Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Actuarial Experience Studies and Assumption Setting

Abstract. Actuarial experience studies convert observations about insured risks, benefit payments, persistency, expenses, and other contract behavior into evidence about future contingent cash flows. Their scientific object is a defined population under observation; their practical output is a defensible set of prospective assumptions for a specified financial decision. This monograph develops the relationship between those two objects through exposure measurement, event definitions, observation windows, actual-to-expected diagnostics, heterogeneous risk classes, credibility, trend projections, uncertainty, and independent review. The analysis covers life and annuity mortality, morbidity, policyholder lapse behavior, property and casualty claims frequency and severity, claims development, expenses, and pension obligations. It derives exposure-weighted rates and A/E ratios, explains why their apparent simplicity can conceal selection and denominator bias, and tests prospective translations with transparent hypothetical calculations. The study distinguishes experience-based central estimates from contractual specifications, measurement-specific risk adjustments, regulatory prudence, and stress assumptions. International regulatory comparisons address U.S. actuarial standards, EU Solvency II, IFRS 17, and model international practice standards. Contemporary mortality evidence through October 2026 demonstrates that experience emerging from a common historical event can support different conclusions depending on the portfolio, measurement horizon, and credibility of an inferred structural change. The resulting synthesis identifies actuarial assumption setting as an accountable forecasting judgment: observable experience supplies disciplined evidence, while the decision purpose determines the relevant prospective estimand, horizon, and financial effect. Numerical illustrations are hypothetical and independently derived; no proprietary insurance records, newly estimated empirical parameters, or new actuarial theory are claimed.

Keywords: Actuarial experience study; actuarial assumptions; actual-to-expected ratio; exposure to risk; mortality; morbidity; claims frequency; claims severity; persistency; lapse; expense inflation; credibility; selection bias; censoring; claims development; mortality improvement; best estimate; risk adjustment; Solvency II; IFRS 17; assumption governance; prospective validation.

Introduction

A mortality rate is measured from deaths and time at risk; a claims frequency is measured from eligible insured events and units of insured exposure; a lapse rate is measured from contract terminations and the relevant population of continuing contracts. These quantities appear self-contained. In use, however, each represents choices about contracts, recognition, inclusion, timing, and the purpose of the projection. The same observed deaths may imply a materially different expected death claim amount when the distribution of sums insured changes. The same observed claims may imply different ultimate loss costs after correcting for reporting delays, policy limits, inflation, and shifts in the insured business. An experience study must therefore begin with an explicit statement of what experience represents and what forecast must be produced.

The scientific object of this Research Dossier is the inferential movement from historical contract and claim records to prospective actuarial assumptions. The analysis starts before statistical estimation, with the conditions under which an event belongs to a risk set, and ends after estimation, with the conditions under which the selected assumption is appropriate for pricing, valuation, reserving, pension measurement, or risk management. This is a narrower and deeper question than general predictive modeling: it concerns the defensible interpretation of observed insurance experience for a specified economic obligation. The practical distinction between a historical rate and a future assumption is embedded in the scope of the ASOP 53 prospective cost guidance and the Solvency II assumptions requirements.

The account has six connected investigations. The first determines the estimand and unit of study. The second establishes the observation window, exposure denominator, censoring, and data selection. The third examines observed-versus-expected comparison and the resulting statistical diagnostics. The fourth translates these observations into prospective assumptions, including credibility, margins, cash-flow consistency, and contractual effects. The fifth deals with trend and emerging evidence, using the latest available 2026 mortality developments. The sixth defines the evidence trail, challenge process, and conditions for approval. Each conclusion follows from a specified model, source, or independently calculated illustration. The monograph does not assert that one estimation procedure is optimal for every insurance contract or jurisdiction.

The estimand precedes the model

An estimand is the quantity the decision requires. A life insurer may require a one-year conditional death probability for insured lives with particular underwriting status; an annuity provider requires survival probabilities across a long horizon; a motor insurer requires projected ultimate claim cost for the coming policy year; a pension actuary requires a sequence of decrement and economic assumptions for a benefit measurement date. The distinction matters because assumptions that appear interchangeable at the level of aggregate historical experience often have opposite financial consequences. Mortality deterioration typically increases the expected cost of death benefits but may decrease the expected duration of life-contingent annuity payments. A higher cancellation rate may change near-term fee income, future claims exposure, acquisition expense recovery, and reinsurance recoveries in competing directions.

For this reason, an experience study should declare, for every output, the population, event, time origin, measurement horizon, covariates, exposure unit, weighting convention, valuation purpose, and standard of uncertainty. The study should also state the geographic and legal setting: an assumption selected for U.S. statutory life valuation is not automatically transferable to an EU best estimate or an IFRS 17 accounting estimate. The IFRS 17 insurance contracts framework provides a contract cash-flow measurement setting; the international actuarial practice standards index records model professional standards whose binding effect depends on local adoption.

Experience Study Design

A study is a defined measurement system

An actuarial experience study comprises a population rule, event rule, exposure calculation, historical observation period, adjustment protocol, estimation method, comparator, and forecast destination. In life insurance, the population might comprise individually underwritten policies issued in specified years, excluding surrendered or converted policies after the date of termination. In disability insurance, the population and event may differ between incidence, termination, and claim recovery studies. In casualty insurance, the analysis may distinguish policyholder exposure, accident occurrence, report and settlement processes, and incurred losses. These are different statistical experiments even when they draw from the same administration database.

The numerator, denominator, and time reference must describe a coherent risk process. If a claim count is recognized on accident date while the denominator is measured by active policies at year-end, the ratio may mix two incompatible populations. If premium units are used to approximate time at risk, the analyst must determine whether changes in rates, deductible, limit, payroll, insured asset value, or inflation undermine the assumed proportionality. The discipline of ASOP 23 data quality requires the actuary to consider relevant data limitations, sources and the intended use; it supplies a professional check on an apparently complete but substantively misaligned denominator.

A practical study specification can be represented as

S=(P,B,E,T,W,X,L,V),\mathcal S=(P,B,E,T,W,X,L,V),

where PP is the eligible population, BB the contractual benefit or exposure boundary, EE the event definition, TT the observation interval and time axis, WW the exposure or amount-weighting convention, XX the available explanatory characteristics, LL the treatment of lags and missingness, and VV the valuation or decision purpose. This notation is merely a descriptive checklist for the established actuarial study elements; it is not presented as an original theoretical model.

Event identity and competing mechanisms

Death, disability incidence, recovery, surrender, partial withdrawal, paid-up conversion, policy expiration, claim reporting, claim reopening, and expense occurrence should be coded separately. Consider a policy with active coverage from January 1 to August 20 and a paid death claim reported in September: the death event may be counted in the exposure year, while its reporting date belongs to a different operational clock. If cancellation precedes death, the risk interval generally closes at contract termination. If a policy is reinstated, separate intervals may be appropriate. If one episode produces multiple payments, an event-count study should avoid counting each payment as a new incidence event.

A multi-state process expresses these distinctions with a state-dependent transition intensity μij(t)\mu_{ij}(t) from state ii to state jj. For a life-and-lapse illustration, an active policy has transitions to death and lapse; under a simple competing-risks model, the probability of remaining active through time tt is

SA(t)=exp⁡ ⁣[−∫0t{μA,D(u)+μA,L(u)} du].S_A(t)=\exp\!\left[-\int_0^t\{\mu_{A,D}(u)+\mu_{A,L}(u)\}\,du\right].

The cumulative death incidence for active policies is ∫0tSA(u)μA,D(u) du\int_0^t S_A(u)\mu_{A,D}(u)\,du. A crude death-event ratio therefore depends on both death and lapse behavior. Estimating a death intensity from observed deaths and total active exposure is appropriate under its specified risk-set definition; interpreting a crude proportion as a stand-alone death probability in a world with changing lapse rates requires additional assumptions. For income protection, the relevant state transitions may include active employment, disability, recovery, retirement, and death, each with distinct exposure and benefit consequences.

Segmentation and study power

Heterogeneity may arise from attained age, sex where legally permissible and actuarially relevant, smoker status, underwriting class, product, duration since issue, distribution channel, policy size, geography, occupation, indemnity design, deductible, and socioeconomic determinants. Splitting by every available variable yields small and unstable groups; excessive aggregation conceals real gradients. Segmentation must be guided by plausible risk mechanisms, contractual differences, substantive law and predictive value. The historical statistical foundation of generalized linear models by Nelder and Wedderburn (1972) permits an appropriate link function and exposure offset to estimate structured variation, while still requiring validation of model assumptions and transfer to the target portfolio.

Insurance data also contain institutional selection. Underwriting admits some applicants at one price and rejects others. Renewal changes the portfolio through customer choice and insurer action. Claims data are affected by deductibles, reporting incentives, and loss-adjustment practice. A historical experience study consequently observes a selected population under a particular contract, not an unconditional risk distribution. This distinction underlies any credible adjustment from the observed portfolio to future business.

Study populations and counterfactual boundaries

An insurer may observe P(Y∣X,S=1)P(Y\mid X,S=1), where S=1S=1 denotes entry into the insured portfolio, while an expansion strategy requires P(Y∣X,S∗=1)P(Y\mid X,S^*=1) for a different future acceptance rule. If SS depends on unmeasured risk characteristics, reweighting by observed XX alone cannot identify the new portfolio risk. The selection problem has a classical counterpart in Heckman's sample-selection analysis and in the broader literature on observational selection. That source is not relied upon here as an insurance-specific empirical result: it supports the general inference that the selection process changes what can be identified from observed samples.

When a pricing or valuation assumption is intended for new business, the experience-study report must explicitly explain how the historical underwriting and distribution filters compare with the prospective ones. When a valuation covers an in-force block, the correct population is the in-force block as of the measurement date, with appropriate selection and duration effects. A broad industry mortality table may offer a useful external reference without being a substitute for the insurer's specific insured-life mix.

Governance of the study specification

A study design is approved when its outputs can be traced to a particular contract population and decision, and its unresolved gaps are identified. Revision control should cover eligibility rules, event-code mappings, adjustment decisions, aggregation methods, and data extracts. The ASOP 56 modeling guidance frames model use in relation to intended purpose, input appropriateness, testing, and communication. The central methodological finding of this chapter is that model sophistication cannot restore comparability to a study whose measured event and denominator lack a coherent contractual definition.

Data Period and Exposure Selection

Observation windows and the actuarial clock

Historical experience is indexed by several clocks. Calendar time captures macroeconomic and seasonal events. Policy duration captures selection wear-off, lapse behavior and contract changes after issue. Attained age captures biological and demographic aging. Accident time captures the occurrence of an insured loss. Report, paid and valuation dates capture the maturation of information and cash payments. A study cannot exchange one clock for another without a transformation. A mortality estimate aggregated by calendar year may disguise a change in policy durations; a claim cost per earned exposure by underwriting year may differ from the cost emerging in a calendar accounting year.

An observation window should be long enough to produce relevant evidence and short enough to avoid blending incompatible regimes without adjustment. The choice is linked to forecast horizon. For next-year automobile pricing, recent claims costs, underwriting terms and repair inflation carry immediate relevance. For long-term life and pension liabilities, a few recent mortality observations are insufficient by themselves to estimate decades of age-specific improvement. A fixed five-year window is therefore not a universal principle. The sensitivity of results to start and end dates is an empirical diagnostic, and material differences should appear in the review record.

Person-time and policy-time exposure

Let contract ii be observed as eligible during [ai,bi][a_i,b_i] within a study period. Its earned exposure measured in years is

ei=∫t0t1I{i is eligible and at risk at t} dt,E=∑iei.e_i=\int_{t_0}^{t_1}I\{i\text{ is eligible and at risk at }t\}\,dt, \qquad E=\sum_i e_i.

If dd is the number of events satisfying a specified definition and events are reasonably represented by a homogeneous counting-process rate within the cell, the central frequency estimate is λ^=d/E\widehat\lambda=d/E. Its units are events per exposure-year. A simple Poisson working model gives Var⁡(λ^)=λ/E\operatorname{Var}(\widehat\lambda)=\lambda/E and an estimated standard error d/E\sqrt d/E, provided the exposure is measured correctly and the independent-event variance assumptions are adequate. Overdispersion, repeated episodes, calendar clustering, and claim dependence require different uncertainty treatment.

A person insured for nine months contributes 0.750.75 person-year; an insured policy covering two persons may contribute two life-exposures rather than one policy-exposure, depending on the event process. A motor policy covering multiple vehicles similarly requires a vehicle-year definition if the underlying rate is per insured vehicle. For mortality, refined central-exposure techniques allocate time at risk consistently with death and withdrawal conventions. The Human Mortality Database methodology illustrates the role of harmonized deaths, population exposure, mortality rates, and life tables in population analysis, while insured-life studies must also account for underwriting and contract selection.

Consider a hypothetical file with five insured incidents. A naive denominator of 100 full policy-years yields 0.05000.0500 claims per year. Eligibility audit shows only 80 earned policy-years; the corrected rate is $5/80=0.0625$, or 25% higher than the naive rate. Nothing about the five claims changed. The difference is entirely due to the risk denominator. This elementary calculation shows why exposure reconstruction is a substantive scientific check rather than a formatting task.

Amount-weighted and count-weighted experience

A count-weighted mortality study estimates event probability among lives, while an amount-weighted study estimates death claims per unit of insured amount. For policy ii, let wiw_i be face amount or another economically meaningful weight. Define

RN=∑idi∑iei,RW=∑iwidi∑iwiei.R_N=\frac{\sum_i d_i}{\sum_i e_i},\qquad R_W=\frac{\sum_i w_i d_i}{\sum_i w_i e_i}.

The two coincide only under special conditions. Their difference is evidence about the relationship between policy size, exposure, and mortality. A mortality table for benefit-amount valuation should be evaluated against the appropriate economically weighted population. But a large face amount can make an amount-weighted estimator volatile, and a very small number of deaths in high-face-amount policies may dominate it. Report both measures and their uncertainty when each addresses a distinct purpose.

Hypothetical illustration: 900 insured lives each have face amount $100{,}000$ and mortality probability 0.0010.001; 100 lives each have face amount $1{,}000{,}000$ and probability 0.0030.003. Expected deaths equal $0.9+0.3=1.2,or, or0.0012$ per life. Expected death claims are $90{,}000+300{,}000=390{,}000$ on $190{,}000{,}000$ of exposure amount, giving an expected claim-amount rate of approximately 0.002052630.00205263 per unit face amount. This difference is mathematically attributable to the higher share of insured amount in the higher-risk group; it would be obscured by a single count-weighted assumption. The SOA valuation basic tables provide real examples of the need to define insured-life classification and valuation basis before interpreting mortality rates, without validating the hypothetical numbers used here.

Left truncation, right censoring, and incomplete follow-up

A life enters a study when eligible and observed, possibly after contract issue or biological birth. That is delayed entry or left truncation, and the risk set must start on the actual eligibility date. Right censoring occurs when observation ends while a contract remains in force, or when coverage terminates without the studied event. Under appropriate independent censoring assumptions, exposure contributions before censoring remain valid. If censoring depends on unobserved future event risk, simple survival methods may be biased. For lapse and disability studies, competing decrements must be treated explicitly, because death or recovery changes which events remain observable.

The same logic applies to claims severity. A file closed at the extract date may be fully paid; an open file contains an incurred or paid amount that can increase, decrease, or reopen. Using paid-to-date amounts as ultimate severities gives a different quantity from the cost of fully developed claims. Claim reporting and settlement processes should therefore be represented by delay distributions, development factors, case estimates or other methods suited to available data. The foundational Mack (1993) chain-ladder analysis concerns reserve variability under specified development assumptions; it does not make unobserved future claims automatically identifiable from any incomplete triangle.

Missingness and data lineage

Data missingness is not a unitary phenomenon. An unavailable age may be nearly random within a homogeneous import batch, while an unavailable diagnosis may follow from provider practice, claim severity, or coverage restrictions. Complete-case analysis can change the population under study. Rubin's inference and missing-data framework (1976) formalizes why inferential validity depends on how missingness relates to observed and unobserved variables. In operational experience studies, every critical field should have a completeness rate by source, cohort, and time; reason codes should distinguish genuine absence, rejected records, and nonapplicability. Imputation may be justified under a defined model and should be accompanied by a sensitivity analysis against plausible alternative mechanisms.

Source lineage is central to reproducibility. The controlled input should identify the originating policy, claim, finance and reinsurance systems; transformations between event and snapshot records; record counts and monetary totals before and after every material filter; currency conversions; policy endorsements; duplicate handling; and reconciliation to authoritative ledgers. A study can reconcile in total and still misclassify experience by product or segment. Both aggregate and subgroup reconciliations are required where material.

Historical representativeness and transport

A historical block is often unlike its future successor. Changes in underwriting appetite, policy limits, medical screening, distribution, regional composition, claim management, medical care, inflation, or cancellation behavior alter the data-generating process. If the target distribution of observed covariates changes from p0(x)p_0(x) to p1(x)p_1(x), a reweighted estimator for an observable outcome may use w(x)=p1(x)/p0(x)w(x)=p_1(x)/p_0(x) where both densities are positive and the conditional outcome mechanism is stable. The assumptions of positivity and conditional transport must be stated. When new target risks lack historical analogues, simple reweighting is unjustified and external evidence or explicit judgment is necessary.

An illustrative portfolio has 900 low-risk lives with annual event probability 0.050.05 and 100 high-risk lives with probability 0.300.30. Its population expected count is $900(0.05)+100(0.30)=75,arateof7.5, a rate of 7.5%. If the extracted sample includes only 90 of the low-risk lives and all 100 high-risk lives, its expected rate is[90(0.05)+100(0.30)]/190=18.1579%$. The sample does not demonstrate that the entire portfolio's risk has risen by that amount. It demonstrates a change in mixture weights. Selection adjustment must be based on a defended relationship between the sampling process and observed risk class; where information is insufficient, the assumption uncertainty remains explicit.

Calendar alignment, currency and financial basis

Expense and claims costs must be expressed on an economically coherent price basis. Medical costs, wage indices, building-material prices, and generic consumer inflation may have materially different dynamics. Earned-premium denominators may require rate-level adjustments before use as a frequency or cost base. Historical foreign-currency claims require a decision about exchange-rate dates and prospective currency exposure. Reinsurance attachment, coinsurance, deductible and policy-limit changes affect insurer net losses and should not be confused with a biological mortality or physical hazard change. A study of total ground-up losses answers a different question from a study of retained insurer payments.

The chapter's principal inference is that time and exposure selection construct the empirical population to which a historical rate refers. Every later A/E ratio, credibility coefficient, and trend projection inherits that construction. Changing the denominator or inclusion window changes the scientific claim even if the displayed model formula remains identical.

Actual versus Expected Experience

Definition, decomposition, and interpretation

An actual-to-expected ratio compares an observed amount or event count with the amount predicted by a clearly specified benchmark. Let OgO_g denote observed events in group gg, EgE_g their expected count under benchmark assumptions and realized exposure, and rg=Og/Egr_g=O_g/E_g when Eg>0E_g>0. The aggregate ratio is

R=∑gOg∑gEg=∑gEgrg∑gEg.R=\frac{\sum_g O_g}{\sum_g E_g} =\frac{\sum_g E_g r_g}{\sum_g E_g}.

This weighted-average identity matters. An aggregate A/E ratio cannot be interpreted without the expected-event weights. A shift in the age, duration, region, or product mix changes those weights and may change the aggregate ratio even if the within-cell ratios remain constant. Nor is R=1R=1 proof that every subgroup is well calibrated: adverse deviations in one class may offset favorable deviations in another. The correct unit of comparison may be count, insured amount, claim cost, policy-year exposure, present value, or another target quantity, but the numerator and denominator must always share it.

For life mortality, a conditional probability benchmark qx,dq_{x,d} can give E=∑iqxi,diE=\sum_i q_{x_i,d_i} when each policy contributes a full specified risk interval. For fractional intervals or competing terminations, expected events should instead be calculated under the same event-time process used for the observed risk set. If a counting-process intensity λi(t)\lambda_i(t) is employed, an appropriate expected count is E=∑i∫aibiλi(t)dtE=\sum_i\int_{a_i}^{b_i}\lambda_i(t)dt. The benchmark can reflect underwriting selection, duration, attained age and calendar period; where it does not, a large A/E ratio may measure benchmark mismatch rather than an unexpected biological change.

A/E illustration with changing mixture

Assume a hypothetical two-group study with expected and observed deaths as follows:

Group Expected deaths Observed deaths A/E
A 80 88 110.0%
B 20 16 80.0%
Total 100 104 104.0%

The aggregate 104% is $0.8(110%)+0.2(80%). If the target future portfolio instead assigns half of expected deaths to each group, standardizing the cell ratios to that mix gives \0.5(110%)+0.5(80%)=95%$. The historical aggregate factor of 104% and the future-mix factor of 95% are each arithmetically correct for different expected-event weights. The difference is a portfolio-composition effect, not a contradiction. This example demonstrates why a raw aggregate ratio should not mechanically become a future assumption without target-portfolio standardization.

Precision: counts are not always independent

Under an independently distributed Poisson count model with known expected count EE, Var⁡(O)=E\operatorname{Var}(O)=E under perfect calibration and the approximate standard error of R=O/ER=O/E is $1/\sqrt E.Moregenerally,anempiricalstandarderrorcanuse. More generally, an empirical standard error can use\sqrt O/E$ if the Poisson approximation is adopted. With E=100E=100, this is approximately 10 percentage points; with E=10,000E=10{,}000, it is about one percentage point. But correlated catastrophe losses, repeated morbidity episodes, household clustering, high face amounts and temporal shocks can invalidate independent-event precision. Reporting a Poisson interval for a dependent portfolio would create false certainty.

For claims frequencies, an overdispersed Poisson working model may use Var⁡(O)=ϕE\operatorname{Var}(O)=\phi E where ϕ>1\phi>1 is estimated from suitable residual diagnostics. A negative-binomial specification can accommodate additional heterogeneity under its variance structure. For severity outcomes, the claim size distribution, truncation by deductibles and limits, and the tail affect uncertainty; the number of claims alone is insufficient to quantify it. A company with a small observed death count but a highly concentrated benefit amount may have materially different credibility for count-weighted and amount-weighted ratios. The credibility procedures standard calls for appropriate relevant experience and attention to heterogeneity, not a single universal exposure threshold.

Credibility as an explicit estimation compromise

Let Yˉ\bar Y be a mean outcome for a particular risk class and mm a comparable reference estimate. A linear credibility estimate takes the established form

θ^=ZYˉ+(1−Z)m,0≤Z≤1.\widehat\theta=Z\bar Y+(1-Z)m,\qquad 0\le Z\le 1.

The coefficient ZZ expresses the evidentiary weight of subject experience. In a simple Bühlmann setting with nn observations, expected process variance EPV\mathrm{EPV} and variance of hypothetical means VHM\mathrm{VHM}, the classical credibility factor is

Z=nn+K,K=EPVVHM.Z=\frac{n}{n+K},\qquad K=\frac{\mathrm{EPV}}{\mathrm{VHM}}.

The formula follows by minimizing mean squared error under the stated random-effects assumptions. It does not prescribe ZZ independently of the heterogeneity and process noise of the portfolio. The underlying tradition is documented in Bühlmann's credibility research (1967) and his later experience-rating account (1969). In operational use, the relevant reference population must be compatible with the subject portfolio or adjusted with credible evidence.

Suppose a hypothetical block has an observed mortality A/E factor of 1.201.20, a suitable external reference of 1.001.00, and estimated credibility Z=0.30Z=0.30. The combined factor is $0.30(1.20)+0.70(1.00)=1.06.Ifavalidatedalternativestudysupports. If a validated alternative study supportsZ=0.60,thecombinedfactorbecomes, the combined factor becomes1.12.Theresultchangesbecausetherelativeevidentiaryweightschange,notbecauseonesetofobserveddeathschanged.Theanalystshoulddocumentthesourceanduncertaintyof. The result changes because the relative evidentiary weights change, not because one set of observed deaths changed. The analyst should document the source and uncertainty ofZ$ and the compatibility of the reference experience. A mismatch between subject and reference cannot be solved by choosing a cosmetically smooth factor.

A/E calibration is multidimensional

For a model with risk cells, inspect ratios by attained age, issue age, duration, sex where appropriate and permissible, product, underwriting class, policy size, territory, calendar time and claim amount. Report both unweighted aggregates and relevant economic weights. Examine residual patterns, time changes and outliers. If one segment materially dominates expected claims, its calibration may control the liability even when it is a small fraction of contracts. If a pricing model is fitted with an offset log⁡ei\log e_i for exposure, a Poisson GLM can write

log⁡E[Ni∣Xi,ei]=log⁡ei+β0+Xi⊤β,\log \mathbb E[N_i\mid X_i,e_i] =\log e_i+\beta_0+X_i^\top\beta,

where NiN_i is the event count. The model includes exposure rather than treating unequal observation time as interchangeable records. Link and variance assumptions should be supported by out-of-sample residual behavior and the intended estimand. Nelder and Wedderburn's formulation supplies a general estimation structure; insurance-specific variables, admissibility and pricing implementation still require an independent actuarial decision.

Claim cost A/E and development bias

For non-life losses, observed claims at valuation date tt usually include only part of ultimate cost for the relevant accident years. If UaU_a is the ultimate cost for accident year aa and Ca,tC_{a,t} is the current cumulative amount, a working development factor fa,tf_{a,t} estimates U^a=fa,tCa,t\widehat U_a=f_{a,t}C_{a,t} under its assumptions. A naive ratio of current paid claims to prospectively expected ultimate claims understates recent accident-year loss cost if immature claims are systematically incomplete. Conversely, a calendar-year cash-flow A/E may deliberately use paid claims because actual payment timing is its object. The discrepancy is resolved by specifying whether the study concerns incurred ultimate losses, paid cash, reported counts or unpaid obligations. The ASOP 43 unpaid-claim guidance concerns methods and communication appropriate to unpaid claim estimates in the U.S. context.

Hypothetical illustration: accident year A has 100 ultimate claims but only 60 recognized by the observation date on 100 exposure units. Accident year B has 100 ultimate claims but 80 recognized on 120 exposures. The observed ratios are 0.600.60 and approximately 0.66670.6667 per exposure, suggesting an increase. Corrected ultimate ratios are 1.001.00 and approximately 0.83330.8333, implying a decrease. The apparent reversal arises entirely from reporting maturity. It is a reason to validate the development process before interpreting calendar changes as underwriting deterioration.

Attribution and experience decomposition

A change in projected claim cost can be decomposed into components. Let NN be exposure units, ff frequency per unit, and ss expected severity. Expected aggregate cost is C=NfsC=Nfs. For two periods,

C1C0=N1N0⋅f1f0⋅s1s0,\frac{C_1}{C_0}= \frac{N_1}{N_0}\cdot\frac{f_1}{f_0}\cdot\frac{s_1}{s_0},

provided the factors use consistent, comparable definitions. For a heterogeneous portfolio, such simple multiplicative decomposition must first control for changing segment mix, limits and inflation. An analogous life-claim decomposition separates insured amount, mortality probabilities and covariate distribution. This calculation is useful for explaining whether a liability change reflects more business, altered risk, altered payment amounts, or the assumptions used to measure them. Where interactions are material, an explicit standardization or counterfactual decomposition is preferable to assigning all changes to the last modified variable.

Benchmarks should be frozen and interpretable

The comparator for an A/E test must be the expected basis that was valid for the relevant measurement, or a separately labeled updated comparator. Recalculating the denominator with newly fitted assumptions and presenting the resulting A/E as independent validation creates circularity. A credible review maintains a versioned baseline and shows what changed between old and new assumptions. For a sufficiently material study, compare historical data against a frozen baseline, a challenger model, and external evidence, recording differences in units and time basis. This is an operational consequence of measuring prediction quality, not an independent actuarial doctrine.

Assumption Selection and Margins

From a historical estimate to a prospective distribution

An actuarial assumption is selected for a specified future period, population, contractual state and measurement purpose. If θ^0\widehat\theta_{0} denotes a historically estimated parameter, a prospective assumption θt∗\theta_{t}^{*} may require a documented transformation of portfolio composition, contractual benefits, operational process, medical or economic trend, and model uncertainty. The essential scientific question is whether each adjustment changes the expected outcome under the future conditions, rather than merely improving fit to an historical dataset.

For a group gg with mortality benchmark qg,tbaseq_{g,t}^{\mathrm{base}}, a working assumption may take the form

qg,t∗=qg,tbase rg sg,t mg,t,q_{g,t}^{*}=q_{g,t}^{\mathrm{base}}\,r_g\,s_{g,t}\,m_{g,t},

where rgr_g is a risk-level adjustment derived from standardized historical A/E and credibility, sg,ts_{g,t} is a prospective trend factor, and mg,tm_{g,t} captures a separately documented portfolio or contractual transformation. Multiplicative combination is appropriate only where the resulting probability remains in [0,1][0,1] and the conceptual factors represent compatible conditional risks; another link function or hazard-rate specification may be preferable. This expression describes a widely recognizable construction from established actuarial elements, not a newly proposed assumption-setting theory. The analyst must avoid counting the same improvement in both ss and mm, or using a base table whose implicit selection effect duplicates rr.

The expected future benefit cash flow from a contract can be written abstractly as E[Ct(θ∗)]\mathbb E[C_t(\theta^*)]. For a simple one-year term death benefit BB with survival-conditioned death probability qq, expected benefit is BqBq. For a level annuity paying BB at the end of each year while alive, with survival probabilities tpx{}_tp_x and discount factors vtv_t, expected present value is

EPV⁡annuity=B∑t=1Tvt tpx.\operatorname{EPV}_{\mathrm{annuity}}= B\sum_{t=1}^{T}v_t\,{}_tp_x.

Changing mortality shifts both types of claim values through different functions. The translation from mortality evidence to financial effect therefore requires the contract benefit structure and a consistent discount basis, rather than a generic assertion that higher mortality always increases insurance liabilities.

Central estimate, adjustment and margin are distinct decisions

A central or best-estimate assumption ordinarily represents an assessment of expected experience conditional on the measurement framework. A margin or risk adjustment represents additional treatment of uncertainty, prudence, or nonfinancial risk according to that framework. A stress scenario asks what financial outcomes arise under specified adverse conditions. These are different operations. If a model embeds a conservative mortality rate, a separate risk adjustment for the same uncertainty may duplicate protection; if a valuation requires an explicit risk adjustment, embedding unknown implicit margins in each assumption impairs interpretation and comparability.

The IFRS 17 measurement framework includes estimates of future cash flows and an explicit adjustment for nonfinancial risk within fulfilment cash flows, subject to its measurement model and other requirements. The Solvency II Article 77 rule addresses probability-weighted best estimates and a separately determined risk margin under the EU prudential framework. Neither regime authorizes the analyst to treat an unexplained historical loss loading as a substitute for the designated measurement component. The term margin must always be defined by purpose, valuation basis, and jurisdiction.

A useful conceptual separation in an actuarial workpaper is the baseline prospective expected value, explicit and documented experience adjustments, model and parameter uncertainty, prescribed prudence or risk adjustment, and decision-specific stress tests. These can be reconciled to the reported figure while preserving their different meanings. For pricing, commercial loadings for expenses, capital, profit, tax and distribution may be relevant, but these are not necessarily the same as best-estimate risk parameters. The ASOP 54 life pricing guidance and ASOP 53 prospective cost guidance provide purpose-specific professional contexts for these distinctions.

Credible selection among conflicting sources

Suppose an insurer's own experience suggests a 12% deterioration against an external benchmark, while a comparable pooled industry study suggests only 2%. It is tempting to average the factors, but a transparent decision asks: Are the event definitions identical? Does one population include recently underwritten lives and the other older policies? Are observation periods equally recent? Are claims incurred or paid? Does one source cover a different jurisdiction or benefit type? Are the apparent differences statistically material? Only after these questions can a credibility weight be meaningfully assigned.

Where reference rates are available, a study should report baseline table selection, table construction or adjustment, reason for using internal versus external data, amount or count weights, the period to which each rate applies, and the specific risk classes where judgment changes empirical results. Some assumptions may be substantially prescribed. A legal or contractual prescription is an attribute of the valuation rule, not empirical confirmation of the prescribed number. The distinction between prescribed and actuary-selected pension assumptions is treated in the current ASOP 27 pension assumption standard.

Non-life frequency, severity, inflation and limits

For prospective property/casualty pricing, the underlying expected loss cost per exposure can be approximated as λtE[Xt]\lambda_t\mathbb E[X_t] under a frequency-severity construction. Historical severities should be made comparable for accident year, payment year, price level, benefit limits, deductibles, legal environment and claims practice. Trend assumptions can differ between frequency and severity. If claim frequency declines by 3% while average severity rises by 8%, their combined cost factor is $0.97\times1.08=1.0476$, an increase of 4.76%, before exposure, coverage and mix changes. This follows from multiplication and does not require a new empirical claim about any actual insurance market.

A fixed monetary deductible reduces reported indemnity in a way that depends on the severity distribution. A proportional inflation adjustment cannot generally reproduce the effect of a fixed deductible or limit if the whole claim distribution changes. For example, if ground-up severity XX rises with inflation factor kk and a fixed deductible is dd, insurer payment becomes (kX−d)+(kX-d)_+, not k(X−d)+k(X-d)_+ in general. Consequently, an actuarial on-level study must consider the contract payment function, rather than applying one cost index mechanically to every coverage structure.

Suppose a hypothetical claim cost is $10{,}000$ at baseline. A selected annual claim-cost inflation rate of 4% for three years projects $10{,}000(1.04)^3=$ $11{,}248.64.A2. A 2% assumption gives \10{,}612.08,adifferenceof, a difference of636.56$ per claim and $127{,}312$ across 200 otherwise identical claims. The arithmetic is transparent; the selection of 4% or 2% must be supported by the relevant claim-cost evidence, benefit design and horizon. If these hypothetical costs are already expressed at a projected payment-date price level, adding another calendar inflation factor would double-count trend.

Life mortality, longevity and product asymmetry

Life mortality assumptions often separate base-table level, risk-class effects, duration effects and calendar mortality improvement. An age-year mortality surface can be expressed using a forecast mortality index in the tradition of Lee and Carter (1992), though that original formulation alone does not resolve insured-life selection, changing underwriting, pandemic regimes or parameter uncertainty. A constant annual improvement rate ii applied to a base probability has the simplifying form qx,t=qx,0(1−i)tq_{x,t}=q_{x,0}(1-i)^t under defined age and cohort conventions. It is not an identity of biological mortality; the improvement path may change by age, calendar year and cohort and must be checked against available evidence.

For a death-benefit portfolio, lower expected mortality can reduce expected near-term death claims; for an annuity portfolio, lower mortality can increase survival-contingent payment duration. For term life insurance, a change in lapse may reduce exposure to claims but also change the insured mix through selective persistency. For annuities, mortality improvement interacts with benefit guarantees and the distribution of starting ages. The relevant prospective mortality assumptions may therefore differ by product and function even when the same population dataset informs them.

United States insured-life assumptions also require distinguishing tables for general-population mortality, fully underwritten insured lives, simplified issue, guaranteed issue, and specific policy classes. The SOA 2015 Valuation Basic materials establish an insured-life valuation evidence source, while subsequent mortality improvement developments must be separately assessed. Importing general-population death rates into an insured-life valuation without explicit mapping of underwriting and demographic composition creates a transport error.

Morbidity, disability and recovery

Morbidity is not one event. Incidence of a condition, insurance claim incidence, hospitalization, disability onset, recovery, recurrence, severity and claim duration can each demand different assumptions. A disability income valuation requires at least an incidence process and a termination or recovery process, conditioned as needed on benefit duration and claimant characteristics. Medical expense assumptions require claim frequency, unit cost, utilization, benefits, provider practice and inflation; their product often matters more than either component separately. Long-term care experience adds benefit triggers, policyholder utilization, lapse, mortality and care-setting transitions.

A portfolio with observed claim incidence of 5% does not necessarily have a 5% expected cost ratio: claim amounts, duration and benefits matter. For a stylized incidence-duration model with annual incidence λ\lambda per life and expected discounted benefit duration DD, expected annualized benefit outgo per life is approximately λBD\lambda B D under suitable stationarity and rare-event assumptions. If duration risk is strongly affected by age or time since onset, direct multi-state or survival modeling is preferable. Trend assumptions should separately articulate incidence, duration, utilization and unit cost whenever material.

Lapse, surrender, conversion and dynamic behavior

Persistency experience is often interpreted through annual conditional termination probabilities. Yet lapse is driven by contract value, premium affordability, distribution, lapse-support features, interest rates, surrender charges, tax rules and policyholder heterogeneity. Surrender behavior can change materially when market rates or alternative savings opportunities change. A constant historical lapse factor may therefore be unrepresentative under a different future rate environment. Contractual options and guarantees must be modeled together with plausible behavioral responses, with explicit attention to anti-selection and the purpose of the cash-flow projection.

A simple expected in-force count after one year is N(1−qdeath)(1−qlapse)N(1-q_{death})(1-q_{lapse}) only when the stated discretization and conditional probabilities justify the sequential product. When death and lapse are competing intensities over time, the appropriate survival expression is the integrated intensity from the study design chapter. The choice is a modeling assumption with measurable financial effects. If a surrender event extinguishes future death cover but triggers a surrender payment, the amount of future liability depends on both the reduction in risk exposure and the settlement value. This is why lapse studies require policy-level benefit reconciliation rather than event rates alone.

Expenses and operational assumptions

Actuarial expense studies distinguish acquisition costs, maintenance costs, claim handling, investment expenses where relevant, fixed versus variable overhead, and costs by transaction, policy, claim, premium or insured amount. A unit expense selected from a historical ratio may become obsolete after a distribution change, system migration, restructuring or inflation shock. For example, if total maintenance cost is $1{,}000{,}000$ for 100,000 policies, a naive allocation produces 1010 per policy. If half the cost is fixed and policy count rises to 150,000 while variable cost per policy remains 55, total expected maintenance cost becomes $500{,}000+150{,}000(5)=$ $1{,}250{,}000,orapproximately, or approximately8.33$ per policy; mechanically carrying forward 1010 would overstate the scalable component. The conclusion depends on the stated fixed-variable decomposition, whose accuracy must be established from operational records.

Expense assumptions should reconcile to finance ledgers, use consistent capitalization and expense classification, and reflect the allocation basis required by the destination valuation. A change in expense accounting policy can create an apparent experience trend with no change in resource consumption. Currency, wage inflation, technology costs and distribution commissions may have distinct forecast drivers.

Pensions and the current U.S. assumption standard

Pension measurements link demographic assumptions—mortality, retirement, disability, withdrawal, optional forms of benefit, and member behavior—to economic assumptions such as discount rate, salary growth and inflation. Selection considers the measurement objective, plan characteristics, data credibility, relevant economic and demographic evidence, and the combined effect of assumptions. The current U.S. professional reference for this selection is the revised ASOP No. 27, adopted in December 2023 with its specified January 1, 2025 effective conditions.

An important historical correction is that the formerly separate demographic standard ASOP No. 35 was formally repealed for work from 2025. The ASB's June 2024 repeal announcement confirms the integration into ASOP No. 27 for the stated work scope. Citing the old ASOP No. 35 as a currently applicable stand-alone assumption standard for an in-scope 2026 pension report would therefore misstate the professional standard. The change also illustrates the need for version control over regulatory and professional source material, not merely actuarial data.

Comparison of measurement frameworks

Framework Purpose and jurisdiction Relevant assumption discipline Interpretation boundary
U.S. ASOP Nos. 23, 25, 27, 41, 52–54 and 56 Professional actuarial services within specified U.S. scopes Data quality, credibility, pension assumption selection, communications, life or P/C methods and modeling Effective dates and actual standard scope matter; these are professional standards, not interchangeable accounting bases
EU Solvency II Prudential valuation of insurance and reinsurance obligations Probability-weighted best estimates, data quality, realistic assumptions, risk margin and documentation EU insurance supervisory regime, subject to relevant consolidated law and implementation
IFRS 17 Accounting measurement for insurance contracts within its adopted reporting scope Current fulfilment cash-flow estimates, explicit nonfinancial risk adjustment and contract-specific accounting components Financial-reporting objective distinct from solvency capital and price setting
IAA ISAP model standards International professional model texts General actuarial practice and specialized topics subject to local applicability Model standards acquire authoritative local status through relevant adoption or incorporation

For the EU, Solvency II Article 82 addresses data quality for technical provisions, while Delegated Regulation Article 22 requires realistic assumptions that can be justified, with circumstances of falsification identifiable and consistent treatment by homogeneous risk group. That is a stronger and more specific proposition than the generic instruction to use a recent dataset. For a financial reporting valuation, the relevant IFRS 17 contract measurement requires considering future cash flows within the accounting standard's contract boundary and measurement structure. A technically accurate mortality table may still be inappropriate if applied across an inconsistent cash-flow boundary or with duplicate margins.

The integrated finding is that assumption selection is not one numerical update. It is a traceable translation of appropriately adjusted experience into a purpose-specific future distribution and its financial consequences, including the contractual and regulatory constraints that determine which quantities belong in the final measurement.

Trend and Emerging Evidence

Historical trend and structural change

An experience trend is an inference about the evolution of the underlying risk process, conditional on the observability of that process. It may reflect gradual exposure change, age composition, medical progress, economic inflation, legal or claims-handling changes, climate conditions, or an abrupt event such as a pandemic. Simple extrapolation of the latest A/E ratio assumes a stable relationship between the observed deviation and future risk; that assumption is an empirical hypothesis requiring justification. A trend analysis should separate level changes, persistent slope changes, temporary shocks, changes in observation or reporting, and selection effects.

For a stationary historical annual rate λ0\lambda_0, an assumed constant proportional trend gg produces λt=λ0(1+g)t\lambda_t=\lambda_0(1+g)^t. This identity is an explicit projection convention. It does not prove that gg will persist, nor that frequency and severity share the same trend. Different horizons may require different transition paths, and beyond the data-supported horizon an assumption may rely increasingly on external evidence and documented expert judgment. The risk is especially acute for long-duration insurance and pension liabilities where small annual rate differences compound over decades.

An analyst can compare a pre-event baseline, event-period level, post-event recovery, and prospective structural scenario without pretending that all periods are realizations of the same stationary distribution. Before interpreting a break as biological or economic, confirm that the data extract, age standardization, exposure denominator, claim definitions and reporting lags remain comparable. The contribution of each new observation should be evaluated against its statistical precision and relevance to the target portfolio, rather than its recency alone.

Latest U.S. mortality improvement evidence through October 2026

The temporal distinction between a historical analysis and a future assumption is directly visible in the 2026 U.S. mortality evidence. The Society of Actuaries Research Institute's RPEC 2026 mortality update (October 22, 2026) reports an age-standardized A/E ratio of 98.7% for ages 65 and older during July 2025–June 2026 when expected deaths are based on projecting 2019 population mortality with MP-2021. The same report states that a new mortality improvement projection scale will not be released in 2026 and continues to regard MP-2021 as a reasonable best-estimate scale for relevant retirement plan valuations, particularly at older ages. The latest result is a measured comparison against a specified historical benchmark, not evidence that an actuarial assumption should automatically be scaled by 0.987 across all ages or insurance products.

The earlier RPEC 2025 mortality assessment also reported insufficient post-pandemic evidence to replace its projection scale. Read together, these reports show that improving recent observed mortality can coexist with caution about changing a long-term trend parameter. The 2026 report also notes dependence of the findings on reporting adjustments, population estimates and age standardization, with some materially different results at younger ages. These features belong in the source-to-assumption bridge. The actuarial inference is that an A/E ratio is a validation diagnostic for a particular population, observation period and benchmark; it does not uniquely identify a decades-long improvement slope.

The insured-life use case supplies an instructive contrast. The SOA's September 2026 insured-life improvement scale provides updated historical and future mortality improvement assumptions for AG38/VM20 valuation, using insured-life evidence for much of the adult age range and different treatments where credibility is limited. The earlier 2025 insured-life scale used a different historical observation period. These are prescribed-application research outputs for a U.S. valuation context, with distinct insured-life populations, data windows and regulatory use. The difference from the RPEC retirement-plan position should not be read as a scientific conflict. Different studied populations and destinations can rationally support different assumption updates.

A further observational reference is the SOA's 2025 mortality improvement survey, published October 15, 2026 by Kara Clark. Its reported survey responses show substantial variation in prospective improvement assumptions by product, company, business function, age and horizon and no uniform post-pandemic industry direction. A survey of practices documents what respondents selected; it does not prove that their selections are unbiased forecasts or quantify predictive accuracy. This distinction is important when a practitioner considers a market consensus as a substitute for an independently justified actuarial estimate.

Three distinct questions raised by the same evidence

The 2026 mortality sources permit at least three analytically separate questions. First, are current deaths above or below a fixed benchmark for a particular population? This is an A/E measurement. Second, do the available data support a lasting change in the annual mortality improvement process? This is a trend-identification problem with sensitivity to time windows, data reporting and cohort behavior. Third, does the relevant professional or regulatory valuation basis adopt, permit or require a particular mortality scale at the measurement date? This is a question about applicable valuation rules and effective dates. A defensible assumption decision answers each one rather than moving directly from the first to the third.

For illustration, consider a hypothetical base mortality of 1% and two long-term improvement rates: 0.5% and 1.0% per year. After 20 projected years, the corresponding probabilities are $0.01(0.995)^{20}\approx0.009046$ and $0.01(0.99)^{20}\approx0.008179$. The latter is about 9.6% below the former on the same starting mortality level. This sizable difference results from different compounding rates over a long horizon even though the annual rates differ by only half a percentage point. It explains why a modest change to a long-term improvement assumption can matter financially without implying that the evidence automatically supports either rate.

Trend attribution in casualty and health experience

Historical casualty claims can rise because risk events become more frequent, individual event costs rise, insured exposure grows, underwriting mix changes, policy terms change or reporting becomes faster. Health insurance costs can increase because utilization rises, price per service changes, covered treatments change, populations age or provider networks shift. A historical aggregate pure premium is unable by itself to distinguish these mechanisms. A transparent study decomposes measured cost where the data allow and identifies components that remain confounded. For example, a 10% increase in average paid claim cost could be a real change in severity, faster settlement of expensive claims, changing deductibles, or a combination; the trend factor selected for future ultimate losses should reflect the appropriate target mechanism.

A fixed recovery or development pattern is especially vulnerable to operational changes. If the insurer speeds up claim payments, a calendar paid-loss triangle may exhibit faster development even when ultimate cost is unchanged. Applying old paid-development factors can then overstate ultimate losses. Conversely, a shift to slower claim settlement can mimic falling current costs. For a material change, analysts should use claim-count, incurred, paid and close-rate triangulation, test alternate tail assumptions, and report the cost effect. Historical results are diagnostic evidence, not a deterministic claim about a single causal explanation.

Calendar splits, holdouts, and proper forecast scoring

Validation should approximate the decision environment: a model intended to forecast the next calendar year should be fitted on an earlier data cutoff and tested on data that subsequently emerged. Randomly dividing policy rows can leak temporal information when the same customer, event, claim development path or medical episode appears in both training and test sets. Rolling-origin evaluation, cohort holdouts and prospective frozen-baseline comparisons are often more relevant. The validation design must distinguish the prediction of observable current events from projection of ultimate or long-horizon outcomes that remain immature at the test date.

For binary outcomes, a Brier score evaluates probabilistic forecasts as the mean of (pi−yi)2(p_i-y_i)^2 across observations. The properties of strictly proper scoring rules analyzed by Gneiting and Raftery (2007) make them appropriate for assessing probability predictions under defined conditions. If event probability is truly 0.200.20, the expected Brier loss for prediction pp is $0.20(1-p)^2+0.80p^2.Thisequals. This equals0.16$ at p=0.20p=0.20 and 0.170.17 at p=0.10p=0.10. The improvement is small but precisely interpretable: the probability forecast closer to the true conditional event chance has lower expected squared-probability loss. In a real actuarial study the true event probability is unknown and must be inferred from observed holdouts, with adequate attention to uncertainty and subgroup calibration.

Validation should also include economic sensitivity. A model may have marginally better per-life probability scores but materially worse amount-weighted error for the contracts driving the liability. Report both where relevant. Forecast quality for aggregate payouts may depend on correlated risks and exposure concentration not captured by individual conditional predictions. Severe events require separate tail diagnostics and stress analysis; a good mean A/E ratio does not validate the tail.

Parameter uncertainty and alternative futures

An experience assumption can be highly precise conditional on a model and deeply uncertain about that model. Model error arises from wrong event definitions, omitted variables, unsuitable distribution, extrapolation, or changing institutions. Parameter uncertainty arises from finite observations and estimation noise. Scenario uncertainty arises from future states not determined by history. These should be separately described when possible. Sensitivity analysis varies assumptions within defensible alternative ranges and translates them to monetary consequences. Simple perturbations often expose vulnerabilities that are invisible in point estimates.

The ASOP 56 modeling requirements frame testing and model governance in relation to intended use. The Solvency II realism provision specifically emphasizes circumstances under which assumptions would be considered false, consistency, and underlying cash-flow uncertainty. A scientifically meaningful assumption record should therefore include falsification conditions and monitoring triggers. Examples include an observed shift outside a justified prediction interval, sustained divergence by age group, a new data coding regime, a legally mandated benefit change, or a material alteration in insured composition. The trigger calls for review, not automatic selection of the newest observation as a permanent assumption.

Judging emerging medical and social evidence

Emerging interventions, demographic transitions, extreme-weather exposure, technology and societal behavior can change insured experience. Their prospective impact should be evaluated through the chain from mechanism to population exposure to measurable claim or survival outcome, with attention to adoption, countervailing effects and time horizon. For example, improvement in a medical biomarker does not itself establish the magnitude of future insured mortality reduction; the effect requires evidence on clinical outcomes and the insured population's exposure. Likewise, a change in modeled climate hazard cannot be inserted directly as a claim severity factor without policy terms, geographic accumulation, adaptation and financial loss mechanisms.

The RPEC 2026 report discusses possible medical influences but does not claim sufficient evidence for a quantified long-term adjustment based on any single emerging therapeutic category. This is a useful example of disciplined uncertainty rather than analytic inaction: available data can justify closer monitoring and scenario analysis while remaining insufficient to reset the central long-run trend. The scientific conclusion of this chapter is that the evidentiary weight of emerging observations depends on identifiability, forecast horizon, population comparability and financial materiality, not simply their date of appearance.

Documentation and Review

Reproducibility is a scientific property of the assumption

An actuarial assumption is auditable when its decision context, data lineage, calculation, source evidence, responsible review and effective date can be reconstructed. A numerical factor alone is insufficient. The minimum documentation should connect the final selected value to the risk population, event definition, exposure basis, observation period, historical experience, A/E comparator, credibility treatment, prospective adjustment, alternative assumptions and final financial impact. A different analyst should be able to distinguish the observed result from the selected prospective number and identify each adjustment between them.

The ASOP 41 communications standard addresses actuarial communication and relevant disclosures within its U.S. scope. The Solvency II documentation provision adds detailed EU requirements concerning technical provisions methods, inputs and assumptions. Both underscore, in their respective contexts, that method and assumption transparency is integral to professional work. In an international encyclopedia, the governing requirements must be assigned to their actual jurisdiction and measurement purpose, with the model IAA professional standards index used as a comparative reference rather than as a substitute for local law.

A complete source package should preserve extracted record counts, date of extraction, source system identifiers, version-controlled transformation specifications, reconciliation controls, adjustment files, calculation environment and independent test results. Where data are confidential or protected, the public research report may contain reproducible aggregate exhibits and mathematical reasoning while detailed records remain in controlled internal archives. The inability to publish individual records does not justify inventing an empirical replication. Analysts should state what was actually tested and what remains a limitation.

An assumption register with purpose-specific fields

For each material assumption, a practical register records: name and version; contract/product/portfolio; unit and event definition; valuation date and projection horizon; selected central value or function; comparator; historical data period; credibility and weight; trend method; material overrides; prescribed elements; jurisdiction and reporting basis; link to the supporting evidence; quantitative sensitivity; challenge and approval; implementation version; and next review trigger. The structure makes interdependencies visible. Mortality, lapse, claim expenses and discount factors may be individually plausible yet jointly inconsistent if they rely on different projection dates or mutually incompatible state transitions.

A scientific review should include a reconciliation between the prior and newly selected assumptions. Let V(θ)V(\theta) be the relevant valuation functional and θ0,θ1\theta_0,\theta_1 the former and proposed assumptions. The aggregate change ΔV=V(θ1)−V(θ0)\Delta V=V(\theta_1)-V(\theta_0) is unambiguous once valuation conventions are fixed. Attributing changes among assumptions requires a stated method and ordering because interactions may make contributions order-dependent. For a first-order local sensitivity, ΔV≈∇V(θ0)⊤(θ1−θ0)\Delta V\approx\nabla V(\theta_0)^\top(\theta_1-\theta_0) under differentiability and sufficiently small changes, with explicit acknowledgment of higher-order effects. For a major life, pension or catastrophe portfolio, the full nonlinear revaluation is ordinarily necessary to assess material consequences.

Independent numerical and qualitative checks

A reviewer should rerun a sample of exposure calculations and test boundary cases: a policy active for one day, a terminated contract, a reinstatement, a duplicate claim, a missing birth date, a claim paid after the observation period, a negative recovery, a currency change and an extreme benefit amount. The appropriate test set depends on product and data. Summary controls should reconcile person-time and premium exposure to authoritative systems and detect large unexplained movement from previous study versions. A model with a good mean fit but unexplained cell deviations, unreasonable hazard rates, or contract cash-flow inconsistencies should be challenged before approval.

The quantitative checks in this monograph are independently specified hypothetical examples, not fitted insurer studies. The exposure correction yields 0.05000.0500 versus 0.06250.0625 per exposure-year; the portfolio-selection example yields 7.5% versus approximately 18.16%; the mixed A/E illustration yields 104% historically and 95% under the declared new mix; the credibility illustration yields 1.06 and 1.12 for the two assumed credibility weights; the claim development example reverses the apparent direction of trend; the inflation example yields $127{,}312$ of total difference across 200 claims; the mortality projection example yields different year-20 probabilities under two annual improvement rates. These calculations establish numerical coherence of the illustrations. They do not establish that their hypothetical parameters are empirically appropriate for a real insurer.

Scientific challenges and decision rights

Internal review should ask which statements come from direct primary rules or original publisher records, which from academic abstracts, which from mathematical derivation, and which from author interpretation. A reviewer should reject unsupported precision and claims of global legal applicability that exceed the cited standard. An assumption backed only by external general-population data should state that limitation for a selected insured portfolio. A professional standard whose effective date changed must be described with its current scope. The ASOP No. 27–35 transition is one concrete example; the distinct 2026 SOA life and retirement mortality updates are another.

The outcome of substantive review is an editorial decision about the exact version of a manuscript, not a generic judgment that a title appears complete. Internal scientific approval should confirm event and population definitions, mathematical coherence, appropriate sourcing, uncertainty disclosures, comparative-law accuracy and bibliographic traceability. Editorial approval should confirm clear argument flow, stable terminology, short contextual hyperlinks, complete references and the absence of unresolved material factual contradictions. External peer review, if separately commissioned or required, has an independent status and must never be represented by automated checks or an author's internal review.

Ethics, confidentiality and the interpretation of differences

Demographic and behavioral differences in insurance experience may be material to projected cash flows, yet their use raises legal, ethical and statistical questions. An observed subgroup difference may reflect underlying exposure, selection, access to medical services, claim management or record completeness. The study should identify which explanation is supported and avoid presenting association as causation. Whether a characteristic may legally enter underwriting, rate setting or valuation depends on jurisdiction and product; the statistical presence of a predictor does not grant authorization to deploy it.

Experience studies also process sensitive records. The appropriate evidence package retains data provenance and integrity without disclosing identifiable health, financial or policyholder information. Aggregation, access controls, reproducible nonpersonal test examples, and independent review preserve both auditability and confidentiality. The distinction between a public scientific synthesis and proprietary operational data is particularly important for an international research dossier: a mathematical method can be examined publicly while actual portfolios remain protected by their applicable governance rules.

Research limitations and scope of verification

This monograph uses authoritative published rules, standard-setting bodies, public research-institute reports and foundational academic literature. Primary legal provisions and published standards were checked against their institutional publication pages; selected historical journal articles were evaluated through publisher abstracts, catalog metadata and available descriptions, with the difference from full-text examination recognized. The document contains no proprietary insurer data, novel empirical mortality table, international census of national regulations, or independently reproduced mortality study. The 2026 SOA reports are interpreted as publications of their issuing research institute; their aggregate findings are not claimed to be replicated here. Hypothetical arithmetic and symbolic derivations are presented with explicit assumptions. Comparative treatment of IFRS 17, Solvency II, U.S. ASOPs and IAA model standards is intended to clarify their different functions, not to pronounce on every national implementation.

Integrated interpretation

The six investigations produce a connected account. Observed experience first acquires meaning through the definition of population, event, denominator and time. Statistical estimation then compares that experience with a frozen reference, accounting for heterogeneity, credibility and reporting maturity. Prospective selection asks which historical components can be transported to the future population and which require explicit trend, contract and institutional adjustments. Financial measurement then places the selected assumption within a specified reporting and risk framework. Finally, documentation makes the chain from observation to financial decision reproducible and challengeable. Each element has an established scientific or professional foundation; their integration is an original analytic synthesis of the available literature and standards rather than a new doctrine.

Conclusion

Actuarial experience studies and assumption setting form a continuous inferential process whose validity rests on the compatibility of the observed population, measurement method, prospective risk and financial purpose. The first finding is that exposure is constitutive of the estimate: wrong or inconsistent denominators alter measured rates before model selection begins. The second is that actual-to-expected ratios are conditioned on their benchmark, time window, risk mix and claim maturity. They are useful diagnostics and can be misleading as automatic prospective multipliers. The third is that credibility expresses the evidentiary relationship between subject and relevant experience; it requires justified reference populations and does not mechanically resolve portfolio selection or structural breaks.

The fourth finding concerns forecasting. A prospective assumption demands explicit treatment of underlying level, trend, policyholder behavior, contractual design, operational process and relevant changes in the insured population. A short-run shock or an improved recent A/E ratio cannot, by itself, identify a permanent long-run improvement rate. The 2026 Society of Actuaries evidence gives a concrete current example: retirement-plan mortality research retains MP-2021 while insured-life valuation materials introduce their own 2026 improvement scale for a distinct population and purpose. Both positions become intelligible when the study object and assumption destination are stated precisely.

The fifth finding is institutional. Best-estimate cash-flow assumptions, pricing loads, accounting risk adjustments, solvency risk margins and stressed scenarios have distinct functions. The current U.S. ASOP No. 27, the repeal of the former pension ASOP No. 35, the EU Solvency II assumptions and documentation requirements, IFRS 17 and locally adopted professional guidance must be interpreted within their actual scope. The sixth is evidentiary: a defensible actuarial assumption is linked to a reproducible study, a versioned benchmark, a stated forecast horizon, transparent uncertainty, material sensitivity and a documented internal approval.

The resulting research judgment is that actuarial assumption quality depends on the explanatory continuity between historical observation and future contingent cash flow. Sophisticated estimators can improve precision within a defensible study. The fundamental scientific responsibility remains the same across insurance lines and jurisdictions: to establish what was observed, why it is relevant, how its limitations affect inference, what future condition is being assumed, and how that assumption changes the financial result. This monograph supplies a source-anchored synthesis and independently calculable examples for that responsibility. Its conclusions neither constitute a proprietary actuarial valuation nor propose a new authorial theory.

References