Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Longevity Model Backtesting for Annuity Liabilities

Abstract. Longevity models determine how long annuity payments are expected to continue, how their uncertainty accumulates, and how future benefits are reflected in present liabilities. This monograph establishes a source-critical, reproducible approach to evaluating these models with evidence unavailable at their original forecasting dates. Backtesting is defined as an ex ante, information-set-preserving comparison of mortality and contractual cash-flow forecasts against subsequent experience. The appropriate test is specific to the forecast target: exposure-weighted death counts assess near-term demographic calibration, cohort survival evaluates the accumulation of mortality changes, and discounted liability errors measure financial relevance. Drawing upon the Human Mortality Database, published annuitant experience, the Lee–Carter and Cairns–Blake–Dowd model families, Continuous Mortality Investigation projections, United States mortality improvement research, and international valuation standards, the study distinguishes retrospective fit from prospective performance. It specifies rolling-origin designs, vintage control, exposure and benefit weighting, probabilistic scoring, cohort diagnostics, uncertainty attribution, and contract-sensitive reserve tests. An independently calculated hypothetical annuity illustrates how plausible differences in mortality improvement assumptions change present values despite identical initial death rates. Recent experience studies offer important validation benchmarks, yet their aggregate publication summaries do not furnish the historical cell-level data required to substantiate a new empirical backtest; this work therefore presents an analytical synthesis and replicated calculation rather than fabricated portfolio-level test statistics. The principal conclusion is that longevity backtesting is complete only when forecast-origin integrity, annuitant representativeness, distributional calibration, and the financial significance of errors are examined together. The result is an operationally reproducible evaluation specification within established actuarial science, without claiming a new mortality theory or original empirical discovery.

Keywords: Longevity risk; annuity liabilities; mortality forecasting; backtesting; stochastic mortality; cohort survival; annuitant selection; model validation; forecast calibration; proper scoring rules; actuarial reserves; IFRS 17; Solvency II; mortality improvement; vintage data.

Introduction

A life annuity converts survival into a stream of legally specified benefit payments. For the individual recipient, surviving another payment date generally increases the total income received. For the issuer of a life-contingent annuity, survival beyond the expected pattern increases expected future outflows relative to a valuation that assumed earlier deaths. A model used to estimate the cost of this promise must therefore account for the age distribution of insured lives, the timing of cash flows, future mortality improvement, contract guarantees, and the financial basis on which projected payments are discounted. Mortality assumptions are among several interacting inputs, but their cumulative errors can remain hidden for years while emerging through changes in reserves and capital requirements.

The OECD's mortality study (2023) describes the development of standard tables for retirement-income provision and stresses the match between a reference population and the population actually exposed to the obligation. The CMI Mortality Projections Model is widely used in United Kingdom actuarial practice to translate assumptions about mortality improvement into projected rates. The SOA payout annuity experience study supplies more directly relevant United States observations for a covered annuitant market. These resources address related but distinct questions: general-population mortality, future improvement assumptions, and selected insured experience are not interchangeable observations of one risk.

The method of backtesting asks a precise historical question. Given only the data and methodological choices accessible on valuation date v, what distribution or trajectory of deaths, survival counts, and annuity payments would a model have predicted for subsequent years? How did those forecasts compare with the later observations, and which components of any discrepancy are financially material? The question requires a sequence of sealed information sets. Estimating a model in 2026 and pretending that its modern parameters were available in 2010 produces a retrospective demonstration, not a genuinely out-of-sample assessment. Likewise, showing a close match between observed deaths and a model refitted to those deaths establishes fit and provides little evidence about future forecast performance.

There are several degrees of validation. A calculation can be internally correct and still rely on an unsuitable population. A mortality curve can fit population death rates but be unrepresentative of policyholders. A model can correctly forecast average deaths while understating tail risk. A model can make reasonably calibrated mortality predictions and nevertheless generate poor liability estimates because the payment schedule, survival option, guaranteed period, benefit indexing, or discounting basis is misspecified. The ASOP 56 modeling standard (Actuarial Standards Board, 2019) connects model testing, inputs, validation, controls, and intended purpose. This monograph follows that principle through the complete chain from source observation to payment obligation.

The study is an original synthesis of documented demographic, actuarial, statistical, and supervisory methods, together with a reproducible synthetic reserve illustration. The available reviewed sources include official institutional methodology pages, publicly described industry experience, published research abstracts, and applicable regulatory explanations. No confidential company mortality extracts, complete annuity policies, or new cross-country panel have been provided. Claims about actual predictive superiority or empirical coverage are consequently reserved for a fully specified test with authentic historic vintages. Numerical results labeled illustrative are computed from stated artificial assumptions and never presented as measured annuitant experience.

The dossier belongs simultaneously to Annuities and Retirement Insurance, Insurance Statistics and Predictive Modeling, and Insurance Pricing and Reserving. That common scientific ownership is consequential: a demographic forecast becomes an insurance conclusion only after contract cash flows, portfolio composition, and the relevant valuation basis have been specified. The following six substantive chapters implement this sequence, ending with separately identified conclusions, verification limits, and future empirical tests.

Historical, Mathematical, and Institutional Foundations

The annuity as a survival-contingent claim

Let a closed cohort of N lives be observed at integer age x on valuation date v. Denote by q(x,t|v) the conditional probability of dying between ages x+t and x+t+1 implied by the information available at v. The corresponding one-year survival probability is p(x,t|v)=1−q(x,t|v). For annual payments occurring at the end of each year, the probability that the original annuitant reaches payment date k is

kpx(v)=∏j=0k−1[1−q(x+j,j∣v)].{}_kp_x(v)=\prod_{j=0}^{k-1}\left[1-q(x+j,j\mid v)\right].

For a fixed amount B payable only while alive at the end of each year, a finite-horizon best-estimate present value under a deterministic discount curve D(v,k) is

Vx(v)=B∑k=1TD(v,k)kpx(v).V_x(v)=B\sum_{k=1}^{T}D(v,k){}_kp_x(v).

For independent lives with homogeneous benefits this expression scales by N. Real annuity portfolios require an individual contract sum because entry ages, commencement dates, guarantees, amounts and benefit schedules differ. The displayed formula assumes that the cash-flow and mortality bases are mutually consistent, does not include expenses or guarantees, and is an expected discounted benefit value rather than a complete accounting or statutory reserve.

The key distinction is between a period life table, which combines age-specific death rates observed during one calendar period, and cohort survival, which follows an aging generation through different calendar years. A period table is a valuable snapshot; pricing a future lifetime stream requires future rates for the cohort as it ages. The Human Mortality Database overview describes the officially sourced death counts and population estimates used to construct comparative time series. The OECD account of modeling choices (2023) explains age–period–cohort structures and the distinctions among model families. An annuity projection should therefore specify both its initial table and its improvement path.

Mortality is often represented through central death rates m(x,t), one-year death probabilities q(x,t), or a continuous force of mortality μ(x,t). They are related under a stated exposure convention; they are not numerically identical. A model fit to log central death rates must be converted consistently before inserting annual probabilities into an annuity-payment formula. In one common approximation, q=1−exp(−m) under a piecewise constant force over the year. The approximation is appropriate only when its exposure and timing assumptions match the data and the model. At old ages, or where death-rate smoothing and age splitting differ, unexamined conversions can produce measurable projection changes.

From mortality tables to stochastic projections

The Lee–Carter study (Lee and Carter, 1992) represented age-specific log mortality using an age level, an age-specific loading, and a time index. A stylized form is log m(x,t)=a_x+b_x κ_t+ε(x,t), with identification constraints required because the factors are otherwise not uniquely determined. Forecast uncertainty is generated by projecting κ_t and by recognizing further sources of uncertainty, rather than by merely reproducing the observed time series. The study established a leading family of long-horizon mortality forecasting methods; it did not make any one index specification optimal for every annuity population.

The two-factor mortality paper (Cairns, Blake, and Dowd, 2006) instead examines mortality at older ages with two time-varying terms influencing the level and age slope of the logit of death probabilities. Its explicit treatment of parameter uncertainty is especially relevant when pension or annuity liabilities extend over many decades. More flexible age–period–cohort structures can represent persistent cohort features, but greater parameterization increases estimation and identification demands. The OECD international comparison gives a useful map of these structures and their interpretation.

Neither family directly supplies a universal annuitant projection. An insurer must select or estimate the mortality level for its own population and determine which general improvement process is transferable. Published model comparisons often differ in age ranges, historical periods, countries and evaluation scores. A defensible backtest therefore preserves each model's exact forecast origin, training sample and forecast target instead of treating a universal ranking of model names as an empirical result.

Population, insured lives, and historical exposures

Observed population mortality uses deaths and population exposure. Annuity experience typically refers to a selected group of customers, to contract-year exposure, and sometimes to amount-weighted rather than life-weighted deaths. One person may own multiple contracts; a contract-year is consequently not automatically a person-year. The SOA 2020–2024 individual payout study reports participation by 23 parent-company groups across 26 companies, with approximately 3.1 million contract-years of exposure and 143,190 deaths. These aggregate counts document the scale and relevance of the source, while the detailed company and age-cell records are not reproduced in the public study summary. Dividing total deaths by total contract-years would blend ages, product types, benefits and calendar years and cannot support an age-specific calibration or portfolio backtest.

The SOA deferred annuity study reports about 80 million contract-years over 2015–2024 among 26 contributing companies. Deferred contracts have distinct participation, account-value, surrender, and payout characteristics. Their experience may help evaluate selection or transition questions, but it must not be substituted for deaths among annuitants already receiving life-contingent benefits. In a similar way, pension member data and retail payout annuity data require cohort and contract matching before their comparative results can be interpreted.

Institutional relevance and boundaries

The IFRS 17 accounting standard uses current measurement of fulfillment cash flows, an explicit risk adjustment and, where applicable, a contractual service margin. A mortality model changes one set of actuarial expectations entering that accounting measurement. The full recognized liability also depends on contract grouping, services, discounting and accounting rules. The IAIS supervisory principles provide an international supervisory framework, while the Solvency II longevity submodule sets an EU regulatory stress specification. These frameworks address different objectives; a scientific backtest reports the underlying forecast behavior separately from each jurisdiction's prescribed financial calculation.

The historical lesson is that survival probabilities, mortality forecasts, insured population selection and reserves form connected layers of one liability valuation. Reconstructing their vintages and definitions is the prerequisite for evaluating what a past model could have known.

Projection Backtesting and Source-Consistent Research Design

Forecast-origin integrity

A defensible experiment begins by freezing a forecast issuance date v, its accessible source versions, and the exact model implementation. Each vintage record should preserve the data extraction date, country and population definition, covered ages, death registration cutoffs, exposure estimates, base table, trend model, estimation window, parameter values, projection horizon, code version and benefit mapping. Later corrections to historical deaths or exposures are especially important: using today’s revised data in an allegedly historical estimate introduces look-ahead information.

The Human Mortality Database data policy notes that its statistics are updated and that country input sources and conditions require attention. As a result, a contemporary download of historic years is not by itself a preserved 2014 vintage. Where original public vintages are unavailable, the researcher should explicitly label the exercise pseudo-out-of-sample on revised historical data and perform a vintage sensitivity analysis; any inference about genuine real-time predictability remains conditional. The distinction matters particularly for recent observations affected by registration lags, population denominator revisions, and retrospective COVID-19 classification changes.

At each training cutoff v, estimate the model using information available no later than v. Issue forecasts for horizons h=1,…,H, then compare each with the realization at v+h. Shift v forward on a predetermined calendar schedule and repeat. A rolling window keeps training length fixed; an expanding window adds information cumulatively. The specification should contain at least one naive benchmark, such as a fixed age-specific mortality-rate basis, and a trend benchmark, such as continuation of the last documented improvement assumption. Comparative scores must be evaluated on exactly the same ages, dates and exposures.

Overlapping forecast horizons create correlated errors: the 15-year forecast made in one year and the 14-year forecast made the next year share many outcome years and common mortality shocks. Confidence intervals for pooled forecast errors should account for serial dependence and repeated observations, for example through a defensible year- or origin-block bootstrap. Treating all overlapping outcomes as independent makes precision appear greater than the experiment supports.

Calibrating death counts without confusing exposure

For a cell i describing age, sex or other permitted grouping, calendar year and policy duration, let E_i be the relevant exposure measure, D_i actual deaths and q^i\widehat q_i the appropriately aligned mortality probability. A simplified expected count is D^i=Eiq^i\widehat D_i=E_i\widehat q_i, assuming that E_i has the exposure definition to which q applies. If exposure is central person-years, a force-of-mortality formulation may be more natural. The observed-to-expected ratio is

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

An A/E below one indicates fewer observed deaths than expected under the specified cell weights; for a pure life annuity it can indicate greater persistence of benefit payments. One overall ratio can conceal offsetting age-, cohort-, duration- and amount-specific errors. At least four complementary views are therefore required: age-standardized deaths; cohort survival to specified horizons; amount-weighted expected payments; and the present value of the actual benefit stream within its observable period. A claim of unbiased prediction based only on a total A/E near unity is analytically incomplete.

When individual deaths are modeled as conditionally independent Bernoulli outcomes, a binomial log score is available. For each cell with n independently exposed equivalent lives and d deaths, its score, omitting terms constant across candidate models, is −dlog⁡q−(n−d)log⁡(1−q)-d\log q-(n-d)\log(1-q). Independence is often imperfect because of shared trends, epidemics and insurer selection. Aggregate predictive distributions should reflect dependence where the purpose is longevity capital rather than solely estimating the mean. The proper scoring rule literature (Gneiting and Raftery, 2007) explains why coherent distributional scores reward honestly specified probabilities and allow models to be compared without rewarding artificially narrow uncertainty bands.

Survival, coverage and probabilistic evaluation

A survival backtest reconstructs realized survival proportions for a closed inception cohort, taking account of death, censoring and benefit status. In life-insurance portfolios, policy termination and mortality must be distinguished; cancellation of a contract is not a recorded death. A valid estimator identifies what is known about lives lost to observation and which cash flows stop under actual contract provisions. If data support Kaplan–Meier estimation under plausibly independent censoring, that estimator can be useful, but comparing it with a model projected for a different coverage state changes the estimand.

For each horizon h and vintage v, evaluate both the point forecast and predictive uncertainty. Point diagnostics include errors in survivor count, mortality rate, expected cash flows and reserve value. Distributional diagnostics include nominal coverage of forecast intervals, probability integral transform behavior when a continuous predictive variable is appropriate, tail-event frequency, and proper scoring rules. A 95% interval with 95% coverage in pooled historical periods might still be unreliable for older cohorts or periods of abrupt trend change; report conditional coverage by important strata.

An informative minimum output table contains forecast origin, horizon, source vintage, model and parameter specification, projected survivor fraction, observed survivor fraction, predicted deaths, observed deaths, benefit-weighted forecast error, reserve error and interval score. Benchmark comparisons must be paired by valuation and target. Where a horizon has not yet fully elapsed, label it as censored and avoid calculating a supposed full-horizon success rate.

Separating calibration, discrimination and economic loss

A fitted model can rank cohorts correctly by mortality while remaining systematically biased at all ages. Calibration asks whether predicted probabilities match realized frequencies over comparable groups. Discrimination concerns the ability to distinguish different mortality risks. Liability prediction is a third criterion: a small error for a high-benefit older annuitant may have a different financial consequence from the same probability error for an individual receiving a much smaller benefit. Define, for forecast v and evaluation horizon H,

Lv(H)=∑k=1HD(v,k)[Cv+krealized−C^v,k],L_v(H)=\sum_{k=1}^{H}D(v,k)\left[C_{v+k}^{\mathrm{realized}}-\widehat C_{v,k}\right],

where realized cash flows are on the same contract and discount basis as forecast cash flows. The expression measures a discounted observed-window payment error; it is not automatically the entire ultimate reserve error while remaining payments are unobserved. Over a partially observed horizon one may alternatively value the remaining book at the end date under a stated independent reference basis; this extension introduces a new model assumption and must be separated from observed-only outcomes.

Tests should cover policy counts, benefit amounts, payment frequency and guarantee features. For an annuity with a guaranteed payment period, mortality does not determine the guaranteed-period obligation in the same way as it determines post-guarantee survival-contingent cash flows. A model that ignores the contract can falsely attribute a benefit schedule mismatch to mortality. This is one reason statistical backtesting and contract-aware cash-flow reconciliation must be reported as two connected but separately interpretable results.

Ex ante model competition

The research design should fix candidate models and evaluation metrics before observing the test outcomes. Choices made after examining failures may be useful model development but belong in a subsequent vintage; using the repaired model to grade the original historical prediction creates selection bias. A benchmark's value is explanatory: a sophisticated model that produces less reliable cash-flow forecasts than a transparent and properly calibrated baseline offers limited evidence of decision value. Rankings should be accompanied by loss differences and uncertainty ranges rather than single ordinal positions.

The recommended experiment thus uses historical information discipline, actual exposure definitions, comparable forecast targets, explicit baseline forecasts, coherent dependence treatment and separate demographic and financial scores. Its validity depends on these controls, not on the complexity of the forecasting algorithm.

Cohort Calibration, Annuitant Selection, and Forecast Drift

Why a population forecast is not automatically an annuitant forecast

Population-based models are attractive because national death registrations can provide long, largely consistent time series. The HMD methodological overview explains that coverage requires sufficiently complete vital registration and census information, limiting the populations that can be treated as harmonized. Even for those countries, an insured annuity portfolio differs from its entire population through income, occupational history, access to health care, underwriting, retirement age, benefit design and purchase timing. The OECD report on retirement mortality observes systematic differences between pensioner or annuitant mortality and population mortality, and recommends choosing base assumptions representative of the covered group.

Selection also changes over time. New annuity sales can attract people with different expectations of longevity than earlier cohorts. Changes in distribution, retirement regulation and product features can shift the mix of participants independently of any biological change. The Finkelstein–Poterba annuity study (2004) identified relationships between ex post mortality and contract characteristics in the United Kingdom, while finding no substantial mortality differentiation by annuity size in that particular dataset. Its methodological implication is that selection can operate along contract features even when a familiar single measure fails to reveal it. A modern backtest should therefore stratify by payout design, product election and duration where source access and sample credibility permit.

A general-population improvement rate could be transferable even when its base mortality level is unsuitable; conversely, two groups with similar base mortality may follow different trends. Fit and validate these dimensions separately. For instance, a multiplicative adjustment to period mortality at v may improve the A/E ratio at inception while leaving the long-term forecast materially biased if the population improvement trajectory is misestimated. Recalibrating the level each year can hide that structural error in a one-year test, whereas frozen multi-year cohort projections expose its cumulative effect.

Cohort and period effects

A cohort is defined by birth year b=t−x. Repeatedly lower mortality for members of a birth cohort as they age can reflect childhood conditions, socioeconomic composition, smoking history or other enduring influences. A period shock affects multiple ages in a calendar interval, although its impact may vary considerably by age. Distinguishing cohort and period effects requires identification choices: in an age–period–cohort relation, birth cohort equals calendar period minus age, so unrestricted additive effects are not uniquely identified without constraints. A fitted cohort term should be tested for stability across training windows and against simpler specifications.

An annuity book has additional cohort dimensions: issue year, age at purchase and years since annuitization. Two individuals born in the same year who started payments at different dates may not share the same selection process. A backtest that aggregates by calendar age alone may combine entry selection with mortality improvement. The correct test design reports explicitly whether a cohort refers to birth year, policy issue, retirement, or first payment.

The Cairns–Blake–Dowd mortality model was developed for older-age stochastic mortality and helps motivate attention to both age levels and age gradients. The Lee–Carter forecast structure highlights a common period factor. In a backtest, each method should be assessed on its own specified age range and forecast object. A cohort-sensitive model should be credited for incremental predictive value only if it outperforms a simpler model on genuinely withheld cohorts and economically relevant outcomes.

Forecast drift and pandemic-era revisions

Forecast drift is the systematic development of errors as the horizon lengthens or as new observations repeatedly shift expected mortality. It may result from a slowly changing improvement trend, a structural break, a sequence of period shocks, revision of base exposure, or changing annuitant selection. The distinction matters because the remedies differ. A level adjustment changes today's mortality table; a drift adjustment changes the anticipated trajectory; and a contract selection adjustment changes the insured-group mapping. Applying all three indiscriminately can double-count the same observed deviation.

UK and United States evidence illustrates how source vintages can encode different responses to recent events. The CMI_2024 methodological explanation describes full weighting of recent years plus an explicit pandemic overlay, replacing the earlier reduced-weight approach. The CMI_2025 published explanation states that the model uses 1985–2025 data and retains the CMI_2024 methods apart from data updates. Working Paper 211 separately emphasizes that death-registration dates can yield differences from occurrence-date data. These are exactly the kinds of information-state and data-definition changes that a real-time backtest must archive rather than remove retrospectively. A further source-integrity example is the CMI notice of corrected tables, which records a corrected second version of Working Paper 211 after several sensitivity tables had originally compared the wrong historical rather than prospective improvement interval. A credible backtest records the paper version and the correction date separately: calculations published before an erratum must be distinguished from their corrected successors when reconstructing the contemporary information set.

The 2026 RPEC mortality update (SOA Research Institute, October 2026) reports an age-standardized actual-to-expected mortality ratio of 98.7% for ages 65 and older during July 2025–June 2026 against its specified 2019 baseline projected using MP-2021. It also states that the committee did not issue a new mortality improvement scale in 2026. The 98.7% figure is an external population-based reference calculation with its own period, comparator and exposure conventions. It is neither an annuity-company reserve result nor the empirical score of a model independently reconstructed in this monograph. The 2026 MIM update clarifies that updating the tool's data and capabilities does not create a new MP scale. A rigorous study must not conflate model-tool version, historical data revision and adopted prospective assumptions.

Diagnosing drift without rewriting the record

For each forecast vintage, construct an error surface indexed by evaluation age and horizon. Persistent errors of one sign among young-old ages, older ages, issue years or benefit amounts may reveal population or cohort mismatch. If errors change sign across ages, an overall A/E can be misleading even when total projected deaths are correct. If one-year probabilities are accurate while long-term survival forecasts depart progressively, review trend extrapolation, shared-factor dynamics and cohort terms. If age-specific demographic forecasts are good but cash-flow projections err, review annuity contract data, amount weights and policy options.

Diagnostic labels should remain separate from causal conclusions. A backtest can demonstrate that a forecasting regime failed to anticipate a period shock; proving that a specific mechanism caused the shock normally requires further empirical evidence. Ex ante stress scenarios for pandemics, socioeconomic divergence or trend breaks should likewise be presented as scenarios, not as backdated forecasts that a historical model somehow issued.

Cross-population portability

The same estimator can be tested on several jurisdictions, but its observations must be reconstructed independently. A population coverage rule appropriate for the Human Mortality Database is different from eligibility for a proprietary insurer dataset. Country differences in pension decumulation, guaranteed annuitization, voluntary purchase and underwriting make experience comparability a substantive research question. The OECD comparative guidelines supply a basis for asking whether mortality assumptions reflect the population receiving retirement payments; they do not justify importing one country's A/E ratio into another without adjustment.

The chapter's central inference is that cohort calibration and forecast drift cannot be meaningfully interpreted without a preserved population and contract definition. Long-run improvement error and changing annuity selection can lead to similar aggregate symptoms. Their discrimination requires segmented data and genuinely historical forecast vintages.

Reserve Sensitivity and Financial Materiality

What a mortality prediction means for liability value

A pure life annuity has a direct exposure to future survival: higher survival probabilities increase expected life-contingent payments, holding contract terms and discount rates constant. The strength of this effect depends on the age distribution, benefit shape, projection horizon and timing of payments. In a benefit stream extending decades into the future, small changes in annual death rates can compound into sizeable changes in expected late-year payments. Discounting dampens distant cash flows, but long-duration obligations remain sensitive to systematic improvement.

For a change in mortality forecast from q to q', define the expected-payment liability difference under a fixed financial basis as

ΔV=∑k=1TD(v,k)Bk[kpx(q′)−kpx(q)].\Delta V=\sum_{k=1}^{T}D(v,k)B_k\left[{}_kp_x(q')-{}_kp_x(q)\right].

The sign is clear for a pure life annuity when q' is lower at all relevant future ages: every affected future survival probability is weakly greater and the discounted value increases. This monotonicity may change for net asset-liability positions or products with death benefits and other contingent features. A guaranteed period changes which payments respond to mortality. The IFRS 17 key-terms guidance emphasizes that fulfillment cash flows and the risk adjustment are elements of the insurance contract measurement; the simple expected-benefit formula isolates one input to that broader measurement.

Independently replicated synthetic annuity calculation

Consider one hypothetical annuitant aged exactly 65 at valuation, with a level annual payment of 10,000 currency units at the end of each year conditional on survival and a 30-payment evaluation horizon. Assume a constant 3% effective annual discount rate, no expense or taxes, no minimum guarantee, no joint-survivor option, and death probabilities at attained age 65+j specified by

q65+j,j(g)=0.012(1.09)j(1−g)j,j=0,…,29.q_{65+j,j}(g)=0.012(1.09)^j(1-g)^j,\qquad j=0,\ldots,29.

Here g is the annual improvement factor applied after the first year. The age-escalation factor 1.09, starting mortality 0.012, and values of g below are hypothetical construction inputs; they are not estimates from SOA, CMI, OECD or a national population. The formula maintains the same initial one-year mortality at age 65 across scenarios while changing its future development. For each payment year k, survival is the product of one minus each preceding annual q, and the projected benefit present value is

V(g)=10,000∑k=130(1.03)−k∏j=0k−1[1−0.012(1.09)j(1−g)j].V(g)=10{,}000\sum_{k=1}^{30}(1.03)^{-k}\prod_{j=0}^{k-1}\left[1-0.012(1.09)^j(1-g)^j\right].

Straight double-precision iteration gives the following replicated results, rounded for display:

Annual assumed improvement g Survival to year 10 Survival to year 20 Survival to year 30 Benefit present value
0.0% 0.831851 0.534771 0.181286 138,430.30
1.0% 0.839720 0.575405 0.251288 142,925.93
1.5% 0.843510 0.594641 0.287844 145,112.33
2.0% 0.847207 0.613142 0.324735 147,244.89

Relative to the 1.0% improvement scenario, the 1.5% assumption increases the expected present value by approximately 2,186.40 currency units, or 1.53%. Moving from zero improvement to 2.0% increases the displayed value by 8,814.59 currency units, approximately 6.37% of the zero-improvement value. These are scenario differences conditional on the artificial formula and discounting. They measure neither the predictive accuracy of a real mortality model nor the regulatory reserve change of a real insurer. The illustration demonstrates why a forecast that looks close on single-year mortality can have materially different implications over a 30-year annuity horizon.

This calculation was independently executed with a loop over j and k and checked by comparing the monotonicity of survival and present value across scenarios. All annual q values remain between zero and one for the illustrated horizon. The exact calculation can be replicated with the pseudocode in the reproducibility chapter; reporting the formula, schedule and rounding convention prevents a black-box numerical example from being mistaken for an empirical test.

Distinguishing valuation systems

The IFRS 17 standard is effective for annual periods beginning on or after January 1, 2023, and uses a current, risk-adjusted measurement of relevant future insurance cash flows. For a group of contracts, changes in assumptions and prospective services can have effects on the contractual service margin and recognition pattern; these cannot be inferred directly from a one-policy best-estimate present value. An appropriate financial backtest should preserve the actual reporting period, currency, discount curve, contract grouping, risk adjustment method and cash-flow classification. For comparability, present the pure mortality effect with the discount basis held fixed, followed by the full reporting-basis reconciliation.

In the European Union, the Solvency II longevity stress rule (Delegated Regulation (EU) 2015/35, Article 138) describes an instantaneous permanent 20% decrease in mortality rates for policies where that decrease raises technical provisions without the risk margin. This is a prescriptive capital stress, not a historical estimate of next-year deaths or a 20% probability assigned to a permanent mortality change. It should not be represented as a backtest score. The life underwriting risk module additionally recognizes distinct mortality, longevity and other underwriting risk components. A valuation analysis should identify the contract segment exposed to each direction of mortality change and the interaction with prescribed capital aggregation.

In the United States, the NAIC principle-based reserving overview describes the use of company-specific experience and prescribed requirements, including the transition to VM-22 for non-variable annuities. Application of any particular version or effective date requires the controlling valuation manual and relevant jurisdictional adoption. This monograph uses the official NAIC framework as contextual evidence and does not infer a company statutory reserve from the simple mortality illustration. The same restraint applies to IAIS supervisory principles, which frame reliable valuation and supervisory objectives without dictating one national annuity contract's accounting entries.

Liability attribution, capital and hedge basis risk

For a portfolio, calculate the difference between successive liability valuations and decompose it under a clearly specified order into realized experience, mortality-level basis, improvement assumptions, portfolio roll-forward, benefit changes, expenses, financial discount effects and management actions. Because nonlinear models have interacting terms, allocation of a cross-effect depends on the decomposition convention; report the ordering or use a symmetrical allocation if interpretation requires it. Do not attribute the entire movement in a recorded insurance liability to a longevity model just because mortality assumptions were updated in the same period.

A longevity hedge may reference a national cohort index, whereas the obligation is payable to a selected insured portfolio. Even a well-calibrated national forecast can leave mismatch between index and portfolio survival. Evaluate hedge effectiveness with joint scenarios for reference and annuitant populations and with the actual payoff terms of the hedge, not with population forecasts alone. The Cairns–Blake–Dowd analysis connects mortality-curve dynamics and parameter uncertainty to longevity-linked financial instruments; its subject matter clarifies why multi-year uncertainty matters to the finance of annuity promises.

A useful report therefore offers a mortality backtest, a contract-weighted benefit forecast reconciliation, sensitivity to alternative improvement assumptions, and the jurisdiction-specific financial measurement side by side. The four results answer different questions, and their joint interpretation is the valid form of actuarial materiality assessment.

Scientific Validity, Model Uncertainty, and Reproducibility

Data quality and revision risk

Source quality is part of forecast quality. An annual death count may be revised when late registrations are processed. Population exposure may change after census rebasing. A policyholder dataset can retain a valid contract ID but contain a changed commencement date, death-date adjustment or repeated amount record. A reproducible test must retain the original data snapshot, report the transformation from raw to analytic fields and reconcile counts to the contributing institution's published totals where possible.

The WHO Mortality Database documents the national data and technical file structure available for mortality research. The HMD source documentation details its death/population inputs and harmonization limits. These materials offer potential inputs, but the analytical dataset must choose a population, years and access permissions before any measured performance can be claimed. Even when original observations are accurate, choosing an updated series while claiming a historical forecast cutoff is a separate information-leakage problem.

The United Kingdom CMI_2025 methods record is especially revealing about data revision and timing. Its documentation explicitly notes that registration-date deaths and occurrence-date deaths yield different calibrations. A future validation file should record the specific denominator estimates and date convention used at each model release; otherwise a comparison may accidentally attribute data processing to genuine change in mortality trend.

Parameter, model and process uncertainty

For a stochastic mortality projection, three distinct uncertainty components require attention. Process uncertainty concerns actual future fluctuations conditional on a correctly specified model and fixed parameters. Parameter uncertainty concerns the limited precision with which those parameters were estimated from finite historical observations. Model uncertainty concerns differences among plausible specifications, constraints, age ranges and trend dynamics. These components interact: a forecast interval based only on a random period-factor innovation can be too narrow if parameter and structural uncertainty dominate at long horizons.

The Cairns–Blake–Dowd study explicitly incorporates parameter uncertainty. The OECD methodological comparison discusses fit, parsimony, and the representation of cohort effects. Their combined implication is methodological rather than a claim about which named model wins: the valuation purpose dictates which uncertainties must be preserved in the projected distribution. Where reliable historic out-of-sample scores are insufficient to rank models, a transparent alternative-scenario analysis communicates the limits of model choice more faithfully than an unjustified numerical probability assigned to each specification.

Proper scoring and interval evaluation

An illustrative probabilistic evaluation can employ the log score for deaths, a Brier score for individual death indicators, and the continuous ranked probability score for a continuous future liability or aggregate payment amount. A prediction interval should be assessed both for nominal coverage and width. A model can widen every interval until coverage appears adequate but thereby lose decision value; proper scores evaluate more of the forecast distribution. The Gneiting–Raftery scoring-rule treatment provides the statistical rationale. All score comparisons must use common observations and consistently identify whether larger or smaller values mean better prediction.

For sparse old-age age cells, exact binomial or explicitly pooled intervals can be preferable to approximate normal intervals. Dependency from common period shocks should be included when predicting total annuity outflows. Test multiple calibration levels: cell-level outcomes, age blocks, cohort totals and benefit-weighted portfolio sums. An insurer may have a large number of policy contracts but limited information on systematic annual improvement because all contracts share a short historical calendar span.

A frequency-versus-severity analogy helps prevent a measurement error. Death-event calibration affects the number and timing of benefit cessations. Benefit size and escalation affect the value of remaining payments. For a pure annuity, higher longevity increases the duration of benefit payment; for death protection, lower mortality can reduce a different liability. These directions are contract-specific and should always be demonstrated rather than inferred from an aggregate company mortality number.

Backtest audit trail

The following pseudocode is an illustrative replicable skeleton. It does not claim that proprietary annuitant histories have been analyzed.

for valuation_date in predeclared_forecast_origins:
    historical_snapshot = retrieve_frozen_vintage(valuation_date)
    assert latest_observation(historical_snapshot) <= valuation_date
    fit_or_restore_model(historical_snapshot, model_version, assumptions)
    for horizon in predeclared_horizons:
        forecast_deaths, forecast_survival = project_mortality(horizon)
        forecast_benefits = apply_frozen_contract_terms(forecast_survival)
        forecast_value = discount(forecast_benefits, frozen_discount_curve)
        if fully_observed(valuation_date, horizon):
            actual = reconstruct_later_observations_using_declared_rules()
            save_demographic_and_financial_scores(forecast, actual)
            save_predictive_interval_scores(forecast_distribution, actual)
compare_models_on_paired_vintages_with_dependence_aware_uncertainty()
retain_versions_and_exact_input_hashes()

For the synthetic numerical example in the previous chapter, a minimal independent replication is:

for g in (0.0, 0.01, 0.015, 0.02):
    survival = 1.0
    present_value = 0.0
    for j in range(30):
        q = 0.012 * (1.09 ** j) * ((1.0 - g) ** j)
        survival *= (1.0 - q)
        present_value += 10000.0 * survival / (1.03 ** (j + 1))
    print(g, present_value)

An executable implementation should retain exact library and runtime versions, source checksum, input tables, expected output tolerances, and documented random-number seeds for simulations. The ASOP 56 adopted standard provides a direct professional basis for reviewing inputs, validating output, monitoring assumptions and maintaining governance controls. The IAA international standards register situates model-related work alongside general actuarial practice, insurance accounting and enterprise risk analysis; applicability depends on the particular jurisdiction and assignment.

Negative controls and counterfactuals

Backtesting benefits from tests that a model should predict approximately as a simple benchmark would. One is a placebo revaluation in which a contract's guaranteed-period cash flows are isolated: mortality-model revisions should not change payments contractually guaranteed irrespective of life status over that period, subject to exact settlement rules. A second is a retrospective comparison at the original one-year horizon: errors that appear only after a software upgrade may reveal implementation drift rather than mortality movement. A third is an age- or country-matched baseline forecast with identical portfolio weights: its performance clarifies whether a sophisticated trend specification adds useful information beyond population selection and base-table alignment.

These diagnostic controls strengthen inference but are not causal experiments. A backtest can reject a forecasting specification under documented evaluation conditions; it cannot, by itself, prove that a single demographic mechanism generated the divergence. Particularly after pandemic disruptions, competing explanations must remain visible, and long-horizon claims must acknowledge the small number of genuinely independent forecast-origin periods.

Boundaries of the completed analysis

This dossier verifies source identities and directly described official findings, specifies backtesting formulas and implementation, and reproduces a hypothetical liability sensitivity calculation. It does not claim successful retrieval of each original commercial annuity contract, fully accessible paid data package, privately held experience cells, or long original historic source vintages. Therefore no actual A/E panel, out-of-sample model leaderboard, confidence band from observed experience, or causal population effect is reported. Such numerical claims require data and access evidence beyond the public materials described here. The limitation is a substantive guard against overclaiming, and it defines the next empirical research task precisely.

Integrated Results and International Research Agenda

Established findings from the evidence

Several conclusions follow directly from the reviewed literature and official documentation. First, stochastic mortality approaches encode different age, time and cohort structures, as shown by Lee and Carter's forecast specification and the older-age two-factor model. Their out-of-sample liability performance is an empirical proposition, not a property guaranteed by the model name. Second, insured annuity populations have distinct selection mechanisms: the UK policyholder study and the OECD mortality guidance together justify explicit demographic and product alignment.

Third, mortality assumptions are vintage-sensitive. The CMI_2025 update and the SOA 2026 improvement update illustrate that institutions respond to new death and exposure data through distinct data and methodological processes. Neither published model version supplies an empirically verified ranking for the specific annuity portfolio of this dossier. Fourth, financial materiality is not the same as mortality-model fit: IFRS 17, the EU longevity stress provision, and the NAIC reserving framework identify different reporting and supervisory objectives. The backtest must preserve these distinctions.

The illustrative calculation provides a fifth result of a different epistemic kind: under the declared synthetic death-rate curve and 3% discounting, shifting annual mortality improvement from 1.0% to 1.5% raises the value of 30 annual conditional payments by about 1.53%. That arithmetic is directly replicable and supports the sensitivity argument. It cannot be interpreted as an observed 1.53% under-reserving rate in any country or insurer.

A complete empirical implementation specification

A subsequent data-backed study should select a concrete, legally and statistically defined annuity segment: for example, an identified cohort of domestic single-life payout annuities with known annual benefits, entry ages, contract commencements, observed deaths, terminations, guaranteed periods and reliable exposure. The study should freeze multiple dated source snapshots before each forecast and reconcile revisions to deaths, exposures and contracts. It should compare a baseline period table, a population trend, a documented annuitant adjustment, and at least two stochastic model structures under paired and predetermined historical forecast origins.

The evaluation should publish four output layers. The demographic layer would report observed and expected deaths, standardized A/E, conditional age-cohort calibration and realized survival. The probabilistic layer would report proper scores and uncertainty coverage. The contractual layer would show forecast versus observed payments and exposure changes. The valuation layer would calculate the mortality component of liability change and then reconcile it to the jurisdiction's reporting or supervisory basis. Each layer should include the number of observed cells, excluded records, time horizon, estimation uncertainty and versioned source code.

The most consequential unresolved empirical question is how much incremental reserve prediction a complex stochastic mortality model provides after a simple base table has been correctly aligned with the annuity portfolio. Additional questions concern stability when model training incorporates pandemic years, differences between registrations and occurrences, the effect of source rebasings, socioeconomic segmentation, old-age mortality sparsity, and the sensitivity of rankings to benefit weighting. These are open tests within established scientific methods; the present analysis makes no claim to have answered them with proprietary experience.

International comparison with scientific limits

Comparative work should begin with a country-specific source inventory, not an unqualified aggregation. The HMD comparability rules supply a basis for many general-population comparisons. The OECD report describes how retirement and annuity table practices differ. The SOA industry series offers one route to U.S. annuity experience. UK CMI mortality projection resources supply model-vintage documentation. A harmonized research dataset should contain explicit mapping between national age conventions, exposure types, death-registration practices, benefit types and source revisions before model-score comparisons are calculated.

Special attention is necessary for amount-weighted exposure. A high-benefit contract may contribute disproportionately to the present value of liabilities, while count-based mortality tables weight the contract as only one unit. A model can improve per-life death prediction and worsen the amount-weighted liability forecast if it performs poorly for a small number of high-payment lives. Consequently the research report should retain both measures, state whether benefits are nominal or inflation-linked and document currency and discount conversions. Cross-country comparisons based on one shared discount curve should be labeled standardized illustrative comparisons; those reproducing actual balance-sheet values require each country's effective accounting and solvency rules.

Interpretation and authorial synthesis

The original synthesis of this monograph is the connection of information-preserving forecast evaluation with contract-specific annuity liabilities in one traceable research design. It identifies the precise places where apparent forecast accuracy can fail to measure the decision actually faced by an insurer: revised data can contaminate a historical forecast, population mortality can misrepresent selected beneficiaries, an aggregate A/E can conceal cohort offsets, and an accurate survivor count can still map to the wrong payment liability. The reasoning follows from existing mortality theory, published empirical evidence, proper scoring methodology and insurance valuation standards.

The practical scientific judgment is that a validation committee should read the mortality forecast, distributional diagnostics, annuitant basis and contract valuation together, while retaining the separate meaning of each test. The judgment is an application and synthesis of established scientific practice. It does not create a new doctrine, statistical model, named principle or claim of intellectual priority.

Conclusion

Longevity model backtesting for annuity liabilities is an ex ante comparison of historically feasible mortality forecasts with later survival experience and the corresponding contractual financial outcomes. Its scientific validity depends on frozen forecast information, representative exposures, the correct distinction between period and cohort mortality, genuine out-of-sample dates and evaluations that preserve dependency across time. Its actuarial validity further depends on the distinction between selected insured lives and the general population and on an explicit mapping from survival to actual payable benefits.

The reviewed evidence supports several concrete judgments. Lee–Carter and Cairns–Blake–Dowd models offer established but structurally different ways to forecast mortality. OECD guidance and annuity selection research require calibration to the retirement-income population under study. Current CMI and SOA releases show the importance of mortality-improvement assumption vintages, methodological changes and data revisions. Proper scoring rules strengthen evaluation of predictive distributions, while IFRS 17, Solvency II and United States reserving guidance locate longevity forecasts within distinct financial and supervisory decisions.

The hypothetical 30-year annuity calculation verifies the mathematical direction and materiality of a mortality improvement assumption change under controlled terms. It makes no empirical prediction about any actual insurer. The public original-source material reviewed here does not include the full vintage-controlled annuitant panel needed to demonstrate genuine historical superiority among competing longevity models; empirical performance remains a defined research objective with explicit input, reproducibility and uncertainty requirements.

The completed scholarly result is a source-traceable research synthesis, a reproducible backtesting specification and a verified synthetic liability sensitivity illustration. Its value is the precision with which it separates observable mortality outcomes, model forecasts and insurance obligations, preserving the evidentiary threshold for every quantitative and comparative claim.

References