Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Stochastic Mortality Forecasting

Abstract. Stochastic mortality forecasting estimates a probability distribution over future age-specific death rates and their associated survival paths. For insurers, its decisive object is the joint distribution of mortality experience over ages, calendar years, birth cohorts, and populations. This monograph examines the mathematical and statistical architecture of such forecasts, beginning with reliable death and exposure data and extending through Lee–Carter age–period models, Poisson likelihoods, two-factor adult-mortality specifications, cohort adjustments, coherent population projections, forecast uncertainty, backtesting, and long-duration insurance applications. It derives the identification constraints of classical models, distinguishes process, parameter, estimation, and structural uncertainty, and gives an auditable numerical illustration of horizon-dependent forecast dispersion. It then develops a scientific evaluation sequence connecting data vintage, model selection, interval calibration, held-out experience, population transfer, and liability cash flows. The analysis finds that stochastic sophistication is meaningful only when the mortality quantity, future-age path, exposure definition, distributional assumptions, and target insured population are aligned. Long-horizon prediction intervals can expand materially even when expected improvement is smooth; model agreement at a single age does not establish agreement on cohort survival, annuity liabilities, or aggregate mortality risk. Original contributions consist of source-grounded comparative synthesis, algebraic derivations, illustrative calculations and actuarial interpretation, without proposing a new named theory or claiming empirical re-estimation of inaccessible datasets.

Keywords: Stochastic mortality forecasting; mortality improvement; Lee–Carter model; two-factor mortality models; Cairns–Blake–Dowd; Poisson regression; mortality rate; cohort effects; parameter risk; model risk; longevity risk; forecast distribution; mortality backtesting; actuarial science; annuity valuation.

Introduction

Mortality forecasts are claims about the distribution of future death experience conditional on historical information, modeling choices, and a defined population. The word stochastic specifies that the projection contains random future variation and an explicit probability structure. It does not by itself establish that all material uncertainty has been quantified. A model may generate thousands of sample paths while excluding uncertainty about its drift parameter, age pattern, structural breaks, data quality, or transfer from population mortality to a selected insurance portfolio.

The scientific origin of modern parsimonious forecasting is Lee and Carter's (1992) age-specific mortality projection study. The central innovation was to represent a matrix of log death rates using a common time index and age-specific sensitivities, then forecast that index as a time series. The subsequent literature demonstrates why alternatives are needed: death counts have heterogeneous exposures, mortality improvement changes with age, cohort experiences can be distinctive, elderly rates require careful treatment, and the random variation of time indexes is only part of long-term prediction risk. The OECD's international actuarial assessment places this statistical work in the wider setting of pension and annuity population selection, future improvements, and public oversight.

This dossier uses six analytically connected stages. First, construct defensible data and measurement definitions. Second, fit and understand the single-index Lee–Carter structure. Third, evaluate alternative two-factor, cohort, and coherent multi-population specifications. Fourth, propagate process, parameter, model, and population-transfer uncertainty to distributional outcomes. Fifth, test models on historical forecasts actually made using information available at each origin. Sixth, translate future age-specific paths into the survival and payment quantities relevant to life insurers, annuities, reserving, and capital.

Mathematical formulas are derived for clarity. Numerical figures later in the text are synthetic illustrations using declared parameters, not empirical estimates for a country or insurer. The referenced primary articles establish the methods and their reported empirical settings. Contemporary institutional reports establish their own bounded findings as of the research cutoff, rather than universal mortality laws.

Mortality Time-Series Data

The estimand and its units

A death count is not a mortality rate. Let D(x,t) be deaths at completed age x during calendar year t and E(x,t) be exposure to risk, conventionally expressed in person-years under a central-rate approach. The observed central death rate is

mx,t=Dx,tEx,t.m_{x,t}=\frac{D_{x,t}}{E_{x,t}}.

Its units are deaths per person-year. The probability of death between exact ages x and x+1, denoted q(x,t), is a different quantity. For a constant force of mortality mu within that interval,

qx,t=1−exp⁡(−μx,t).q_{x,t}=1-\exp(-\mu_{x,t}).

The approximation q approximately equals m is most accurate at low rates and under compatible exposure and hazard assumptions. Models fitted to log central death rates cannot automatically be substituted into a logit probability equation. The chosen scale changes the likelihood, sensitivity at old ages, and permissible forecast domain.

The Human Mortality Database overview describes how national death registration, population counts, reconstructed exposures, and life-table calculations are harmonized. This explicit separation matters because an unexplained change in denominators can resemble a mortality trend. A data user must retain population definition, sex classification, age convention, calendar boundary, version date, completeness, and any extrapolation or closure at advanced ages. Year of death and birth cohort are related by c=t-x under a consistent age convention, but calendar-year versus age-at-last-birthday grouping can complicate cohort assignment.

A central age–period model assumes an observed panel over ages x in X and years t in T. A survival projection for a person now aged a instead follows the diagonal x=a+j, t=t0+j. That diagonal is the contractual quantity. A forecast of mortality at one fixed age in successive future calendar years is insufficient to calculate the survival of an aging cohort.

Counts, likelihood, and the observation process

The simplest exposure-based count model is

Dx,t∣Ex,t,μx,t∼Poisson⁡(Ex,tμx,t).D_{x,t}\mid E_{x,t},\mu_{x,t}\sim\operatorname{Poisson}(E_{x,t}\mu_{x,t}).

The conditional mean and variance both equal E times mu. The Poisson log-bilinear formulation of Brouhns, Denuit, and Vermunt (2002) embeds Lee–Carter-type systematic structure in this likelihood and thereby weights sparse cells differently from high-exposure cells. A Poisson count equation is a modeling assumption; overdispersion, correlated shocks, reporting artifacts, and dependence across cells can challenge its fit. For a bounded death probability measured over lives followed completely through an interval, a binomial likelihood is also natural. A population exposure reconstructed from migration and period estimates is not literally the identical statistical object as a cohort of individually followed insured lives.

The distinction between expected deaths and observed deaths becomes critical near old-age tails. A rate of 0.05 inferred from 20 expected deaths has much greater sampling uncertainty than the same rate inferred from 20,000 expected deaths. Equal-weight least squares on logged observed rates can give very influential weight to noisy observations unless graduation, exposure weighting, or an explicit likelihood is used. A zero observed count cannot enter an ordinary log-rate without treatment; adding an arbitrary small constant changes the implicit statistical model. An analyst should choose a count likelihood or justified correction, record it, and examine sensitivity.

Annual series, short shocks, and vintages

Long-run forecasts depend on whether a year's unusual experience is interpreted as temporary mortality shock, a level change, or a permanent change in improvement. The HMD's Short-Term Mortality Fluctuations data provide a distinct weekly evidentiary resource for short-horizon mortality changes. Weekly counts need comparable denominators, seasonal handling, dates of occurrence versus registration, and a revision record. Their latest years are provisional in many systems; weekly patterns cannot simply be stitched into an annual historical time series without reconciling timing and exposure. A current publication date is not a substitute for a frozen data vintage.

The statistical unit of comparison also matters. A general-population series may reflect public-health interventions, migration, socioeconomic composition, or reporting improvements. An insured-life portfolio differs in underwriting, attained-age mix, policy duration, benefit size, lapsation, geography, occupation, and distribution channel. The OECD retirement-income analysis emphasizes matching baseline assumptions to the population whose payments are modeled. Thus population forecasts supply a reference improvement process, while portfolio-level calibration still requires an explicitly justified bridge.

Data architecture for reproducibility

A defensible mortality research dataset preserves the source and retrieval date; years and ages used; exposure method; final versus provisional status; exclusions; treatment of zero rates; age closure; sex and region definitions; and train/test cutoffs. It also preserves both raw death/exposure counts and derived rates, with transformation code or equations. As a result, a second researcher can rebuild the input panel without silently replacing a past information set with revised observations.

For international comparisons, harmonization is valuable only where registered deaths and population exposures meet adequate completeness. The HMD's documented inclusion conditions impose substantial coverage constraints. Extrapolating the same model to settings with weak vital registration demands additional measurement-error and demographic-reconstruction analysis. Reporting a narrow stochastic interval conditional on an unreliable denominator communicates precision the observed data do not support.

Lee–Carter Framework

Age–period factorization

In the classical age–period specification,

log⁡mx,t=ax+bxkt+εx,t,\log m_{x,t}=a_x+b_x k_t+\varepsilon_{x,t},

where a(x) is the age profile of mean log mortality, b(x) measures the age-specific response to the period index, k(t) is the time-varying overall level of mortality, and epsilon is a residual. Lee and Carter (1992) estimated a low-rank representation of the historical log-rate matrix using singular value decomposition and projected the period index through a random-walk time-series model. The rank-one structure is attractive because a high-dimensional age–year panel is summarized by one common time factor and one vector of age sensitivities. Its interpretability is also a restriction: systematic changes outside that single direction cannot be captured by the core index.

Identifiability requires constraints because scaling and shifting the index can be absorbed into the age parameters without changing fitted mortality. A common convention is

∑xbx=1,∑tkt=0.\sum_x b_x=1,\qquad \sum_t k_t=0.

To verify the symmetry algebraically, consider any nonzero scalar c and constant d. The transformation b'_x=b_x/c, k'_t=c(k_t-d), and a'_x=a_x+b_x d yields a'_x+b'_x k'_t=a_x+b_x k_t. Constraints select a representative parameterization; they do not add information about future mortality. Different conventions can be translated, and forecasts must consistently transform all parameters.

The original workflow has at least three modeling operations: estimate age coefficients and index; optionally calibrate the fitted period index to reproduce an aggregate feature of death experience; then estimate a time-series law for the index. Confusing the calibration step with the time-series forecast can make published parameter values apparently incompatible. Rates reconstructed from a rank-one fit need not reproduce every annual death count exactly.

Time series and forecast laws

A common stochastic continuation is

kt=kt−1+d+ηt,ηt∼(0,σk2),k_t=k_{t-1}+d+\eta_t,\qquad \eta_t\sim (0,\sigma_k^2),

with mutually independent innovations for the simplest formulation. Conditional on known d, sigma and the current index,

E[kT+h∣FT]=kT+hd,Var(kT+h∣FT)=hσk2.\mathrm E[k_{T+h}\mid\mathcal F_T]=k_T+hd,\qquad \mathrm{Var}(k_{T+h}\mid\mathcal F_T)=h\sigma_k^2.

The corresponding conditional log-rate at fixed age x has a forecast mean a_x+b_x(k_T+hd), with period-innovation variance b_x^2 h sigma_k^2 when residual and parameter uncertainty are excluded. This is a conditional statement, not the total uncertainty of a real forecast. If the fitted drift d itself has sampling variance s_d^2, an elementary independent-uncertainty approximation adds b_x^2 h^2 s_d^2. Growth proportional to h squared explains why long-horizon parameter uncertainty may become larger than annual innovation uncertainty.

A random walk with drift encodes a specific future extrapolation of historical improvement. If b(x)>0 and d<0, the fitted log mortality rate declines at the annual pace b(x)|d|. That pace could change as age-specific medical progress, epidemics, or population composition evolve. A stationary ARIMA formulation, damped trend, or alternative process must be assessed on its forecast performance and substantive plausibility. Assuming the drift estimate remains exactly correct over fifty years would be a much stronger claim than estimating its value from a finite historical sample.

Count likelihood, residuals, and aging

Brouhns, Denuit, and Vermunt's (2002) exposure-based mortality regression models the stochastic counts directly, while respecting the log-linear factorization of the mortality intensity. Under a Poisson fit, Pearson residuals and deviance residuals can reveal age–year structures left unexplained by the common period factor. Parallel diagonal bands in the age–year residual heatmap suggest cohort effects; alternating stripes at fixed calendar years suggest unmodeled period shocks. These are diagnostic patterns, not conclusive evidence of their causes.

Life-table quantities require a fresh conversion of each simulated mortality surface into age-specific survival. On a piecewise-constant hazard assumption, the survival probability from age a through n successive years is

npa,T=exp⁡(−∑j=0n−1μa+j,T+j).{}_{n}p_{a,T}=\exp\left(-\sum_{j=0}^{n-1}\mu_{a+j,T+j}\right).

For annual death probabilities estimated directly, the corresponding expression is the product of (1-q_{a+j,T+j}). The two formulas are consistent under q=1-exp(-mu) but should not be mixed with an unannounced approximation. Taking the expected value of each year's rate and then calculating survival from that deterministic average generally differs from averaging survival across joint stochastic paths.

Interpretation and scientific limits

The Lee–Carter model's principal advantage is a transparent common-factor interpretation with a probabilistic extrapolation. It also has an empirical track record and a reproducible baseline function. The Lee–Miller historical evaluation documented forecast errors that differ by horizon, including imperfect interval calibration, and Booth, Maindonald, and Smith (2002) found that Australian data could require a modified calibration period. These findings show that estimation-window selection and stability of the age response are part of forecast validity, rather than technicalities reserved for model fitting.

A single age–period index may be well supported for a broad population while a financial contract is highly sensitive to an exceptional age range or cohort. Accordingly, goodness of fit for all ages is not sufficient evidence that the forecast distribution is fit for every insurance purpose.

Two-Factor Mortality Models

Adult mortality level and slope

A two-factor model introduces more flexibility in how mortality changes with attained age. The foundational Cairns–Blake–Dowd (CBD) form for older adults can be written as

logit⁡(qx,t)=kt(1)+(x−xˉ)kt(2),\operatorname{logit}(q_{x,t})=k_t^{(1)}+(x-\bar x)k_t^{(2)},

where logit(q)=log(q/(1-q)) and x-bar is a fixed centering age. The first factor moves log-odds of mortality at every modeled age equally, while the second changes their slope across ages. Its origin in the analysis of post-60 mortality and parameter-sensitive longevity pricing is established in the CBD two-factor investigation. The centering age is a numerical and interpretive choice. The model does not assert that the older-adult age slope is linear across childhood or all human ages.

The two factors are stochastic and may be correlated. A basic vector random walk,

kt=kt−1+d+ηt,Cov(ηt)=Σ,\boldsymbol{k}_{t}=\boldsymbol{k}_{t-1}+\boldsymbol d+\boldsymbol{\eta}_t,\qquad \mathrm{Cov}(\boldsymbol{\eta}_t)=\Sigma,

produces h-step conditional covariance h Sigma when parameters are fixed. The covariance of predicted logits at ages x and y is

Cov(zx,T+h,zy,T+h)=h wx⊤Σwy,wx=(1,x−xˉ)⊤.\mathrm{Cov}(z_{x,T+h},z_{y,T+h}) =h\,\boldsymbol w_x^\top\Sigma\boldsymbol w_y, \quad \boldsymbol w_x=(1,x-\bar x)^\top.

Consequently, a distribution of mortality at one age cannot be projected independently of the distribution at another without discarding the systematic factor dependence. When the financial object is joint survival across many years, factor correlations determine cash-flow tails.

The CBD model's logit link guarantees 0<q<1 for finite logits and is attractive at advanced ages. The Lee–Carter log-rate link guarantees positive rates but does not constrain a rate to be below one, which is appropriate for a force/central rate and would be wrong if interpreted as a probability. These domain properties follow directly from the mathematical links; they do not establish empirical superiority.

Cohort extensions and identifiable structure

Calendar period and birth cohort capture different historical patterns. Period effects change mortality simultaneously at many ages during a given year. Cohort effects follow individuals born in a particular year as they age. A generic extension can include

g(mx,t)=ax+bxkt+γt−x,g(m_{x,t})=a_x+b_x k_t+\gamma_{t-x},

subject to model-specific specifications, constraints, and identification conventions. Renshaw and Haberman (2006) study cohort-sensitive mortality reductions. Adding a cohort function can improve representation of diagonal residual patterns but creates substantial inferential demands. The three indices x, t, and c=t-x are linearly dependent; unrestricted linear trends cannot be assigned uniquely to age, period, and cohort. Constraint choices therefore influence component interpretations even when fitted surfaces are unchanged.

A cohort term should be forecast, not merely carried forward without a law. An estimated cohort deviation for birth years already observed does not tell the model what unobserved future cohorts will experience. At old ages, sparse early cohorts and changing population definitions can make long-horizon extrapolation unstable. Cohort patterns can also be confounded with age misstatement, migration, or registration practices.

The eight-model empirical comparison by Cairns and colleagues (2009) is important because a model that wins an in-sample criterion can have weak parameter stability. It identified robustness concerns for one cohort-rich formulation and showed that fit and forecast suitability answer different questions. Plat (2009) offers a broader stochastic model construction with attention to cohort effects and uncertainty. Neither result licenses a single global ranking across all ages, periods, populations, and insurance cash-flow functionals.

Coherent populations and compositional changes

Sexes, regions, countries, and socioeconomic subpopulations may share common historical forces while retaining genuine differences. Independent univariate extrapolations can lead to implausibly increasing relative differences when strongly related populations have shared drivers. Li and Lee's (2005) coherent multi-population extension combines a common trend with population-specific deviations so their long-run divergence is limited under the model. Hyndman, Booth, and Yasmeen (2013) develop a product-ratio coherent forecast using functional time-series structure for common and relative components.

Coherence is a modeling restriction on relative mortality trajectories, not an empirical fact that every difference must disappear. Permanent differences in baseline mortality, unequal health conditions, or population-specific policies can remain. If an insurer sells in several countries, shared drivers may transmit correlation across mortality assumptions, while local effects create basis risk. A globally identical trend would erase meaningful heterogeneity; mechanically independent national trends would ignore common causes. The relevant objective is a joint distribution whose dependence structure is defensible.

Choosing the right number of factors

Adding factors raises in-sample flexibility but also increases the number of stochastic laws, cross-covariances, parameter estimates, and identification restrictions. A two-factor model may capture movement in both level and slope, yet miss exceptional cohort effects. A richer model may approximate historical experience precisely, yet produce unstable long-term out-of-sample curves. The model-structure review in the comparative actuarial literature demonstrates why complexity needs an observable performance rationale.

Model selection should therefore examine fit by age and cohort, residual autocorrelation, prediction interval behavior, rolling-origin accuracy, parameter stability, cohort survival, and financial outputs. It should include explicit competing approaches: a parsimonious one-factor benchmark, an adult two-factor specification, an age–period–cohort alternative if indicated, and a coherent structure where cross-population comparison is fundamental. The model selected for pricing term-life death claims need not be the model selected for long-term annuity longevity risk.

Forecast Uncertainty

Distinct uncertainty sources

Predictive uncertainty has separable origins that should be disclosed, even where exact additive decomposition is impossible.

Innovation or process uncertainty concerns as-yet-unrealized mortality changes given the fitted stochastic process. Sampling uncertainty comes from the limited deaths and exposures used to infer the mortality surface. Parameter uncertainty concerns drift, covariance, age response and other fitted quantities. Model uncertainty concerns competing statistical structures and assumptions about the future. Data uncertainty concerns incomplete counts, revisions, and population definitions. Transfer uncertainty concerns applying general-population forecasts to insurance policyholders. Event and regime uncertainty concerns shocks outside the historical stochastic law.

Let Y be a future mortality-derived outcome and theta an unknown parameter vector estimated from information D. The law of total variance gives

Var(Y∣D)=Eθ[Var(Y∣θ,D)∣D]+Varθ(E[Y∣θ,D]∣D).\mathrm{Var}(Y\mid D) =\mathrm E_\theta[\mathrm{Var}(Y\mid\theta,D)\mid D] +\mathrm{Var}_\theta(\mathrm E[Y\mid\theta,D]\mid D).

The first term is conditional predictive variation averaged over parameter uncertainty; the second is variability in the conditional prediction resulting from unknown parameters. A nested distribution over model classes M adds another layer. This identity holds when the stated conditional distributions and posterior or uncertainty measures are well-defined; it is not a universal proof that one estimated model captures every relevant structural break.

Kleinow and Richards (2017) investigate parameter risk in mortality forecasts with direct implications for insurance regulation and report that some components can be financially material. Their result corrects the intuition that thousands of process paths automatically represent uncertainty about estimated long-term improvement. Reusing a single drift and a single age sensitivity across simulations is conditional simulation, however large the number of draws.

A reproducible thirty-year illustration

Consider a synthetic Lee–Carter projection at one fixed attained age with today's central rate m0=0.010 deaths per person-year, sensitivity b=0.020, expected annual index drift d=-1, annual innovation standard deviation sigma_k=0.60, and drift-estimation standard error s_d=0.12. Assume the innovation sequence is independent and Gaussian, the drift estimate independent of future innovations, and all quantities are in the same index normalization. These parameters are chosen solely to expose sensitivity; they are not fitted mortality estimates.

Over h=30 years, the expected change in log mortality is b h d = -0.60. The conditional median rate ratio is exp(-0.60)=0.5488, giving an indicative fixed-age median rate 0.005488. Conditional innovation variance on the log-rate scale equals

b2hσk2=(0.020)2(30)(0.60)2=0.00432.b^2h\sigma_k^2=(0.020)^2(30)(0.60)^2=0.00432.

The associated standard deviation is 0.06573. A symmetric normal 95% interval on the log scale transforms to an asymmetric rate-ratio interval approximately [0.4825, 0.6243]. It is a conditional, process-only interval.

Now add independently estimated drift uncertainty. The drift contribution is

b2h2sd2=(0.020)2(30)2(0.12)2=0.005184.b^2h^2s_d^2=(0.020)^2(30)^2(0.12)^2=0.005184.

The combined log-rate variance is 0.009504, with standard deviation approximately 0.09749. The corresponding 95% interval for the fixed-age rate ratio becomes approximately [0.4534, 0.6644]. The widening arises even though the median rate remains unchanged. Moreover, for lognormal mortality rates, the arithmetic predictive mean ratio under this mixture is exp(-0.60+0.009504/2), approximately 0.5514, exceeding the median 0.5488. Reporting only the central path would obscure both tail dispersion and the distinction between mean and median.

The purpose of this calculation is the horizon scaling: innovation variance grows in h, while drift-estimation variance grows in h squared under the stated approximation. It does not imply drift uncertainty must dominate in every empirical sample; covariance between parameter estimates, time-series model choice, non-Gaussian innovations, and changing trend regimes can alter the result. The example is a fixed-age future rate. The death probability for a current sixty-five-year-old thirty years later involves a different attained age, not this fixed-age calculation.

Predictive distribution of survival and liabilities

Simulate a common future mortality surface path for each draw, then walk each birth cohort through successive attained ages and calendar years. For a fixed starting age a, a cash-flow simulation should evaluate all relevant survival probabilities jointly. Expected survival is

E[exp⁡(−∑jμa+j,T+j)],\mathrm E\left[\exp\left(-\sum_j\mu_{a+j,T+j}\right)\right],

which generally differs from exp of minus the sum of expected hazards. Jensen's inequality supplies one direction of this distinction when the exponential is applied to a stochastic accumulated hazard; the exact financial effect still depends on benefit timing, discounting, and joint correlations. For death-contingent insurance, the same path should determine each year's survival-to-start and death-during-year probabilities, avoiding independently sampled inconsistent death events.

Where mortality is modeled with parameter draws or several candidate specifications, draw those uncertainties at the appropriate hierarchy: one model/parameter realization may govern many future years and insured lives, while individual deaths are conditional on shared systematic paths. Drawing a new uncertain drift independently for every policy or year would artificially diversify a common trend risk. Both common shocks and sampling variation belong in the predictive distribution, but at distinct aggregation levels.

Tail behavior and model risk

Forecast quantiles inherit assumptions about innovation tails, parameter distributions, link functions, and treatment of extreme events. A Gaussian period-index random walk may give analytically convenient bands while underrepresenting pandemics, climate-health hazards, conflict, or correlated shocks. The functional-data alternative of Hyndman and Ullah (2007) incorporates robust age-curve forecasting, a useful response to anomalous years, although robustness to outliers is a different objective from estimating the probability of future extremes. Excluding shocks from estimation may stabilize a secular trend and simultaneously omit material mortality catastrophe scenarios. Scenario and stress analysis should therefore complement a statistical baseline.

A predictive distribution can be summarized by means, medians, quantiles, expected shortfall where appropriate, and joint liability measures. Each answers a different decision question. An interval calibrated at each age separately need not attain the nominal simultaneous coverage for a portfolio or a 30-year path. Tail calibration and dependence must be tested at the aggregation used to make decisions.

Backtesting and Period Effects

Reproducing the information set available then

Backtesting asks what a method could have predicted at an earlier calendar date, using the information available at that date. Choose successive forecast origins T1, T2, and so forth. At each origin, freeze the accessible dataset vintage, fit parameters on years no later than that origin, select the allowed model according to a predeclared rule, and forecast horizons h in an agreed set. Compare these forecasts with later realized experience, preserving the distinction between final revised death counts and information originally available to the forecaster.

The original Lee–Carter forecast evaluation by Lee and Miller (2001) demonstrates why in-sample fit cannot stand in for genuine forecasting evidence: calibration varies across horizons and historical episodes. A rolling-origin design should record training windows, whether model tuning uses test information, and whether a single unusual year disproportionately changes estimated drift. Reporting only the window with the best subsequent fit is hindsight selection.

Distributional diagnostics

Mortality forecasting requires evaluation of both point accuracy and the predictive distribution. At the cell level, bias and absolute error in log rate can be calculated, with reasonable exposure weighting. At the count level, the logarithmic predictive score is the negative log probability assigned to the observed death count under the model's forecast mixture. Lower mean log score indicates a better probabilistic prediction under the specified evaluation design. Proper scoring rules discourage a model from claiming artificially narrow confidence intervals.

Prediction intervals can be checked for empirical coverage at each horizon and age group and for interval width. A nominal 95% interval with 95% coverage can still be badly calibrated in the tails, or can achieve coverage by being far wider than necessary. Coverage across many related ages cannot be treated as independent repeated evidence when all depend on a common period index. The evaluation should aggregate by origin and consider block methods or uncertainty estimates that preserve temporal dependence.

A simple constructed diagnostic illustrates the difference between observed deaths and projected mean deaths. If a held-out age–year cell has exposure 100,000 person-years, the forecast central rate is 0.008, and 900 deaths are observed, then the expected count is 800 and the simple Poisson Pearson residual is (900-800)/sqrt(800)=3.54. This calculation is arithmetically checkable, but its nominal reference distribution assumes that the forecast mean is known and count variation is Poisson. Under a stochastic mixture with trend uncertainty, the predictive variance is larger; the simple residual must not be presented as a final backtesting significance test.

Period effects, structural change, and the pandemic

A long-term trend inferred from a particular historical segment can be sensitive to start and end dates. Booth, Maindonald, and Smith's (2002) variable-decline analysis shows how alternative periods can alter the forecast age profile and expected life expectancy. This supports a straightforward falsification test: compare forecasts fitted using several historically defensible training windows and test their stability on later periods. If a conclusion relies on one chosen window, that dependency is a finding to report.

Large mortality shocks pose a layered identification problem. An abnormal year may contain excess deaths concentrated at particular ages, subsequent mortality displacement among frail populations, changes in reporting, and genuinely persistent changes in health risks. For an annual stochastic model fitted on one-factor log mortality, a pandemic can appear as a large period-index innovation; a more persistent alteration may require an intercept, drift, cohort, or regime revision. The mathematical form alone cannot determine which interpretation is epidemiologically correct.

The 2026 evidence is especially instructive. The Society of Actuaries Retirement Plans Experience Committee's October 2026 mortality assessment reports that age-standardized U.S. mortality at ages 65 and above over July 2025–June 2026 was 98.7% of a stated MP-2021-based expected level for males and females combined, while declining to issue a replacement projection scale because robust evidence on the long-run trend remains insufficient. This statistic is conditional on that comparator and age standardization; it is not a universal population mortality probability or proof that a permanent post-pandemic trend has been identified.

A separate 2026 Mortality Improvement Model update describes updated tools and data, while keeping MP-2021 distinct from the MIM-2026 model update and excluding 2020-onward experience from its prospective trend-fitting framework. The coexistence of updated data and a retained projection basis is coherent: an actuary can observe new mortality levels without claiming that the long-term rate of future improvement has been re-estimated reliably. A forecast paper should report which of these decisions it makes rather than conflating new observations, revised software, and a new secular projection scale.

Comparing models fairly

The quantitative actuarial model comparison of Cairns and colleagues demonstrates the possible conflict between statistical fit and parameter stability. A fair validation program should compare at least a transparent benchmark, a suitable multi-factor model, and a model with cohort effects when justified. It should harmonize fitting ages and years, exposure definitions, outlier policy, number of parameters, and forecast horizons. Comparisons should include predicted survival and liabilities, since small log-rate errors at irrelevant ages can dominate a crude all-cell score while materially important senior-age errors are hidden.

A backtest can also suffer evaluation-period selection. Strong historical results in a small number of countries, a narrow time window, or similar policy environments cannot establish global performance. Coherent models should be tested for whether common-factor restrictions improve cross-population forecasts; a single-population score is insufficient to judge their defining objective. Methods to reduce in-sample complexity, including BIC, need not maximize forward predictive performance. Where predictive distributions matter, probability scores and interval diagnostics should be treated as primary evidence.

A practical audit trail

The reporting record for each candidate should contain: observation unit; data vintage; training and holdout years; age and population coverage; link function; likelihood; identifiability constraints; model-selection rule; parameter law; simulation method; random seed if simulation is used; number of forecast paths; horizon-specific scores; interval calibration; sensitivity to shock-year treatment; parameter stability; and insurance-valued outcomes. An analyst who can reproduce a forecast mean but not its stochastic uncertainty has reproduced only part of the model.

Long-Horizon Insurance Implications

From future rates to contractual outcomes

A stochastic mortality forecast is used by an insurer through contractual cash-flow projections, not through the rate surface alone. Consider a term-life death benefit B paid at year-end following a death during policy year j. For a policy aged a at valuation time T, conditional on one future mortality path, the probability of payment in year j is the product of survival through earlier policy years and death probability during year j:

Pj=(∏i=0j−2(1−qa+i,T+i))qa+j−1,T+j−1.P_j=\left(\prod_{i=0}^{j-2}(1-q_{a+i,T+i})\right)q_{a+j-1,T+j-1}.

For deterministic discount factors v_j, the pathwise present value of expected death benefits over n years is the sum of v_j B_j P_j. The full stochastic expected value averages this quantity over the joint mortality scenarios, and any uncertainty in benefit amounts and discount rates requires its own consistent modeling. For a life annuity with annual payments while alive, the corresponding pathwise expected value is a sum of discounted survival probabilities; longer survival increases expected annuity payments, all else equal. For death insurance, lower mortality generally reduces near-term expected death claims; the timing and duration of coverage can modify the value comparison.

These algebraic results explain why mortality improvement can have opposite direct cash-flow signs for life insurance and annuities. An insurer with both exposures may receive some natural offset, but hedge effectiveness depends on age, cohort, policy type, underwriting selection, and timing. A perfect offset cannot be inferred from a common general-population trend.

Population transfer and basis risk

An actuarial model may be calibrated to the general population because it has extensive national historical data, then applied to a portfolio with a distinct mortality level and different trend response. The bridging assumption must be documented: for example, a multiplicative level ratio at each age, a relative improvement factor, or a hierarchical population model. Brouhns, Denuit, and Vermunt (2002) discuss adapting mortality projections to annuitants; the OECD analysis of pensioner assumptions explains why retirement-income populations need representative baselines.

A fixed portfolio multiplier m_port(x,t)=r_x m_pop(x,t) implies the same proportional improvements as the source population. That is a testable structural restriction, not a consequence of multiplication. If r_x varies over future time due to underwriting practices, product mix, changing socioeconomic composition, or lapse selection, mortality basis risk persists even with an accurate general-population projection. Basis risk matters for longevity-linked instruments and reinsurance where the reference population and insured liabilities are different.

A coherent international population model can mitigate implausible divergence between related national baselines while still failing to explain product-level selection. The coherent group forecast of Li and Lee (2005) addresses group-level dynamics; it should not be mistaken for an insured-portfolio experience study. For the same reason, a report claiming an annuity capital benefit from a country forecast must identify the exact portfolio assumption being changed.

Insurance valuation, risk margins, and capital

Economic valuations are functions of discounting, expected cash flows, and risk allowances. Under IFRS 17 insurance-contract measurement, fulfilment cash flows are measured on a current, risk-adjusted present-value basis using relevant information. This accounting framework does not prescribe Lee–Carter or any single mortality model, nor does it replace jurisdiction-specific solvency regulation. The operative scientific task is to understand how mortality assumptions enter claims and benefits, the distribution of their variation, and the entity's actual risk-adjustment methodology.

For a portfolio of N conditionally independent lives exposed to a shared systematic mortality factor K, a simplified aggregate count C has

Var(C)=E[Var(C∣K)]+Var(E[C∣K]).\mathrm{Var}(C)=\mathrm E[\mathrm{Var}(C\mid K)]+\mathrm{Var}(\mathrm E[C\mid K]).

If lives are homogeneous conditional on K with annual death probability p(K), the conditional variance term is N E[p(K)(1-p(K))] and the systematic term is N squared Var[p(K)]. This latter contribution does not vanish by increasing portfolio size. In real portfolios, heterogeneity and dependence must be handled explicitly, but the scaling result identifies the economic difference between diversifiable individual death randomness and shared mortality-trend risk. Thus portfolio size alone cannot eliminate long-horizon longevity uncertainty.

For a life insurer, longevity risk is prominent when promised payments continue while survivors live. For mortality protection, an abrupt increase in deaths can dominate short-horizon claims. A model calibrated only to smooth secular improvements is incomplete for catastrophe accumulation. The appropriate capital distribution may need distinct shock scenarios in addition to stochastic trend models. Kleinow and Richards's insurance-oriented parameter-risk analysis explains why ignoring uncertainty in trend estimates can distort capital results.

Duration, convexity, and common shocks

A long-duration annuity is exposed to the cumulative survival path over many future years. A small persistent change in mortality improvement can alter many payments at once. Because survival and present value are nonlinear functions of the entire mortality surface, the expected liability under uncertain mortality is not generally equal to the liability obtained by simply substituting expected future rates. Discount-rate movements may also be correlated with macroeconomic conditions that influence mortality, requiring a joint economic and demographic simulation where relevant. If discounting is fixed for illustration, the model should say so.

The risk can be examined through conditional derivatives without implying a universal monetary sensitivity. Under a constant force for each year, survival is S_n=exp(-sum mu_j), so the partial derivative with respect to a specific future hazard mu_j is -S_n for j within that horizon. Reducing any one hazard raises survival. An annuity payment at or after that horizon responds to the survival change; an earlier death benefit responds differently. With uncertainty in discount factors, the cross-scenario covariance between survival and discounting becomes relevant.

The same mortality scenario should drive each member of a cohort and each payment period. Re-sampling mortality independently for each annuitant destroys the common factor on which hedging and capital depend. The two-factor longevity-risk study explicitly connects correlated mortality curve movement with longevity-linked pricing and parameter uncertainty. Real insurance users must additionally document contract boundaries, mortality basis, and any pricing measure distinct from the real-world forecasting measure.

Institutional judgement and international application

International guidance distinguishes statistical comparability from direct transferability. The OECD's 2023 report surveys mortality-table development and urges representative baselines and appropriate projections. The 2026 U.S. retirement update offers a more recent national institutional judgement, contingent on its own data and population. Neither supplies a global default process for every insured-life contract. A scientific insurance dossier must report local population characteristics and effective actuarial and accounting requirements where a particular application is made.

For implementation, a sound internal decision record states the model's scientific purpose, target lives, source population, choice of projection law, scenario dependence, uncertainty components, observed out-of-sample diagnostics, governance signoff, and date of the underlying information. A quantitative result becomes institutionally useful when its scope is traceable, especially where two reasonable models yield materially different long-term capital or reserve conclusions.

Integrated analytical synthesis

The progression from a death count to a capital or premium consequence has a sequence of transformations: counts and exposures define an observed rate; a model estimates an age–period–cohort mortality surface; a stochastic forecast supplies a joint distribution over future surfaces; each path becomes cohort survival and death events; contracts map events to payments; discounting and dependence produce liabilities and risk measures. Each transformation can introduce an additional source of uncertainty or mismatch. Accuracy at the first transformation cannot certify correctness at the last.

This sequence yields three analytically supported conclusions. First, the most relevant forecast evaluation depends on the user's liability functional: a senior-age annuity book and a young adult term-life book emphasize different ages and horizons. Second, accurate average mortality improvement is not sufficient evidence of calibrated tail outcomes, because systematic factors and uncertain drifts persist across years. Third, credible international application requires separately documenting population selection and measurement infrastructure. These conclusions are applications of established actuarial, demographic and statistical theory, not a newly named model or doctrine.

Conclusion

Stochastic mortality forecasting is a scientific method for constructing conditional distributions of future mortality surfaces and the survival paths they imply. The Lee–Carter framework establishes a transparent parsimonious benchmark, while Poisson log-bilinear estimation aligns the fit with death counts and exposures. Two-factor and cohort extensions capture additional age and birth-cohort structure; coherent models account for shared population dynamics. Their additional expressiveness requires more disciplined parameter identification, time-series estimation, and forecasting validation.

The decisive scientific property is not the sheer number of simulated scenarios but the credibility and calibration of the underlying joint predictive distribution. The monograph's thirty-year illustration shows how adding even a single drift-parameter uncertainty term widens forecast intervals through horizon-squared scaling. Historical backtesting, information-set integrity, period sensitivity, shock handling, and proper distributional scores are therefore integral to model verification. Contemporary 2026 actuarial reports illustrate why observing mortality's recent return toward a baseline does not automatically justify a newly estimated long-term improvement scale.

For insurance applications, the modeled rate surface must be translated consistently into the diagonal paths of aging cohorts and the associated contractual cash flows. Death benefits, annuities, capital, and longevity hedges can respond differently to the same mortality shift. General-population extrapolations require explicit insured-life and annuitant calibration; systematic trend uncertainty cannot be diversified away merely by pooling a larger number of lives.

The original analytical synthesis of this dossier consists of a common inferential chain from data and identifiability through predictive uncertainty and backtesting to liability outcomes, supported by verified foundational literature and a reproducible synthetic numerical example. No new named model, theory, proprietary mortality dataset, or unpublished empirical forecasting result is asserted.

References

  • Lee, Ronald D., and Lawrence R. Carter. 1992. “Modeling and Forecasting U.S. Mortality.” Journal of the American Statistical Association 87 (419): 659–671. Source record.

  • Cairns, Andrew J. G., David Blake, and Kevin Dowd. 2006. “A Two-Factor Model for Stochastic Mortality with Parameter Uncertainty: Theory and Calibration.” Journal of Risk and Insurance 73 (4): 687–718. Source record.

  • Human Mortality Database. 2026. “Overview.” University of California, Berkeley, and Max Planck Institute for Demographic Research, accessed November 14, 2026. Source record.

  • Brouhns, Natacha, Michel Denuit, and Jeroen K. Vermunt. 2002. “A Poisson Log-Bilinear Regression Approach to the Construction of Projected Lifetables.” Insurance: Mathematics and Economics 31 (3): 373–393. Source record.

  • Renshaw, Arthur E., and Steven Haberman. 2006. “A Cohort-Based Extension to the Lee–Carter Model for Mortality Reduction Factors.” Insurance: Mathematics and Economics 38 (3): 556–570. Source record.

  • Li, Nan, and Ronald Lee. 2005. “Coherent Mortality Forecasts for a Group of Populations: An Extension of the Lee–Carter Method.” Demography 42 (3): 575–594. Source record.

  • Lee, Ronald, and Timothy Miller. 2001. “Evaluating the Performance of the Lee–Carter Method for Forecasting Mortality.” Demography 38 (4): 537–549. Source record.

  • Booth, Heather, John Maindonald, and Len Smith. 2002. “Applying Lee–Carter under Conditions of Variable Mortality Decline.” Population Studies 56 (3): 325–336. Source record.

  • Hyndman, Rob J., and Md. Shahid Ullah. 2007. “Robust Forecasting of Mortality and Fertility Rates: A Functional Data Approach.” Computational Statistics & Data Analysis 51 (10): 4942–4956. Source record.

  • Cairns, Andrew J. G., David Blake, Kevin Dowd, Guy D. Coughlan, David Epstein, Alen Ong, and Igor Balevich. 2009. “A Quantitative Comparison of Stochastic Mortality Models Using Data from England and Wales and the United States.” North American Actuarial Journal 13 (1): 1–35. Source record.

  • Plat, Richard. 2009. “On Stochastic Mortality Modeling.” Insurance: Mathematics and Economics 45 (3): 393–404. Source record.

  • Kleinow, Torsten, and Stephen J. Richards. 2017. “Parameter Risk in Time-Series Mortality Forecasts.” Scandinavian Actuarial Journal 2017 (9): 804–828. Source record.

  • Hyndman, Rob J., Heather Booth, and Farah Yasmeen. 2013. “Coherent Mortality Forecasting: The Product-Ratio Method with Functional Time Series Models.” Demography 50 (1): 261–283. Source record.

  • Organisation for Economic Co-operation and Development. 2023. Mortality and the Provision of Retirement Income. Paris: OECD Publishing. Source record.

  • Society of Actuaries Research Institute, Retirement Plans Experience Committee. 2026. RPEC 2026 Mortality Improvement Update. October 22. Source record.

  • Stern, Larry N., et al. 2026. Overview of the 2026 Updates to the Mortality Improvement Model (MIM). Society of Actuaries Research Institute, June 23. Source record.

  • IFRS Foundation. 2026. “IFRS 17 Insurance Contracts.” Issued Standards, accessed November 14, 2026. Source record.

  • Human Mortality Database. 2026. “Short-Term Mortality Fluctuations (STMF).” Updated October 26, 2026. Source record.