Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Mortality Trend Validation and Backtesting in Life Insurance

Abstract. Mortality improvement assumptions influence life insurance premiums, future death benefits, reserves, reinsurance, capital and the assessment of long-term financial performance. Their credibility depends upon how forecasts perform when confronted with deaths and exposures that were genuinely unavailable when the forecasts were made. This monograph examines mortality trend validation as a reproducible scientific and institutional process: defining insured populations and temporal estimands; preserving information vintages; selecting calibration and holdout windows; assessing exposure-adjusted predictions; measuring interval and distributional calibration; separating estimation error, population change and genuine temporal deterioration; and translating observed predictive failures into assumption governance. It develops an integrated examination of Lee–Carter, cohort and multi-population methods, actuarial graduation and credibility, and recent insurance-specific and population mortality improvement practices in the United States and United Kingdom. The analysis distinguishes mortality of insured lives from general population mortality, and annual rate calibration from cohort survival and liability calibration. It evaluates the disruption of conventional backtests by COVID-19, delayed registration, mortality table revisions and changing underwriting composition. Three transparent numerical exercises demonstrate standardized forecast error, the financial direction of mortality improvement assumptions and the effect of erroneous exposure denominators. The principal conclusion is that a trustworthy mortality backtest requires alignment of the forecast's original information set, the observed insured cohort, measurement definitions, forecast horizon and contractual use. Performance must be judged jointly through predictive likelihood, age-duration-specific calibration, interval coverage, structural robustness and decision consequences. The contribution is a critical integration of established evidence and explicitly derived illustrations; no new actuarial theory, proprietary empirical findings or external peer review is claimed.

Keywords: Life insurance; insured mortality; mortality improvement; trend validation; backtesting; out-of-sample forecasting; mortality projection; calibration window; rolling-origin evaluation; exposure at risk; expected-to-actual mortality; Lee–Carter model; cohort effects; stochastic mortality; interval calibration; predictive scoring; mortality basis risk; model risk; underwriting selection; COVID-19; assumption governance; principle-based reserves; actuarial standards; longevity risk.

Introduction: Scientific Question and Research Design

Mortality is a demographic event, an actuarial rate and a determinant of contractual cash flows. A mortality model is therefore examined at several distinct levels. One question asks whether a fitted age-by-year surface reproduces historical death rates. Another asks whether a forecast formed at an earlier decision date would have anticipated subsequent experience. A third asks whether the forecast would have supported credible premium, reserve or solvency decisions for the portfolio actually insured. The subject of this Research Dossier is the connection among these three questions, with special emphasis on genuine temporal prediction and subsequent insured-life observations.

A forecast made on December 31 of a particular year can use only the evidence then available. Once a later experience study, revised death register, updated policy administration file or mortality scale becomes available, reconstructing the historical forecast with those newer inputs introduces information that did not exist at the original forecast date. This form of look-ahead leakage can make an apparent backtest an ex-post reconstruction. The insistence upon out-of-sample testing is consistent with the actuarial modeling standard ASOP 56 (Actuarial Standards Board [ASB], 2019), which explicitly addresses holdout observations and validation suited to a model's intended use.

The research object is projected mortality change for insured lives. Insurance selection, sex, attained age, underwriting classification, policy duration, distribution channel, product type and reporting practice can make the insured portfolio different from the population from which demographic studies draw data. The SOA individual-life experience collection documents insurance-oriented evidence distinct from the Human Mortality Database methodology (Wilmoth et al., 2025). Both are valuable, but a population-level forecast is a candidate comparator for insured mortality, not a substitute for establishing comparability.

The method of this monograph combines documentary review of published research, actuarial professional standards, supervisory and accounting frameworks, explicit mathematical definitions, conceptual counterexamples, and hypothetical calculations with declared parameters. Academic publications support statistical and demographic claims. Official actuarial or supervisory documents support statements about professional and institutional practice. An insured life observed in a proprietary administrative dataset is not inferred from a population publication. No company-specific mortality backtest is reported, because no verified microdata and archived forecast vintages have been supplied.

The disciplinary relation is deliberate. The Life Insurance Science establishes the contractual context; Actuarial Science supplies survival and contingent-value mathematics; Insurance Statistics and Predictive Modeling supplies general statistical inference; Risk Theory places forecast uncertainty within loss distributions; Annuities and Retirement Insurance clarifies the opposite financial direction of longevity risk. This dossier has its own object: retrospective validation of prospective insured-mortality trends and its evidentiary implications.

Mortality Trend Validation as an Actuarial Research Object

The relevant estimands

Let x denote attained age, t calendar year, d duration since issue, s underwriting or risk class and p the defined insured portfolio. Let D(x,t,d,s,p) be the death count and E(x,t,d,s,p) the corresponding central exposure in person-years. The empirical central death rate is

m^x,t,d,s,p=Dx,t,d,s,pEx,t,d,s,p.\widehat m_{x,t,d,s,p}=\frac{D_{x,t,d,s,p}}{E_{x,t,d,s,p}}.

At an operational level, E is reconstructed from actual time at risk under specified treatment of coverage commencement, termination, lapse and death. Under a piecewise-constant hazard approximation, a model may posit

Dx,t,d,s,p∣Ex,t,d,s,p,mx,t,d,s,p∼Poisson⁡(Ex,t,d,s,pmx,t,d,s,p).D_{x,t,d,s,p}\mid E_{x,t,d,s,p},m_{x,t,d,s,p} \sim \operatorname{Poisson}(E_{x,t,d,s,p}m_{x,t,d,s,p}).

The Poisson model is an approximation conditional on exposure and a specified intensity. Multiple policies on one person, common events, heterogeneity and epidemic dependence can invalidate independent policy-level interpretations. In discrete annual survival calculations, death probability q is different from central rate m. Under a constant continuous hazard within the interval, q = 1 - exp(-m); that conversion is assumption-dependent.

A calendar-year mortality improvement statistic can be expressed as

Ix,t=1−mx,tmx,t−1.I_{x,t}=1-\frac{m_{x,t}}{m_{x,t-1}}.

Positive I indicates a decline in the rate; negative I is deterioration. The statistic presupposes matching age definitions, reliable denominators, sufficiently stable comparability and an appropriate reference population. Aggregating over changing age or class distributions can create an apparent trend even when every within-group rate stays constant.

Three different estimands must be retained throughout validation. The first is the conditional rate for an identified age-duration-class segment. The second is a standardized portfolio rate, holding an explicitly chosen composition fixed. The third is the future survival or death-benefit distribution under actual transitions of a cohort. The third depends on future age advancement, policy duration, lapses, underwriting mix and possibly guarantee or beneficiary structure. Validation that succeeds for one estimand need not succeed for the others.

Historical fit, predictive validity and decision validity

In-sample goodness of fit evaluates a model against observations used to estimate it. Predictive validity evaluates a previously fixed model against subsequently unseen outcomes. Decision validity evaluates whether resulting errors are acceptable for a stated use and risk tolerance. These dimensions have different data needs and failure modes. A heavily parameterized curve may closely fit observed age-period mortality but extrapolate poorly because calendar trends and cohort patterns are unstable. A model can predict aggregate deaths acceptably while consistently underestimating mortality for duration-one policies with large sums assured. A forecast with a small average percentage error can still be inadequate if the misestimated cells drive materially different reserves or capital needs.

The distinction is familiar in stochastic mortality research. Lee and Carter's mortality projection study (1992) decomposes historical U.S. mortality into age and time components, while the mortality-model goodness-of-fit comparison of Dowd et al. (2010) investigates how fit assessment varies across stochastic models. The existence of a fitted structural pattern does not, by itself, certify a forecast generated before later outcomes became known.

The unit and horizon of validation

The validation unit must be chosen before scores are inspected. A single person-year is appropriate for one-year mortality incidence. An insurance policy is appropriate for certain claims and administration analyses, subject to duplicate-life control. A sum-assured-weighted exposure is needed when the monetary distribution of death benefits is the question. A multi-year cohort is appropriate for accumulated survival. A valuation liability is appropriate when testing a reserving assumption at its actual economic consequence.

Time horizons should be reported separately: one-year nowcast, three-to-five-year medium-term projection, ten-year projection, and longer runoff expectations answer different questions. Long-horizon backtests require older forecast vintages and therefore may have fewer independent origins. Repeatedly overlapping ten-year windows do not create ten times the independent information of a single decade.

Data Provenance, Cohort Definition and Exposure Construction

The insured-life observation chain

The recording process starts with application, underwriting decision and coverage issue; continues through classification changes, premium payment, lapse, reinstatement and conversion; and ends, for death exposure, with the insured event or another termination. Date of death, date reported, date claim approved, date paid and date entered in the actuarial warehouse are separate fields. Each has a possible effect on what is counted at a historical cutoff.

A backtest should state whether a death is assigned to occurrence year, registration year, notification year or settlement year. Mortality experience studies ordinarily seek the event occurrence basis, but annual reporting files may lag. If one valuation vintage includes all later-reported deaths from an earlier year while another uses only deaths reported by a comparable date, their differences measure reporting maturity as well as mortality. The issue is visible in population work too: CMI Working Paper 211 (CMI, 2026) explains that its England and Wales data are based on registration dates and contrasts them with occurrence-date approaches.

Policy exposure must be reconciled from effective dates and terminations. A policy active for six months generally contributes approximately half a year of central exposure, adjusted for the exposure convention. A lapse before the end of a year removes later time at risk; a death ends further exposure. A policyholder with multiple policies may contribute several policy exposures but only one biological life. A dataset that silently mixes lives, policies and sums assured can generate apparently systematic forecast bias.

The ASOP 23 data-quality standard (ASB, 2016) emphasizes appropriate evaluation and use of data for actuarial services. In a trend backtest, implementing that discipline entails reconciling observation definitions, tracing imputation, versioning extraction code and documenting limitations of source reliability. Because calendar and age denominators may be revised, even an official population series is not a fixed object without its extraction date and methods version. The HMD Methods Protocol Version 6 (Wilmoth et al., 2025) explicitly revised population-exposure estimation using monthly birth information and revised infant mortality calculations.

Deaths, censoring and matching exposures

For life-insurance mortality validation, right censoring occurs when the observed contract experience ends without a death within the window. It is valid to count observed exposure up to the censoring time under the relevant risk-set definition. It is invalid to treat the entire remaining policy term as observed survival time. Lapse and surrender should be examined as possibly informative censoring: if poor-health policyholders retain coverage more often, survivors of the insured block may differ systematically from those who exit. The estimation problem changes with the estimand. Historical death incidence among policies actually in force is observable conditional on reliable files; counterfactual mortality among all originally issued lives after lapse is generally not identifiable from the insurer's own death records alone.

Investigators must record the following metadata together, rather than relying upon a single mortality-rate spreadsheet: issue-date and attained-age conventions; sex and risk class definitions; exact period at risk; policy-versus-life counting; underwriting and channel changes; death occurrence and reporting dates; claims verification; geographic scope; calendar and cohort identifiers; extract and revision timestamps; source-specific censoring; and the approach to missing or conflicting records.

Representative populations and external benchmarks

A general population mortality table is constructed around resident population counts and registered deaths; a life-insurance experience table is constructed around selected persons under coverage. Underwriting can select healthier lives at issue, but later policy duration, class composition and socioeconomic selection can reshape this advantage. Different risk classes can shift relative proportions because of sales, lapse and mortality. Consequently, comparing an insurer's raw aggregate death rate against population improvement may measure composition rather than trend.

The 2015 Valuation Basic Tables (SOA, 2015) illustrate detailed U.S. insured-life basis differentiation, including risk classification and smoker status. They also carry documentation and corrections that matter to reproducibility. The UN World Population Prospects methodology (UN DESA, 2024) documents population reconstruction and projection with evidence quality varying by country. A reliable cross-population validation should therefore treat source coverage and demographic definitions as inputs, not background footnotes.

A coherent comparison standardizes age and relevant contract composition:

mtstd=∑xwxrefmx,t,∑xwxref=1.m^{std}_{t}=\sum_x w_x^{ref}m_{x,t},\qquad \sum_x w_x^{ref}=1.

The reference weights may reflect an opening in-force portfolio, a stable external comparison or the portfolio relevant to the decision. The choice itself must be disclosed. Standardization removes one source of mix distortion, but it cannot correct for unmeasured changing health selection or incomplete death reporting.

Historical Improvement, Future Improvement and Projection Models

Establishing the decomposition

A mortality forecast usually begins with a baseline at a specified age, calendar time and insured classification, then projects subsequent change. In simplified form,

mx,t+hproj=mx,tbase∏j=1h(1−Ix+j,t+jfuture).m^{proj}_{x,t+h}=m^{base}_{x,t}\prod_{j=1}^{h}(1-I^{future}_{x+j,t+j}).

Here the exact age index depends on the cohort being projected and the improvement convention. The baseline must have its own experience period and exposure adjustments. Historical mortality improvement inferred from past mortality rates is an estimate, while future mortality improvement is an explicit forward assumption or predictive distribution. Confusing them allows a favorable historical fit to stand in for untested claims about future change.

In the U.S. life-insurance setting, the SOA 2025 individual-life improvement scale drew on 2011–2019 insured experience for most adult ages, whereas the SOA 2026 scale and recommendation relied on 2013–2023 insured experience for most adult ages and incorporated different future-looking assumptions. Both identify age ranges where insured evidence is less credible and other evidence is required. These documented changes are powerful evidence that improvement assumptions are vintage-dependent, and also a warning against pretending that a 2026 table was available to a reserve model frozen in 2020.

Lee–Carter and alternatives

The Lee–Carter approach models log mortality as

log⁡mx,t=αx+βxκt+ϵx,t,\log m_{x,t}=\alpha_x+\beta_x\kappa_t+\epsilon_{x,t},

where the age effect alpha, age sensitivity beta and time index kappa are subject to identification constraints. A projected kappa path produces an extrapolated age-time mortality surface. The method provides an interpretable starting point, but validation must account for the model's extrapolation assumption, calibration range, late-age fit and structural breaks. Re-estimating all parameters on the complete evaluation period before calling later years a holdout invalidates the test.

A cohort-based Lee–Carter extension (Renshaw and Haberman, 2006) introduces birth-cohort structure, while Cairns, Blake and Dowd's two-factor study (2006) explores older-age stochastic mortality with parameter uncertainty. Cohort effects may capture systematic diagonal patterns in an age-period table, but identifiability and boundary behavior matter. For any age-period-cohort formulation, calendar year equals attained age plus birth year, creating linear dependence and requiring conventions or constraints. Observed model superiority on a specific historical region should be assessed against sensitivity to those conventions and stability outside the calibration region.

When mortality is projected for several connected groups, separate unrestrained models can produce implausible long-run divergence. Hyndman, Booth and Yasmeen (2013) construct coherent multi-population forecasts using product-ratio functional time-series models, with published out-of-sample comparisons for selected national subpopulations. Their evidence justifies including coherent benchmarks in comparisons, while its population setting does not itself establish that the same gain will occur in a heterogeneous insured portfolio.

Stochastic distributions and scenario methods

A single forecast trajectory omits parameter uncertainty, innovation risk and the possibility that mortality follows a different structural regime. A distribution of simulated future mortality surfaces is preferable when the intended decision depends on tail loss, long-term survival or asymmetric product cash flows. Forecast paths should retain correlation over ages and years; simulating each age independently can exaggerate diversification. A proper comparison holds model complexity, data vintage and prediction horizon constant while evaluating probabilistic calibration.

A projection model may blend internal insured experience, external insured benchmarks and general population information. The weight assigned to each source is a modeling judgment informed by credibility, comparability and objective. The ASOP 25 credibility guidance (ASB, 2013) provides actuarial guidance on the use and communication of credibility procedures. A backtest should measure the effect of the chosen blending rule rather than asserting that more years or larger external datasets automatically improve prediction.

Calibration Windows and the Information-Vintage Problem

What the calibration window actually determines

A calibration window has a starting date, an ending date, definitions of included experience, and a closing information timestamp. Its length affects the bias-variance tradeoff. A long period supplies deaths and may stabilize thin age cells; it also mixes earlier medical, behavioral and underwriting regimes with current experience. A short recent period responds more quickly but may be dominated by atypical events or reporting revisions. There is no universally optimal length independent of mortality segment, horizon and model family.

Calibration sensitivity should be tested prospectively. At an origin t0, fit candidate models with data ending no later than t0, using predeclared windows such as five, ten and fifteen years where feasible. Freeze the fitted coefficients, parameters and transformation choices. Score each against the same subsequent experience. Repeat at other origins with fixed rules. Comparing only the retrospectively best-performing historical window and declaring it optimal uses the holdout as an additional training set.

Period selection requires explicit treatment of 2020–2023 COVID-19 experience. Excluding a shock can produce underestimation when similar risks recur; including it mechanically as a persistent trend can overstate future deterioration. An independent structural-break indicator, scenario model or parameter sensitivity may be defensible, but each must be evaluated as a prespecified alternative. Later experience should inform the next assumption update, with its reason and cutoff documented.

Vintage reconstruction

A rigorous historical vintage comprises the exact input extract, data revisions then available, model code version, parameters, manual overrides, imposed constraints, baseline table, projection scale and release date. If archived versions cannot be recovered, the work may be called a reconstructed historical exercise, with a leakage audit describing which elements are retrospective. That exercise remains informative but does not offer the same evidence as an authentic contemporaneous forecast.

Two errors are especially subtle. First, the historical death series itself may have been revised after the forecast date, so retraining on its modern reconstruction can change fitted rates. Second, researchers may retroactively choose the best parameterization after inspecting the test years. The first is data-vintage contamination; the second is model-selection contamination. Reliable evaluation either prevents each or quantifies it separately.

The SOA 2023 mortality-scale correction notice provides a concrete demonstration of model version risk: published future improvement scale values were corrected after release for specified ages. Likewise, CMI lists versioned releases of its mortality projection models. Version metadata therefore determine whether a backtest uses the model actually in force at the earlier date.

Origin-horizon grids

Each scored prediction should be indexed by (origin, target year, forecast horizon, age, contract duration and segment). A prediction issued at origin 2016 for 2021 is a five-year forecast, whereas a prediction issued at origin 2020 for 2021 is a one-year forecast. Pooling them without horizon adjustment obscures deterioration in longer-term performance. The forecast origin must precede the corresponding observation, and its training end date must precede the outcome's information availability.

A complete origin-horizon matrix also allows investigators to distinguish sequential forecasting errors from errors concentrated in one shock year. Backtests across many overlapping origins share common observations and are dependent; uncertainty intervals for comparative performance should use suitable blocks or origin-level resampling rather than assume that every age-year cell is independent evidence.

Designing Genuine Out-of-Sample Mortality Backtests

Research protocol

The protocol begins with a frozen question: predict deaths, mortality rates, rates of improvement, survival, liability cash flows or a specified combination. It states the business decisions being tested, population eligibility, exposure convention, censoring treatment, geographic territory, underwriting classes and target horizons.

For a candidate model M and origin t0, the procedure is:

  1. Construct the information set available at t0, including the contemporary data release and methodological version.
  2. Select a training window and fit M using only the information set, with preprocessing and tuning confined to that window.
  3. Produce point forecasts and, where relevant, predictive distributions for future cells or cohort paths.
  4. Retain immutable predictions with timestamps, code and data hashes.
  5. After each target period matures, derive observed exposure and death outcomes from a defined and versioned validation extract.
  6. Compare predictions against subsequent outcomes, reporting observation delays, missingness and changes in population.
  7. Aggregate scores by horizon, age, duration, sex, underwriting and product, using uncertainty estimates that respect dependence.
  8. Evaluate predefined challenger models and translate substantial errors into a governance record.

A rolling-origin scheme repeats this sequence for multiple t0 values; an expanding window and a fixed-length rolling window are separate candidate designs. Nested historical validation is appropriate when hyperparameters are selected, so the outer holdout remains untouched by the selection procedure. The validation record should identify whether a forecast was generated prospectively or retrospectively with original inputs, and whether assumptions remained frozen for the intended testing horizon.

Benchmark design

A sophisticated model should be challenged by transparent baselines. Candidates include no further improvement from the locked baseline; constant annual improvement fixed at origin; recent historical trend extrapolation; a nationally published scale that actually existed at origin; a structured Lee–Carter projection; and a scenario mixture with specified shock processes. Each benchmark should use the same mortality basis and comparable information. It is unfair to give a challenger a later revised mortality table and criticize the older model for failing to know it.

The SOA 2026 improvement-scale documentation illustrates that both experience windows and transition assumptions change over time. A valid retrospective challenger must use the version available at the historical origin. If a new scale is introduced in 2026, earlier forecasts can be compared to it only in a clearly labeled hindsight assessment.

Backtesting deaths versus modeled rates

A model forecast m can be converted to expected deaths using the actual realized exposure E in the test year:

D^tconditional=∑x,d,sEx,t,d,sobservedm^x,t,d,s.\widehat D^{conditional}_{t}=\sum_{x,d,s}E^{observed}_{x,t,d,s}\widehat m_{x,t,d,s}.

This is a conditional mortality-rate backtest: actual exposure is supplied to isolate rate adequacy. It is distinct from an end-to-end portfolio forecast, which must also predict exposure, lapse, new sales and mix. Both are useful and should not be confused. A forecast that uses realized exposure may perform well on death counts while a comprehensive cash-flow projection fails because future portfolio volumes were misestimated.

The contract unit should also be stated. A term-life portfolio with substantial underwriting selection may exhibit duration-dependent death rates. An annuity portfolio selects persons who survived to annuitization and may have very different socioeconomic composition. Combining their deaths within a single error score can conceal the failure of each product model.

Statistical Scores, Calibration Diagnostics and Uncertainty

Expected-to-actual analysis

An accessible first diagnostic compares observed deaths O with expected deaths E, where E means an expected count and is distinct from exposure. The expected-to-actual convention must be explicitly stated to avoid confusion. Here actual-to-expected is

A/E=∑iOi∑iD^i.A/E=\frac{\sum_i O_i}{\sum_i \widehat D_i}.

If expected deaths are constructed using the forecast mortality rates and the realized risk exposures, A/E above one indicates more observed deaths than forecast in the evaluated cells. It does not alone identify causes: underlying trend deterioration, selection changes, exposure error, claims delay and model misspecification can produce similar signals. An aggregate A/E near one can result from offsetting age or risk-class errors.

For approximately Poisson counts with fixed expected total, a standardized residual is

Z=O−EE.Z=\frac{O-E}{\sqrt{E}}.

The simple reference variance assumes conditional Poisson variation and ignores parameter uncertainty, extra-Poisson dependence and revisions. It should be treated as a diagnostic, not an automatic significance test for all insured portfolios. Prediction intervals constructed under realistic dependence and estimation uncertainty provide a stronger basis for assessing material deviations.

Likelihood and deviance

For observed cell count O_i and forecast mean lambda_i, the Poisson deviance against a saturated mean is

D=2∑i[Oilog⁡(Oiλi)−(Oi−λi)],\mathcal D=2\sum_i\left[O_i\log\left(\frac{O_i}{\lambda_i}\right)-(O_i-\lambda_i)\right],

with the zero-count term interpreted continuously. Deviance penalizes the divergence between predicted and observed cell means and preserves more information than one portfolio-wide A/E ratio. Its distribution and interpretation require scrutiny when many low-exposure cells, heterogeneity or overdispersion occur.

The log predictive score is especially useful when models issue full predictive distributions. Gneiting and Raftery's scoring-rule analysis (2007) explains why proper scoring rules reward candid probabilistic forecasts rather than arbitrarily narrow intervals. For death count distribution P_i, the negative log score for observation O_i is -log P_i(O_i). Scores should be compared for the same cells and exposure conditions. A model whose intervals are narrow only because it ignores uncertainty should not be called superior on point-estimate accuracy alone.

Coverage and sharpness

Interval calibration asks whether a stated nominal prediction interval, for example 90%, contains the observed outcome at approximately the expected frequency over a suitable collection of genuinely held-out cases. Sharpness asks whether those intervals are informative. An extremely wide interval can cover almost everything while giving little guidance. A useful validation reports empirical coverage across ages, horizons, cohorts and periods, alongside interval width and an appropriate interval score.

Correlation complicates coverage. Ten adjacent forecasts may all fail during one pandemic year, and multiple age categories may be affected by one common shock. A model that treats such outcomes as independent gives false precision. Multivariate scenarios and aggregation across suitable blocks better represent correlated mortality innovations. Tail validation is intrinsically difficult when extreme epidemics are rare: a limited historical sample cannot empirically certify extraordinarily small tail probabilities.

Smoothing, credibility and sparse exposures

Raw rates for thin ages fluctuate strongly, so graduation and pooling are often warranted. But smoothing must be assessed for its effect on the validation target. If a smoothing method is fitted using all calendar years before splitting train and test, leakage has occurred even if the final predictive regression excludes those years. Externally sourced graduation assumptions should be documented with their publication dates.

The actuarial credibility standard ASOP 25 identifies a framework for credibility decisions and communication. Credibility weights should be selected from training evidence and relevant similarity to the external base. A small insurer may need to borrow strength from industry studies, but evidence quality depends on how the selected insured mix compares with that industry sample.

Selection, Portfolio Drift and Population Basis Risk

Underwriting and duration effects

Individual life insured populations are actively selected. Newly issued policies may reflect medical or simplified underwriting; some portfolios have waiting periods, exclusions or distinct smoker classes. Even if external mortality improvement is correct for the population, an insurer can experience deterioration if the proportion of less selected lives increases, lapse patterns shift or underwriting standards change. Trend validation should therefore distinguish calendar effects from policy-duration and issue-cohort effects.

A useful rate decomposition considers

mtportfolio=∑gwg,tmg,t,m^{portfolio}_{t}=\sum_g w_{g,t}m_{g,t},

where g denotes an underwriting-age-duration-product segment. The difference between two years combines within-segment changes and changes in weights. A model that projects only within-group mortality must be coupled with a separate forecast of exposure weights to predict aggregate death counts. Without this separation, realized mix can produce a misleading apparent trend. This is the mortality analogue of a changing statistical population.

Multi-population comparison

Population forecasts provide valuable structural information, especially where insured data are sparse at very young or very old ages. But their role as external benchmarks requires an explicit transport argument. Relevant differences include country, socioeconomic status, smoker prevalence, sex composition, occupational mix, underwriting thresholds and cohort entry. The coherent multi-population forecasting study (Hyndman et al., 2013) addresses connections between related population rates, while insurers must additionally validate whether the projected relationship is stable for their selected policyholders.

An external source may also have revisions and reporting differences. The HMD short-term mortality series provides weekly demographic mortality information, which can help identify period shocks but cannot supply the denominator of an insurer's selected policy block. Treating weekly public mortality fluctuation as an insured portfolio's realized annual death rate would cross the estimand boundary.

Reinsurance and sum-assured weights

Reinsurance and capital depend on claim severity and face amount, not simply death frequency. For policies indexed by j, a sum-assured-weighted diagnostic compares

∑jBjDj∑jBjqjforecast,\frac{\sum_j B_jD_j}{\sum_j B_jq^{forecast}_j},

where B_j is the relevant death benefit, D_j is the death-event indicator during a properly defined period and q is the corresponding probability. A single high-benefit death can dominate results in a small portfolio, so frequency calibration and financial calibration should be reported separately. The validity of sums-assured weights also depends on whether benefit changes, ceded portions and product options are modeled consistently.

Structural Breaks, Epidemics and Data Revision

The methodological shock problem

An epidemic creates correlated death experience, uneven impacts by age and geography, changes in coding and reporting, and potential post-shock recovery. It can disturb both current mortality level and the inferred direction of improvement. Forecasts developed before 2020 can be assessed against subsequent COVID-19 outcomes, but the resulting scores describe performance under an extraordinary realized regime; they do not establish that one fixed long-term extrapolation rule must permanently embed the entire shock.

The SOA 2025 historical-improvement scale used a pre-pandemic insured experience window for many ages, whereas the 2026 insured improvement scale incorporated experience through 2023. The documents show explicit methodological management of recent history, including exceptions where insured observations are sparse. A backtest should compare such choices at authentic decision dates, while distinguishing the standard-setting purpose of a valuation scale from the purely predictive purpose of a scientific forecast.

Registration delay and observation vintage

The CMI_2025 release analysis reports sensitivity to death registration patterns in England and Wales. A researcher using deaths registered by a cutoff to infer deaths occurring before that cutoff risks undercounting recent experience if registration is incomplete. Delays can change under administrative stress. Corrective options include a fixed maturity period for scoring, a vintage-consistent nowcasting model or sensitivity bands for pending registrations. Each option must be applied symmetrically to candidate models.

Document corrections matter as well. CMI Working Paper 201 documents a later version associated with CMI_2024. A useful backtest archive therefore stores both initial and corrected data and explains which one represents the historical information set.

Levels, trends and changes of regime

An elevated mortality level, negative annual improvement and an enduring change in long-run improvement are three different statements. After a shock, the annual improvement rate can be strongly positive merely because an elevated death rate returns toward its earlier path. Projecting that recovery rate indefinitely would produce implausibly low future mortality. Conversely, excluding a structural deterioration from all forward assumptions could underestimate future death risk.

An informative decomposition reports the pre-event base, shock excess, near-term recovery assumption and long-run trend separately. CMI Working Paper 214 discusses interaction between base tables and improvement assumptions. The general implication is direct: when a base mortality table already incorporates a temporary adverse episode, applying an improvement path that independently assumes the same episode will recur or be reversed can distort the projected curve. The precise interaction depends on the chosen baseline and model; it must be recalculated, not asserted.

Comparative Institutional Evidence: United States and United Kingdom

United States: insured-specific evidence and valuation purposes

The U.S. actuarial context combines company experience, industry tables, professional standards and state-based valuation requirements. The 2026 NAIC Valuation Manual includes VM-20 requirements for principle-based reserves and prospective mortality improvement treatment; precise applicability depends on contract type and jurisdiction. This document is a regulatory valuation authority in adopting jurisdictions, while a private research forecast might optimize a different loss function. A backtest should identify which role the assumption serves.

The 2026 SOA scale was accepted for the applicable AG38/VM20 valuation year and differentiates historical and prospective components. Its purpose includes standardized reserving practice. An insurer's internal economic best estimate may differ from a prescribed or bounded prudent estimate, and validation should compare them as separate objects rather than treat differences as measurement mistakes.

The SOA Mortality Improvement Model 2026 update reports underlying historical information through 2023 while maintaining the MP-2021 projection scale. This illustrates how new historical evidence and a future assumption can be updated on different schedules. The 2026 RPEC improvement report concerns retirement and post-retirement experience and did not issue a new mortality improvement scale in 2026. Its insights are relevant as comparative longevity evidence but are not equivalent to an insured individual-life death-benefit assumption.

Professional guidance has distinct scope. ASOP 56 addresses model understanding and validation; ASOP 23 addresses data use; ASOP 25 addresses credibility procedures; ASOP 54 addresses pricing of life and annuity products within its stated application; and ASOP 41 addresses communication. These standards are professional requirements for actuaries subject to their applicable rules, not empirically established universal mortality forecasting algorithms.

United Kingdom: population-based model governance

The Continuous Mortality Investigation provides mortality projections through a versioned model family. CMI Working Paper 211 accompanies CMI_2025, calibrated to England and Wales population deaths registered through December 2025. The CMI publication chronology places this release on March 10, 2026. This concrete dating matters in a vintage-based test. The source explicitly describes sensitivity to recent mortality data, registration patterns and selected model settings.

The CMI model is informative for projection practice; it is not a direct insured-life validation dataset for every U.K. insurance or annuity portfolio. Differences in insured and annuitant selection need independent investigation, and subscription restrictions can limit access to detailed working papers or software. A responsible evidence claim distinguishes public release descriptions from a verified execution of the complete CMI software.

CMI Working Paper 214 specifically examines how base tables and improvement assumptions interact. A project that replaces its baseline but retains an older future improvement path should therefore rerun cohort-level survival and cash-flow sensitivity, including possible double adjustment around exceptional mortality periods.

European accounting, solvency and international frameworks

Under Solvency II Article 77, technical provisions are based on best-estimate cash flows and a risk margin, with the best estimate requiring realistic assumptions and up-to-date credible information. The EIOPA revised technical-provisions guidelines elaborate supervisory expectations within the relevant legal setting. This creates an institutional reason to investigate mortality trend credibility and uncertainty, while the regulation does not prescribe one universal statistical backtest.

IFRS 17 Insurance Contracts (IFRS Foundation, effective 2023) uses current measurement of fulfillment cash flows and contractual service margin concepts. A change in mortality assumptions can affect reporting through the applicable measurement mechanics; the impact depends on contract group, coverage units, financial assumptions and whether the change relates to future service. Financial-reporting measurement and statutory reserve measurement must therefore be reviewed under their respective frameworks.

The IAIS Insurance Core Principles establish international supervisory principles, including valuation and enterprise risk management, used by jurisdictions through their own supervisory systems. They are globally relevant governance benchmarks but should not be cited as directly binding legislation in every country. Backtesting evidence becomes practically useful when it is translated into a clearly attributable decision record under the actual institution's legal and reporting obligations.

Contractual and Financial Consequences

Life protection and annuity cash flows

Under a pure life protection contract, higher mortality than expected usually increases the expected incidence of death-benefit payments, all else equal. Under an immediate life annuity, unexpectedly lower mortality generally extends expected benefit payments. Endowments, participating contracts, universal life accounts, riders and joint-life arrangements can combine these exposures in more complex ways. Consequently, the direction of financial error from an improvement assumption depends on the payoff structure, policy terms, asset effects and policyholder behavior.

For a simplified term protection benefit B paid upon death within n years, with death probabilities q_j and survival S_j before year j, expected discounted death payments are

Vdeath=B∑j=0n−1vj+1Sjqj,Sj=∏k=0j−1(1−qk).V_{death}=B\sum_{j=0}^{n-1}v^{j+1}S_jq_j,\qquad S_j=\prod_{k=0}^{j-1}(1-q_k).

For a simplified life annuity due with annual payment A,

Vannuity=A∑j=0n−1vjSj.V_{annuity}=A\sum_{j=0}^{n-1}v^jS_j.

The formulas are stylized and omit policy expenses, tax, guarantees, lapse, reinsurance and other contract terms. They demonstrate why the same underestimated longevity improvement can reduce future death benefits for one product while increasing lifetime payment obligations for another.

Reserves and capital

Cash-flow sensitivity is the decision-relevant extension of rate validation. For each historic forecast vintage, an investigator can rerun the original model's expected claims, reserve or profit projection, then compare subsequent financial experience under a consistently reconstructed contract and investment basis. That financial backtest has its own attribution problem: differences can arise from mortality, lapse, expenses, investment yield, claims settlement or changes in volume. A mortality-specific contribution should be isolated using an explicit attribution order or counterfactual recalculation.

An end-to-end backtest with realized exposure is useful for mortality rate adequacy; an end-to-end business forecast requires exposure projection. A forecast error attributable to a new distribution channel belongs partly to business-mix prediction rather than pure biological mortality. However, the insurer's financial decision may depend on both, so reporting should include component and total-error perspectives.

The relation to risk management is especially significant for long-term guarantees. A small persistent proportional mortality error can accumulate across many future years. Parameter uncertainty and time dependence may dominate point-rate error. This warrants reporting both expected-value sensitivity and plausible distributional tails, without claiming that a finite historical backtest can identify the complete distribution of future catastrophes.

Policyholder fairness and model oversight

Forecast models often differ in predictive accuracy across subgroups. If an aggregate model hides sustained errors for particular underwriting classes, the consequences may include cross-subsidies, mispricing, inappropriate reserve allocations and access impacts. Evidence of differential error does not independently establish a legal discrimination finding; the relevant legislation and justified risk classification must be examined separately. It does establish a concrete validation question that should not be absorbed into one global score.

Assumption Governance and Reproducibility

Institutional accountability

An actuary responsible for a mortality improvement assumption needs a documented explanation of what the assumption estimates, which experience sources it uses, when they became available, how it extrapolates, how uncertainty is measured, and which decisions it affects. A governance record should carry model owner, challenge owner, approval date, data extraction vintage, code version, adopted baseline, improvement scale, scenario set, intended product use, last reviewed actual experience and next review trigger.

ASOP 56's modeling framework supplies relevant validation and communication expectations for covered actuarial work. ASOP 41's communication guidance reinforces the need to explain results and material limitations. In the insurance context, governance is substantive only when validation results can change assumptions, require targeted investigation or alter the permissible use of a model.

Red flags and escalation

Material triggers include repeated forecast underestimation across consecutive vintages; persistent age or class bias; prediction-interval coverage significantly below target; large mortality level changes after claims data corrections; sudden changes in underwriting mix; unexplained divergence between deaths and exposures; evidence of leakage; and a reserve sensitivity exceeding the organization's materiality threshold. The thresholds must be selected before the result is used to judge success.

An exceedance should cause a defined investigation rather than automatic model replacement. Investigators should ask whether the observed discrepancy is due to event timing, a data defect, an unmodeled shock, population composition or model assumptions. Each cause supports a different corrective action. Retroactive redefinition of testing populations to remove adverse evidence should be documented and independently challenged.

Reproducible evidence package

A reusable mortality backtest can be reproduced if it contains: a population and contract specification; input data extracts with permissible access or hashes; exposure-building code; calibration and validation cutoffs; frozen model versions and parameters; all forecast outputs; subsequent death and exposure outcomes; scoring definitions; exception and correction logs; decision materiality thresholds; independent review findings; and authorization of any model update.

Proprietary insured-life data require confidentiality and privacy controls. A published scientific report can still communicate methods, diagnostics, aggregate results and limitations without exposing individually identifiable records. A scientific result that cannot be replicated from public data must say so, rather than transform an internal validation certificate into a claim of public empirical verification.

Analytical Case Studies and Sensitivity

All numbers in this chapter are constructed illustrations, not results from insurer data, public mortality tables, or historical backtests. They are presented so that a reader can verify their arithmetic and identify the conditions under which the conclusions hold.

Case study: aggregate calibration concealing segment errors

Consider three one-year insured-life groups. Group A, ages 35–49, has 60,000 exposed life-years at a predicted mortality rate of 0.003, producing 180 expected deaths, and 210 hypothetical observed deaths. Group B, ages 50–64, has 40,000 exposed life-years at 0.0075, producing 300 expected deaths, with 294 observed. Group C, ages 65–79, has 20,000 exposed life-years at 0.020, producing 400 expected deaths, with 444 observed.

Hypothetical segment Life-years Predicted rate Expected deaths Observed deaths Actual / Expected
Ages 35–49 60,000 0.0030 180 210 1.1667
Ages 50–64 40,000 0.0075 300 294 0.9800
Ages 65–79 20,000 0.0200 400 444 1.1100
Total 120,000 — 880 948 1.0773

The portfolio's aggregate deviation is 68 deaths, giving A/E = 948/880 = 1.07727. Under an illustrative conditional Poisson approximation, Z = 68/sqrt(880) = 2.292. The approximate 95% reference band for the difference under fixed E is plus or minus 1.96 sqrt(880), or about 58 deaths. This is a heuristic detection calculation; it excludes shared shocks and parameter uncertainty.

A segment-level inspection shows a 16.7% excess in group A and an 11.0% excess in group C, while group B is slightly below expectation. Even a total A/E closer to one could arise through offsetting deviations. The governance consequence is to investigate age-specific assumptions and exposures before revising a uniform mortality improvement factor. A single aggregated adjustment could worsen calibration for group B while concealing selection or measurement problems in the other groups.

Case study: mortality improvement and simplified death-benefit value

Assume a purely hypothetical portfolio with aggregate death benefit of USD 100 million allocated to independent identical ten-year exposures, no lapse, no discounting and initial annual death probability q0 = 0.01. Holding q at 0.01 for ten years gives terminal survival

S10flat=(1−0.01)10=0.904382075,S_{10}^{flat}=(1-0.01)^{10}=0.904382075,

so the undiscounted cumulative expected death benefit is USD 9,561,792.50. This deliberately uses a flat age-neutral baseline so that only trend assumptions differ; it is not a realistic age-related insurance pricing basis.

Now suppose annual death probabilities fall by one percent relative each year: q_j = 0.01(0.99)^j for j = 0,...,9. Then

S10improve=∏j=09[1−0.01(0.99)j]≈0.908392806,S_{10}^{improve}=\prod_{j=0}^{9}[1-0.01(0.99)^j] \approx 0.908392806,

and expected cumulative death benefit becomes approximately USD 9,160,719.36. The difference is roughly USD 401,073 lower than the flat assumption. A two percent annual reduction gives terminal survival approximately 0.912209247 and expected death benefit approximately USD 8,779,075.30. These values arise solely from the stipulated toy probabilities.

The exercise clarifies a contractual asymmetry. Greater mortality improvement reduces expected mortality claims for this simplified death-protection product. The same increased survival probabilities raise expected payment duration for a similarly simplified annuity. Real portfolios require age progression, discounting, policy-by-policy benefits, future sales and lapse, underwriting, expense, credit and reinsurance effects. The direction of pure mortality sensitivity survives this elaboration for straightforward death and lifetime-benefit payoffs, but the total contract valuation effect may include offsetting mechanisms.

Case study: exposure denominator correction

Suppose a mortality forecast applied to a portfolio produces 1,100 expected deaths and the study records 1,000 actual deaths. A/E = 1,000/1,100 = 0.9091. A later audit discovers that exposure at risk was overstated by five percent on a multiplicative basis: recorded exposure equals 1.05 times actual exposure, and the model rate itself was unchanged. Correct expected deaths would be 1,100/1.05 = 1,047.619, yielding A/E = 0.9545.

Without the denominator audit, the apparent under-mortality is 9.09%; after correction it is approximately 4.55%. An investigator who adjusted the future mortality improvement path to eliminate the first discrepancy would have attributed a data defect to biological mortality change. The example is independent of detailed forecasting algorithm and shows why exposure governance is logically prior to interpreting a backtest residual.

Case study: vintage mismatch and misleading superiority

Imagine that Model A is frozen in December 2018 using mortality data published by that date, while Model B is fitted retrospectively in 2026 to a subsequently revised historical series and tuned after seeing 2020–2023 outcomes. Both models are then scored against 2019–2023. Model B may appear superior, but the comparison does not estimate which approach a decision-maker in 2018 could rationally have selected. The test has compared different information sets.

A valid comparison would reconstruct Model B using only the 2018 vintage and select parameters in an earlier inner window. The 2026 fitted version can be retained as a hindsight diagnostic with that label. This counterexample derives from the chronology of information rather than from invented numerical performance figures. It establishes that identical test years are insufficient for a fair backtest without equal historical information.

Counterevidence, Research Limits and Robustness

Why ranking models is difficult

Mortality models can be ranked differently by ages, time horizons and evaluation scores. A model minimizing annual death-count deviance may give poor predictions of extreme old-age cohort survival. A smooth long-term demographic model may be robust over decades yet respond slowly to a sudden event. A highly adaptive model can predict recent annual deaths well but extrapolate an unusual recovery indefinitely. No single reported score can cover all actuarially relevant decisions.

Dowd and colleagues' mortality-model evaluation shows the importance of systematic goodness-of-fit comparison, while Gneiting and Raftery's probabilistic scoring framework explains why full-distribution assessment belongs alongside point accuracy. Both are methodological evidence, not a warrant to proclaim a universally best model.

A comparison should inspect at least a no-improvement baseline, a simple extrapolation, an insurer-calibrated model and an alternative structural form. A robust conclusion reports whether rankings persist when calibration start year, pandemic treatment, exposure corrections, age grouping and forecast horizon are varied. If rankings reverse, that is a substantive finding about fragile identification rather than an editorial defect to suppress.

Three categories of uncertainty

Sampling uncertainty arises because death counts are finite, especially at young ages and in small subgroups. Parameter uncertainty arises because estimated historical age and time coefficients are imprecise and correlated. Structural uncertainty arises because future medical, behavioral, socioeconomic and environmental conditions may follow processes outside the model family. Scenario uncertainty additionally concerns exceptional common events and institutional responses.

Only part of this uncertainty is identifiable from a finite historical holdout. The fact that a 95% interval covered 95% of past tested cells under one set of conditions does not establish calibrated catastrophic tails or future structural stability under different conditions. Model governance should state this limitation, monitor leading evidence and use stress scenarios as distinct exercises.

Jurisdictional and source restrictions

The SOA and NAIC material describes U.S. actuarial and valuation practice, CMI material describes model releases for England and Wales, and European legal material addresses the relevant European insurance regime. None is a single worldwide compulsory mortality method. Published population data are subject to country differences in registration and exposures. Subscription-controlled studies and software may not expose complete reproducibility artifacts in public pages. The present research examines the identified publisher and institutional descriptions and accessible scholarly bibliographic records; it does not claim to have downloaded restricted full texts, executed CMI software, conducted industry interviews or reestimated proprietary mortality tables.

There is a further limit to the available empirical record. No archived insurer data, original time-stamped internal forecasts or comprehensive individual policy death records were supplied. The mathematical case studies establish internal arithmetic and conceptual sensitivity, not empirical outperformance of any model. This distinction is retained in the scientific assessment, evidence ledger and editorial passport.

Research agenda grounded in existing methods

The most consequential unresolved empirical questions concern transportability and stability: how underwriting selection interacts with recent population mortality shocks; how long duration-specific improvement differences persist; how forecast intervals should incorporate joint exposure and mortality uncertainty; and when a stable long-run assumption should be revised after severe short-run deviations. These are testable questions for appropriately governed confidential insurer datasets or genuinely open insured-life aggregates. They can be addressed with established cohort, credibility, time-series and hierarchical methods, without announcing new proprietary theories.

Integrated Scientific Synthesis

The evidence supports five connected results. First, mortality trend validation is a temporal scientific claim: forecast parameters, information availability, experience definitions and outcomes must be ordered chronologically. A fit obtained with future knowledge cannot substitute for a genuine historical prediction. Second, insured mortality is generated by a selected, managed portfolio; improvements observed in national populations are valuable contextual evidence but require documented comparability before they are transported to insurance liabilities.

Third, a mortality-rate forecast and a financial contract forecast have different validation boundaries. Conditioning on realized exposure is often desirable when diagnosing rate adequacy, but it excludes volume and mix error. Full financial consequences require cohort transitions, lapses, death-benefit weights and contract-specific valuation assumptions. Fourth, model quality requires several complementary diagnostics: expected-to-actual deaths, disaggregated residuals, properly scored predictive distributions, interval coverage, horizon-specific accuracy and material financial sensitivity. Each evaluates a different aspect of adequacy.

Fifth, actuarial standards and supervisory frameworks make evidence governance consequential. The 2026 U.S. insured improvement recommendation and CMI_2025 projection documentation demonstrate that data windows, modeling conventions and revisions are real parts of actuarial production. Solvency II's current-information requirement and IFRS 17's current cash-flow measurement show why the consequences extend to institutional reporting and financial obligations.

The original analytical contribution here is to make those established results mutually consistent for the narrower task of testing insured mortality trend assumptions. Its organizing distinctions—historical versus prospective information, rate versus portfolio exposure, point versus distributional prediction, population versus insured cohort, statistical versus financial loss, and published standard versus local legal authority—are analytical classifications of known phenomena. They are not proposed as a new theoretical system. Taken together, they explain why credible mortality projection validation needs an evidence trail extending from individual observations and versioned forecasts through subsequent cohort outcomes to documented actuarial decisions.

Conclusion

Mortality trend validation in life insurance is completed only when the forecaster can demonstrate what was predicted, with which information, for which insured population and time horizon, how later deaths and exposures were reconstructed, how predictive failures were scored, and what those failures mean for the contract and institutional decision. A defensible test preserves original prediction vintages, maintains cohort and exposure comparability, reports disaggregated and distributional performance, and separates extraordinary shocks from claims about persistent long-run improvement.

The comparative record establishes important differences between U.S. insured-life improvement scales, U.K. population-based CMI modeling and international financial reporting or solvency frameworks. Their evidence can be combined through transparent assumptions, while each retains its own scope and governing authority. The numerical exercises demonstrate that segment aggregation, altered trend assumptions and exposure mismeasurement can change the apparent result and its financial interpretation. They are derived examples rather than empirical insurer findings.

The general scientific conclusion is that the quality of mortality projection depends upon the integrity of the entire forecasting and validation sequence. Reliable future use requires reconstructable historical forecasts, credible observations, sound statistical comparison, explicit recognition of portfolio selection and disciplined assumption review. These conditions support more accurate scientific attribution and more accountable insurance decisions without requiring a new doctrine or an unverified assertion of predictive superiority.

References