single historical frequency trend when the underlying exposure network is changing. This observation follows the general mathematics of dependence and model misspecification; it is presented here as analytical application rather than a new authorial doctrine.

International interpretation

Prospective loss costs are expressed in local legal and economic environments. A model applied across jurisdictions may need to account for variations in tort law, compensation systems, health-service costs, inflation indices, monetary currency, public insurance and social security offsets, deductible conventions, claims-reporting practices, catastrophes and available exposure data. Pure premium can retain its general mathematical definition while the insured payment variable changes with the local contract and law.

A U.S. actuarial practice standard establishes professional expectations for work within its scope, not binding legal authority throughout the world. The same distinction applies to statistical research: a claim-frequency predictor found important in one country may have a different coefficient, data meaning or permissible use in another. An international model requires local calibration and independent evaluation, with contracts and legal constraints treated as data-defining features rather than afterthoughts.

European insurance regulation illustrates the difference between future pricing loss cost and financial reporting. The Solvency II Directive and the related rulebook's best-estimate provisions specify a valuation structure for existing insurance obligations. The best-estimate technical provision is a probability-weighted present value of future cash flows related to those obligations. This accounting and solvency purpose is distinct from the expected annual claim cost per exposure of a prospective new-policy portfolio. A common statistical engine may support both applications, but must apply the correct contract boundaries, cash-flow definitions and discounting requirements.

The NAIC's rate regulatory model text illustrates a further distinction: a technically credible loss cost is embedded in a regulatory decision about the permissibility of rates. Legal limits on rating variables, filing schedules and permitted expenses are jurisdiction-specific. It follows that international comparisons of “pure premium” should first reconcile coverage, insurer payment basis, exposure, currency and price level before attempting to compare numerical cost ratios.

Cross-border pooling may improve statistical precision for shared hazard characteristics but can also conceal local heterogeneity. Where geophysical hazards and engineering vulnerability are comparable, international catastrophe or engineering evidence may inform model assumptions. Where claims settlement and litigation practices differ, imported severity experience can mislead. The appropriate criterion is substantive transferability of the insured cost mechanism, not the convenience of combining large data sets.

Integrated analytical synthesis

The research establishes a connected set of conclusions that follow from the definitions and evidence rather than constitute a newly proposed theory. First, prospective loss cost is an expectation defined jointly by the future exposure portfolio and the contract payment function. Any estimator that changes either definition without explicit adjustment changes the object being estimated. This makes risk classification, exposure auditing and benefit reconciliation mathematically integral to pure premium ratemaking.

Second, claim development and trend should be evaluated as distinct transformations with separately documented assumptions. Development concerns the incomplete observation of historical incurred events; trend concerns changes in the future cost environment. A single scalar multiplier may summarize both effects in a simple example, but its interpretation cannot establish that the component assumptions are defensible. This is especially important where claim settlement, litigation and inflation change at different rates.

Third, the frequent presentation of pure premium as frequency times severity is useful because it exposes drivers, but the portfolio cost generally arises from a conditional, exposure-weighted aggregation. Frequency–severity decomposition requires compatible claim definitions and adequate treatment of dependence and policy terms. A correctly weighted overall rate may be wrong for individual classes if it conceals heterogeneity or future mix change.

Fourth, an actuarial indication must be separated from risk loading, expense loading, profit provisions and the final regulated or marketed price. The pure-premium equation produces one component of the insurer's future expected cost structure. The simplified example demonstrated that a $565.94 attritional indemnity cost becomes $668.85 in explicitly defined claims-related cost and $938.47 in illustrative total rate after stipulated additional provisions. Each numerical transition has an economic interpretation and requires its own evidence in live practice.

Fifth, estimation quality is multidimensional: representativeness of the future portfolio, claim maturity, data provenance, contractual comparability, calibration, model adequacy, scenario coverage and implementation governance all matter. Reproducible arithmetic is necessary for the integrity of a numerical demonstration, while empirical validity depends on the substantive evidence supporting the assumptions. A precise estimate without correct contract and exposure definitions has limited decision value.

Sixth, risk uncertainty must remain visible across the workflow. A central expected loss cost, a confidence range for its parameters, a forecast distribution of realized claims, a catastrophe tail outcome and a required capital provision answer different questions. They should inform one another without being collapsed into a single undefined margin.

Seventh, the international transfer of a pure-premium method is possible when the insured loss variable, exposure base, legal environment, future cost level and data provenance are reconstructed locally. An identical formula can legitimately yield different results across jurisdictions; international comparability rests on consistent definitions rather than uniform numerical rates.

These conclusions jointly support a rigorous research position: pure-premium ratemaking is a problem of prospective identification, measurement and inference. Its computation becomes meaningful when the anticipated insured liability and corresponding future exposure are defined, historical claims are transformed transparently and uncertainty is measured within the limits of available evidence. This synthesis provides a general analytical foundation for property, motor, liability, catastrophe and specialty applications without proposing a new theory of insurance pricing.

Conclusion

Prospective pure premium estimation has a precise scientific object: the expected ultimate cost of insured claims attributable to a defined future exposure population, expressed per comparable unit of exposure. The calculation combines three elements that must remain aligned. The claim population specifies which insured events and payment obligations are included. The denominator identifies the units over which the future cost is shared. The information set and forecast assumptions specify what was known at the pricing date and how the observed history is interpreted for future conditions.

The analysis demonstrates why a historical loss-to-exposure quotient is only a starting point. Claim development estimates eventual costs of prior insured events; time trend projects those costs to a future price level; contract and coverage adjustments identify the insurer's actual payment function; class and exposure transformations construct the prospective portfolio; and catastrophe or large-loss components account for risk absent from ordinary experience or materially distorted by rare occurrences. Each adjustment has an identifiable mathematical meaning and can be checked for overlap, timing and data consistency.

The frequency–severity decomposition clarifies the sources of expected cost. Forecast claim count per exposure and cost per claim must refer to compatible definitions. Under heterogeneous exposures, the total portfolio cost is determined by exposure-weighted conditional expectations. The same average loss cost can arise from different class mixtures, and the same class-level expectations can yield a different aggregate result when the future mix changes. An aggregate historical trend can therefore obscure significant changes in portfolio composition.

Statistical methods provide different ways to estimate these quantities, including traditional development-factor approaches, credibility, generalized linear models, compound distribution models and hierarchical techniques. Their quality is determined by the validity of the assumptions and evidence supporting the future risk, not by computational complexity alone. The claim development models used in reserving can inform historical ultimate losses, while the pricing decision requires additional assumptions about new exposure, future cost conditions and contract terms. This distinction remains especially important for long-tail and catastrophe-exposed portfolios.

The independently recalculated fictional study illustrates the full conversion from historical paid losses to developed and trended attritional indemnity, then to claims-related cost and a simplified total rate. A three-year synthetic experience produced a prospective attritional loss cost of $565.94 per exposure after the selected development and trend adjustments. Under separate, clearly stipulated claims adjustment expense and catastrophe assumptions, the claims-related cost was $668.85. Adding the illustrative fixed, variable and profit provisions gave a total rate of $938.47. The example's exact reproducibility establishes arithmetic coherence, not empirical representativeness. It also demonstrates that different assumptions about future trend can materially change an overall indication.

Uncertainty attaches to every important step: immature claims, scarcity and heterogeneity of historical observations, the distribution of extreme events, correlated exposure, unstable economic or legal conditions, and model design and implementation. A scientifically defensible study assesses uncertainty in a form appropriate to its use. Expected cost, process volatility, parameter estimation, model risk, capital requirements and stress losses are distinguishable quantities. The central cost estimate can be precise in definition while appropriately uncertain in magnitude.

International use requires the explicit reconstruction of local insured cost definitions, contractual limits, regulatory obligations and exposure measures. Professional standards and statistical research provide transferable methods within their stated scope, whereas legality and implementation are governed by the rules that apply to each market and product. Regulatory requirements, fairness and customer impact can constrain the translation from actuarial indication to implemented premium without changing the underlying mathematical identity of expected claim cost per exposure.

The original contribution of this dossier is its integrated source-critical analysis of estimation, contractual measurement, temporal projection, statistical validation and the boundary between pure premium and final rate. It draws together established actuarial mathematics and professional guidance, reconstructs their dependencies and develops transparent illustrative calculations. No new authorial theory, model or doctrine is advanced. The resulting scientific conclusion is that credible pure premium ratemaking is accomplished through a chain of auditable definitions, empirical inferences and prospective assumptions in which every transformation remains traceable to the insured risk and the chosen future exposure.

Methodological Appendix

A. Arithmetic verification of the synthetic portfolio

Let the inputs be the vectors of paid attritional indemnity L=(620000,695000,810000)L=(620000,695000,810000), development factors D=(1.28,1.15,1.07)D=(1.28,1.15,1.07), year exponents h=(4,3,2)h=(4,3,2), and earned exposures E=(1600,1700,1800)E=(1600,1700,1800). For annual trend g=0.055g=0.055, the projected numerator is

N(g)=∑i=13LiDi(1+g)hi.N(g)=\sum_{i=1}^3 L_iD_i(1+g)^{h_i}.

The three developed amounts are 793600793600, 799250799250, and 866700866700. Trend multipliers at 5.5% are 1.2388246506251.238824650625, 1.1742413751.174241375, and 1.1130251.113025. Multiplication produces 983131.242736983131.242736, 938512.41896875938512.41896875, and 964658.7675964658.7675, respectively. The sum N(0.055)=2,886,302.42920475N(0.055)=2{,}886{,}302.42920475 divided by total exposure 5,100 is 565.941652785245565.941652785245.

Let α=0.12\alpha=0.12 denote the synthetic claims adjustment expense factor applied exclusively to attritional indemnity and C=35C=35 denote a separately defined catastrophe-and-its-expenses expected cost per exposure. The claims-related loss cost is

Q(g)=(1+α)N(g)5100+C.Q(g)=(1+\alpha)\frac{N(g)}{5100}+C.

At the base trend, Q(0.055)=668.8546511194744Q(0.055)=668.8546511194744. Let F=35F=35 be the fixed non-claim expense per exposure, v=0.18v=0.18 the premium-proportional expense factor and q=0.07q=0.07 the underwriting profit factor. Since 1−v−q=0.751-v-q=0.75, the simplified illustrative rate becomes

P(g)=Q(g)+F0.75.P(g)=\frac{Q(g)+F}{0.75}.

The corresponding central rate is P(0.055)=938.4728681592992P(0.055)=938.4728681592992. The lower and upper trend sensitivities are computed by repeating all three cohort projections with g=0.045g=0.045 and g=0.065g=0.065, rather than applying the trend change only to the already aggregated base loss cost. These produce the three rates shown in the chapter's table.

The frequency and severity illustration similarly follows direct multiplication. Baseline frequency 0.20 and severity $3,000 imply an initial expected indemnity cost of $600. With annual frequency factor 0.98 and severity factor 1.07 over two years, prospective frequency is 0.19208, severity is $3,434.70, and cost is $659.737176, reported as $659.74. This compound result corresponds to two years of aggregate annual factor 0.98×1.07=1.04860.98\times1.07=1.0486: 600(1.0486)2=659.737176600(1.0486)^2=659.737176.

The credibility illustration uses Z=0.65Z=0.65 and selected complement $530: 0.65(565.941652785245)+0.35(530)=553.36207431040930.65(565.941652785245)+0.35(530)=553.3620743104093. With the same 12% attritional claims expense and $35 catastrophe-and-its-expenses provision, the resulting claims-related cost is 553.3620743104093(1.12)+35=654.7655232276585553.3620743104093(1.12)+35=654.7655232276585, reported as $654.77. This calculation is included to verify arithmetic alone; the illustrative credibility weight does not claim empirical derivation.

B. Exposure-weighted and conditional estimands

A portfolio-level pure premium is sometimes written as the product of an overall average claim frequency and an overall average severity. This representation is valid under consistent aggregate definitions:

E[L]E=E[N]EE[L]E[N].\frac{\mathbb{E}[L]}{E} = \frac{\mathbb{E}[N]}{E} \frac{\mathbb{E}[L]}{\mathbb{E}[N]}.

If the portfolio is divided into groups xx, however, group cost is generally π(x)=f(x)s(x)\pi(x)=f(x)s(x) and portfolio cost is ∑xwxπ(x)\sum_x w_x\pi(x), with wx=E(x)/∑yE(y)w_x=E(x)/\sum_y E(y). This expression does not equal the product of unweighted averages of group frequencies and severities. Even when group count and cost distributions are estimated perfectly, an unweighted aggregation creates a composition error whenever class exposures differ.

With stochastic future volume, a pricing model may be required to estimate premium per expected earned exposure, total expected claim cost, or the distribution of an experience-rated aggregate loss ratio. Their estimands are distinct. Under uncertain future volume ETE_T, E[LT]/E[ET]\mathbb{E}[L_T]/\mathbb{E}[E_T] need not equal E[LT/ET∣ET>0]\mathbb{E}[L_T/E_T\mid E_T>0], because low-exposure states can receive very different weights. A prospective portfolio assumption should specify whether exposure is deterministic, forecast as a distribution or conditioned on observable underwriting decisions.

The exposure-weighted treatment also explains why a risk-cost study cannot be fully described by a table of aggregate annual loss ratios. A prospective price is for a particular expected mix and coverage base. Ratemaking may appropriately use separate projections of within-class risk levels and future class volume, with the results recombined at the end. This preserves the interpretability of observed changes.

C. Claim development, trend and the order of adjustments

For simple scalar adjustments on a homogeneous cohort, development and trend commute algebraically: LDT=LTDLDT=LTD. Operationally, however, the underlying quantities may not be independent scalar factors. Development patterns can vary with claim severity, settlement timing, coverage layers and inflation, while trend may act on payments made in different calendar years. A payment-level inflation model and a claim-count development model need not be equivalent to a single aggregate multiplier. The study must therefore identify the historical claim measure, the component being projected and the time at which each change occurs.

The same caution applies to deductibles and limits. In general, E[min⁡{(aX−d)+,ℓ}]\mathbb{E}[\min\{(aX-d)_+,\ell\}] cannot be reconstructed by multiplying an already limited historical insurer payment by a general severity trend aa. Changes in ground-up loss level can shift the proportion of claims piercing a deductible or exhausting a limit. An actuary can model ground-up severity then apply current payment terms, or estimate appropriate limits adjustments under a documented model. The method chosen matters even when the historical average limited loss is known.

Discounting introduces a further dimension. When claim payments occur over multiple future dates, present value depends on both amounts and the assumed time pattern, whereas undiscounted ultimate loss cost is based on eventual nominal payments. A cash-flow analysis may need an explicit allocation by payment date and the applicable discount or investment treatment, consistent with the revised ASOP No. 20. Discounting should not be inserted into an otherwise undiscounted pure premium equation without explaining the matching premium and capital assumptions.

D. Minimum reproducibility and challenge record

A usable record of an empirical prospective loss cost analysis includes the following data and decision elements, insofar as permitted by confidentiality rules:

  1. Defined future insured events, policy payment functions, coverage conditions, exposure units and prospective period.
  2. Historic exposure and claim cohorts with earned and accident or policy-period keys, valuation date, gross/net basis, currencies and claim expense conventions.
  3. Reconciliation of observed paid, case, incurred and ultimate amounts; selection and rationale for development and tail factors.
  4. Trend indices or models, projection periods, expected cost dates, selected future scenarios and evidence for assumptions.
  5. Large-loss or catastrophe exclusions and the independent calculation of corresponding restoration allowances.
  6. Expected mix by class, territory, limit and deductible, with a comparable historical standardization or covariate model.
  7. Source and model provenance, reproducible transformations, out-of-time validation and identification of material model uncertainty.
  8. A reconciliation from prospective indemnity to total claims-related cost, other expenses, profit, capital and final rate indication.
  9. Sensitivities, known limitations, assumptions requiring external verification, and an implementation audit trail.

This checklist describes the evidence necessary to reproduce and challenge a rating indication. It is an editorial aid to established actuarial practice rather than a new rating framework. The appropriate depth and documentation of any particular professional assignment depend on its scope, law, materiality and the standards applicable to the actuary.

The illustrative synthetic computations in this dossier are supplied at a sufficient level of detail to be recalculated independently without access to proprietary claims or policy records. They establish the arithmetic pathway. Their assumptions deliberately omit actual claim observation, fit diagnostics, statutory filing analysis and market validation. Those tasks would be indispensable before applying the numbers to an actual insurer or insured population.

References