RESEARCH DOSSIER
Extreme Old-Age Mortality and Annuity Payment Uncertainty
Abstract. Extreme old-age mortality is a consequential boundary between observation and extrapolation in the valuation of lifetime retirement income. Mortality beyond ages 100, 105, and 110 is observed through thin, age-sensitive population records whose selection and validation rules differ markedly from the experience underlying annuity pricing and reserving tables. The scientific question is therefore both demographic and financial: how much can be inferred about remaining life at exceptional ages, and how does uncertainty about that inference translate into the timing, value, dispersion, and financing of annuity payments? This monograph examines the history and interpretation of extreme-age hazard estimates, population and insurance data provenance, competing late-life mortality specifications, statistical identifiability, period-versus-cohort inference, mortality improvement, annuitant selection, contractual survival-contingent benefits, stochastic cash-flow valuation, portfolio pooling, and systemic longevity risk. It reviews the evidence for late-life mortality plateaus and the countervailing evidence on age errors, without treating the empirical debate as resolved across populations. Original analytical synthesis connects three distinct quantities: conditional survival probabilities, the present value of tail payments, and the covariance of liabilities across lives. A fully specified numerical illustration compares three hypothetical constant annual death probabilities from age 110 and analytically decomposes idiosyncratic and shared mortality uncertainty. The analysis shows why a small conditional extreme-age payout value may coexist with important model governance requirements; why the financial relevance of a tail model depends on survival to the tail, discounting, benefits, and the size of exposure; and why an annuity reserve must not be calibrated by copying a supercentenarian population estimate into a contract-specific mortality basis. International accounting, solvency, and actuarial guidance is interpreted within its institutional scope. The result is a source-based fundamental research monograph; no new authorial theory, proprietary model, empirical database, or externally reviewed finding is asserted.
Keywords: Extreme old-age mortality; supercentenarians; annuities; life-contingent cash flows; longevity risk; mortality plateau; mortality deceleration; age validation; mortality misstatement; Human Mortality Database; International Database on Longevity; annuitant selection; cohort effects; hazard function; truncation; right censoring; discounting; tail uncertainty; mortality improvement; model risk; pension liabilities; actuarial valuation; solvency; IFRS 17.
Introduction: Research Object, Questions, and Method
The duration of a life annuity remains uncertain until the relevant lives have died or the contract has otherwise terminated. An insurer may know the payment amount and frequency precisely while retaining uncertainty about the number of payments, their calendar placement, related expenses, and the accumulated financial cost. At the upper end of the human survival distribution, this ordinary life-contingent property becomes an exacting problem of data quality and extrapolation. Some mortality studies estimate near-constant mortality after roughly age 105 or 110, but the observed population is exceptionally selective and the effective sample is small. Contractual and financial conclusions therefore require a full bridge from the specific population, observation rule, and estimated survival process to the actual benefit formula.
The object of this dossier is the relationship between the distribution of deaths at exceptional ages and uncertainty in annuity benefits that remain payable because an annuitant or another covered life survives. The relevant scientific distinctions concern: (a) mortality intensity versus annual probability of death; (b) age-specific mortality among a population versus mortality conditional on annuity ownership; (c) individual randomness versus uncertainty about a common future mortality path; (d) valuation of expected payments versus risk and capital; and (e) biological longevity versus contractual terminal ages. These distinctions organize the empirical literature and determine the financial interpretation of any fitted tail.
This dossier uses an international documentary and analytical method. Primary materials are an official mortality-table repository, original mortality-data documentation and age-validation institutions, published demographic articles with distinct findings, an actuarial research study, an original actuarial modeling article, and the official texts or explanations of accounting, supervisory, and professional standards. The demographic studies are evaluated within their reported populations and observation windows. No restricted individual-level IDL extract, insurer proprietary experience file, or original longitudinal field study was obtained. Numerical results below are hypothetical computations independently reproducible from stated assumptions; none is presented as an estimate of current mortality or as an insurer's filed reserve.
The methodological reason for considering several evidence layers is direct. The International Database on Longevity documents and validates exceptional ages, while the Society of Actuaries' 2012 annuity reserving work concerns a regulatory insurance-mortality basis. These are different objects, generated under different sampling and purpose conditions. The 2018 Italian study and the 2019 data-error critique also lead to different interpretations of apparent late-life hazard flattening. A credible annuity valuation must preserve those differences rather than average them away rhetorically.
Historical, Mathematical, and Institutional Foundations
The movement from lifespan description to contingent valuation
Annuity mathematics transforms a distribution of future lifetimes into a schedule of expected survival-contingent payments. For a life aged x at valuation, the elementary problem is not to predict a person's exact date of death but to quantify the probability of receiving each payment under a specified contingent benefit. The actuarial present value is a weighted sequence of payments, with survival probabilities and discount factors determining the weights. The distinction matters acutely when a fixed contract is supported by a mortality table built from population experience many years before its final cash flows are due.
Individual survival has been central to economic accounts of annuitization. The original money's-worth study by Mitchell and colleagues shows why annuity valuation depends on the purchaser's mortality distribution and market pricing assumptions. Subsequent Poterba–Solomon annuity estimates, originally issued in 2021 and revised in October 2025, compare population and annuitant prospective mortality for immediate and deferred annuities and show that the applicable mortality basis can materially alter assessed payout value. Their headline values belong to specified U.S. markets, periods, payout forms, and discounting assumptions; they are neither universal annuity loadings nor mortality estimates at age 110.
The annual payout may be simple while the insurance undertaking is complex. A contract may pay while one life survives, after either one or both spouses die, during a minimum guaranteed period, with escalating amounts, or under a deferred commencement. Each design gives the tail of the lifetime distribution a distinct monetary weight. A guaranteed period can keep payments payable after death; a joint-life continuation can keep payments payable because a different life survives. The basic survival function must therefore be matched to the actual contract state space, not merely to the product label.
The population, the insured cohort, and the contract
Three observation units must be maintained. A national population record includes residents or persons covered by the recording jurisdiction, with its own reporting and migration conventions. A validated longevity case describes an authenticated person or event within a frame. An annuity portfolio contains insured lives with underwriting, wealth, occupation, health, purchase-date, and contract-selection characteristics. These units may overlap but need not share the same hazards.
The official International Database on Longevity documentation emphasizes systematic collection of deaths at age 105 or older and verification against birth and death records. Its international assembly improves reliability relative to anecdotal lists but does not itself create an annuity purchaser cohort. The Society of Actuaries' annuitant table is constructed for a different valuation role: its 2012 period male table incorporates margins and includes a prescribed final age. A period table, a cohort projection, a validated exceptional-age sample, and an in-force insurer portfolio can all produce plausible-looking late-life probabilities while answering different questions.
The annuity contract determines an additional layer of exposure. A death during a particular month may stop future monthly payments under one wording, leave the current payment unchanged under another, or trigger a death benefit or a survivor amount. The present investigation's main hypothetical calculations assume an annual life-only annuity payable in arrears while the covered person is alive. Real liabilities require the precise payment commencement, payability date, proof-of-life provisions, guaranteed terms, currency, indexation, and expenses in addition to mortality. Treating this simplified benefit as the contract language of a real insurer would be an analytical error.
Mortality measures and the survival process
Let T_x denote future lifetime after exact age x, S_x(t)=Pr(T_x>t) its survival function, and μ(x+t) the continuous force of mortality, when it exists. Then S_x(t)=exp(-∫0^t μ(x+u)du). For complete-year survival, write {}tp_x=Pr(T_x≥t) and q{x+t}=Pr(x+t≤age at death<x+t+1 | alive at x+t). A discrete annual recurrence is {}{t+1}p_x={}tp_x(1-q{x+t}), starting with {}_0p_x=1. Annual death probability and continuous force are distinct: under a constant force μ during one year, q=1-exp(-μ). An annual death probability of 0.5 corresponds to a constant within-year force ln(2)≈0.6931, rather than 0.5.
Period rates summarize experience across calendar-year age groups; cohort rates follow persons born in a given period as they move across calendar time. This difference is fundamental for an annuitant entering a payout stream now: the payer needs prospective probabilities along the cohort's future diagonal, whereas the latest observed period table is a snapshot of multiple birth cohorts. The Human Mortality Database Methods Protocol describes period and cohort rate construction, exposures, adjustments, and open-age treatment. Its open age 110+ should not be mistaken for observing separate yearly death probabilities to every possible higher age.
A single remaining life expectancy is insufficient to value a contract with a long tail. Two mortality distributions can have equal remaining expected lifetime but different probabilities of surviving beyond an advanced age; an escalating annuity or a delayed annuity can then have different values. The full joint distribution of benefit-contingent states, not merely the mean length of life, identifies payable cash flows.
Historical evidence at exceptional ages
The empirical literature contains results supporting mortality deceleration or near-constant hazard at advanced ages. Gampe (2010), using age-validated supercentenarian records and accounting for country-specific sampling frames, reported human mortality beyond age 110 consistent with approximately 50% probability of death per year. This estimate concerns its validated international sample and design, not a guaranteed universal rate in future cohorts.
Barbi et al. (2018) studied 3,836 documented Italian residents older than 105 across 2009–2015 and reported a plateau in extreme-age mortality beyond approximately 105, along with cohort-level decline. Rootzén and Zholud (2017) studied extreme-longevity records and drew inferences from their extreme-lifetime statistical analysis, including about an approximately 47% annual death probability at ages above 110 within their dataset. The absence of an identified finite lifespan limit in one model does not imply an appreciable probability of survival to arbitrarily distant future years. A finite cohort and a very small positive survival probability are distinct propositions.
These empirical conclusions have been disputed on measurement grounds. Gavrilov and Gavrilova (2019) analyze late-life errors and age misreporting and show why erroneous extreme ages can produce an apparent deceleration. Belzile, Davison, Rootzén, and Zholud (2021) revisit human mortality at extreme age through extreme-value and demographic analysis. Together the literature illustrates why validated ages, entry rules, and likelihood construction must precede pronouncements about a biological upper bound. The evidence remains study-specific even when estimates are superficially similar.
Advanced-Age Hazard Models
Gompertz, logistic, and plateau specifications
The Gompertz law represents adult-age mortality force as μ(x)=a exp(bx), with a>0 and b>0. It is simple, interpretable, and frequently useful within the age range supporting estimation. Extrapolation to ages with almost no exposure can force implausibly fast hazard growth when the underlying population shows deceleration, selection, data error, or genuine biological changes. A Gompertz fit at ages 65–90 is therefore not automatically reliable at ages 110–120.
A common decelerating shape is a logistic-type hazard, for example μ(x)=[a exp(bx)]/[1+c exp(bx)] for positive a,b,c. It approaches a/c as x grows, while its effective slope at intermediate ages depends on all parameters. Here the formula is an illustrative member of an established functional family, not a newly proposed mortality law. A piecewise alternative fits a conventional hazard to an age threshold and then applies a constant death probability q* beyond it. That simpler tail makes sensitivity transparent, although a discontinuous transition may be unsuitable in a production basis without documented smoothing and justification.
For model choice, data richness matters at least as much as curve elegance. Gompertz models have a few parameters and high extrapolation sensitivity; logistic models use a plateau parameter whose estimation is often weak when high-age deaths are sparse. A fixed plateau directly imposes a tail assumption. A flexible spline or semiparametric curve can describe observations but may be underidentified at the edge of the data. Competing fitted hazards should be evaluated by external validity, predictive scoring on supported ages, sensitivity at unobserved ages, and how they alter actual contracted payments.
The 2012 IAM period mortality table offers a useful caution: its male mortality probabilities are 0.38 at 105, 0.40 at 106–119, and 1.00 at age 120, reflecting an annuity reserving design and terminal convention. A terminal q=1 at 120 is a model closure, not independent empirical evidence that every person must die in that year. The documented 2012 table has its own margin and projection conventions; directly substituting its late-age numbers into a generic population study, or the reverse, erases their purposes.
Conditional survival and annuity present value
For a life-only annuity paying fixed amount B at the end of year t if the life survives, with time-specific discount factors D(t), the expected present value at age x is
when a contract basis uses terminal age ω. A corresponding open-tail valuation replaces the finite upper bound by a convergent infinite sum. If D(t)=v^t=(1+i)^{-t}, annual survival probabilities are level p=1-q for the relevant interval, and payments begin after one year, a perpetual conditional annuity value is B vp/(1-vp), provided vp<1. The extreme-tail component for an annuity already in payment at age x, beginning after year K, is
for a horizon of H annual payments after age x+K. The formula explicitly separates survival to the start of the tail, conditional payment survival after entry, and discounting. Inflation-linked or varying benefits replace B with B_t; non-annual payment periods require frequency-consistent probabilities and factors.
The economic interpretation follows immediately. A 10% variation in a contingent tail's value does not imply a 10% variation in the entire annuity liability. The portfolio materiality depends on the fraction of baseline total present value attributable to the tail. Conversely, a portfolio already containing many living centenarians has much greater sensitivity to their immediate conditional survival than a newly retired pool with few expected entrants to the tail.
Period, cohort, and forecast mortality
Historical period hazards cannot be read unchanged as future cohort hazards. Lee and Carter (1992) proposed an influential mortality forecasting specification expressing log age-specific rates through an age component, an age-sensitivity component, and a period mortality index. The forecasting power of that representation depends on estimation range, changing mortality regimes, and constraints on parameters. Its use above 105 requires a separate defensible tail treatment because the data supporting the period index are much stronger at ordinary adult and retirement ages.
Cairns, Blake, and Dowd (2006) developed a two-factor stochastic mortality model focused on older ages, with an age-sensitive factor affecting higher ages more strongly, and analyzed the relevance of both process and parameter uncertainty for longevity-linked finance. Their original application concerns post-60 mortality and longevity bonds. A model calibrated on ages around 60–90 can generate a prospective distribution of survival but should not be mistaken for directly measuring mortality above 110.
A coherent prospective assumption should be indexed by attained age, calendar year, sex or other admissible covariates, population, and valuation date. For an age-x person in calendar year y, write q(x+t,y+t|z,θ), where z is observed cohort information and θ describes model parameters. The survival probability to year t is the product of the corresponding one-year survival probabilities along the forward cohort path. Mortality improvements and periods of excess mortality change that path. The distinction between a historical fit and a prospective scenario should be visible in every model report.
Mortality improvement and structural change
Recent mortality histories demonstrate the danger of mechanical extrapolation from a short window. The Society of Actuaries' RPEC 2025 update states that post-pandemic evidence remained insufficient to issue an updated mortality improvement scale in 2025, although additional U.S. experience was analyzed. It also cautions that national actual-to-expected ratios may not apply to a specific retirement plan. This is a substantive caution about the data-generating process, not an endorsement of zero future improvement.
Generational forecasting and supercentenarian observations answer different questions. Mortality at age 110 today is observed among persons born more than a century earlier, with highly selected exposure histories; a person aged 65 today would reach 110 forty-five years in the future. Medical interventions, environments, recording systems, and selection could all differ. A stationary terminal hazard may be a practical valuation scenario, but its apparent transparency should not be confused with forecast certainty. Good governance uses scenario ranges and documents the sources of projected mortality improvements, reversals, and correlations.
Data Sparsity at High Ages
Age validation and observational frames
Extreme-age estimates can be dominated by a small number of records. If a person is incorrectly recorded as older than they are, they can contribute false survival exposure to ages where actual exposure is scarce. An erroneous age at death can also move the death count across several denominators. Record linkage from a credible birth certificate to an identity-consistent death record is therefore part of the statistical measurement, rather than a ceremonial verification step. The IDL validation protocol describes why only systematically observed and verified events support its research design and why national metadata differ.
The standard scientific controls are documentation of birth and death dates, identity continuity, inclusion criteria, calendar coverage, population boundary, and source-document provenance. For an insurer portfolio, a similar control applies to age at issue, date of birth, beneficiary identifiers, contract status, and payment records. Data quality has two directions: invented extreme longevity exaggerates survival; unregistered or miscoded deaths can distort exposures and apparent rates. An age-validation strategy should report exclusions and consider how missing documents may correlate with place, period, or observed longevity.
The original Italian investigation by Barbi et al. used a national and systematic frame rather than celebrity records. Its documented 3,836-case design helps address certain selection artifacts. Nevertheless the sample belongs to Italian cohorts born 1896–1910 and the observation years 2009–2015. Even strong internal age validation cannot automatically establish the hazard of future annuity buyers in another jurisdiction or birth cohort. External validity and record quality are separate properties of evidence.
Censoring, truncation, and exposure denominators
Right censoring occurs when a person remains alive at the end of observation, so the exact later death time is unknown. Left truncation occurs when an individual enters the observed study only after surviving to an eligibility age or inclusion date. If a death dataset includes only people who died in a particular national window, treating the resulting sample as a random cohort of entrants at age 105 may lead to a biased hazard estimate. Individuals alive before entry and deaths outside the covered calendar window must be handled according to the data collection design.
Under continuous-time survival analysis, with entry age a_i, observed exit age b_i, and event indicator δ_i, an illustrative likelihood contribution is proportional to [μ(b_i)]^{δ_i} exp[-∫_{a_i}^{b_i} μ(u)du], conditional on valid entry. Calendar-year and cohort-specific hazards require the analogous Lexis-time formulation. This expression assumes noninformative censoring given modeled covariates, correct event and entry dates, and a sampling scheme compatible with the likelihood. For some death-only frames, the proper likelihood has additional selection normalization; the illustrative formula cannot be applied blindly.
The key denominator is actual exposure to the risk of death at age and period, not the number of headlines or validated exceptional deaths. The HMD population methods show how deaths, population counts, Lexis cells, and exposures enter mortality-rate estimation. Its procedures also demonstrate why old-age rates can require extinct-cohort methods and adjustments. When mortality data are aggregated into an open 110+ cell, separate age-115 and age-120 hazards cannot be inferred without assumptions or more detailed records.
Model identification and competing empirical findings
A hazard plateau is an empirical proposition about how hazard or annual mortality changes conditional on survival to later ages. It is not identical to a biological ceiling or the possibility of an exceptionally long lifespan. Studies can disagree because they have different entry ages, ages verified, observation windows, countries, and statistical treatment of censored or truncated cases. At such small exposures, the difference between a continuous force and an annual probability also becomes numerically important.
The Gampe (2010) study found an approximate 50% annual death probability after 110 in a validated sample. The Rootzén–Zholud analysis reported about 47% under its frame. The Gavrilov–Gavrilova critique identifies data errors as a mechanism generating spurious deceleration. These should be compared by target estimand and data provenance instead of treated as votes for a single universal constant. The Belzile et al. reassessment broadens the methodological literature on statistical inference near the observed upper endpoint.
There are multiple sources of uncertainty. Sampling uncertainty arises because few individuals survive to the relevant ages. Identification uncertainty arises when models with different terminal behavior fit the observed age interval comparably. Age and coverage errors can introduce measurement bias. Structural uncertainty arises when the future mortality law differs from that in the historical sample. Parameter confidence intervals computed conditional on a single specified model generally do not cover all these sources. A well-calibrated sensitivity analysis therefore includes alternative tail families and data-quality assumptions rather than only sampling error bars around one fitted hazard.
Transferability to annuitants
Annuity buyers are selected by preference, resources, access, and survival expectations, and the in-force population may further be selected by observed longevity. Empirical studies of annuitization consequently compare prices under different mortality bases. The original Mitchell et al. (1999) research and the later Poterba–Solomon study both illustrate the substantial importance of distinguishing general-population mortality from annuitant experience for a payout product. Their studies are not direct estimates of supercentenarian annuitant hazards.
A responsible transfer procedure starts with credible policyholder experience at observed ages, compares it with appropriate external bases, estimates selection and duration effects where data permit, and assigns a separate, transparent terminal segment when actual experience becomes uninformative. Credibility techniques may blend portfolio and wider experience but cannot create high-age observations that do not exist. Covariates that improve fit at 70 or 80 may lose discriminatory power among the few people surviving to 110; any assumed persistence should be tested rather than presumed.
For a long-deferred annuity, the distribution of entrants to advanced ages is shaped by mortality at younger ages. An assumption change between ages 65 and 100 can have a larger financial effect than a dramatic change conditional on surviving to 110. Research attention should therefore follow each assumption's contribution to monetary liability as well as the scientific fascination of the extreme tail.
Late-Horizon Payout Uncertainty
A reproducible tail-payment illustration
Consider a hypothetical cohort of people who are each exactly age 110 at valuation and entitled to 20 possible annual life-only payments of one monetary unit each, payable at the end of years 1 through 20 if alive. Use a flat annual effective discount rate i=3%, a constant annual death probability q throughout the modeled interval, independence of deaths conditional on q, and no expenses, benefit escalation, taxes, guaranteed period, or payments after the final modeled year. These assumptions are solely designed to expose sensitivity; none estimates the actual age-110 mortality of any country or annuity portfolio.
For p=1-q and v=1/1.03, the value of one conditional liability is A(q)=Σ_{t=1}^{20} (vp)^t. The second moment is E[Y²]=Σ_{i=1}^{20}Σ_{j=1}^{20}v^{i+j}p^{max(i,j)} for the random discounted sum Y=Σ_{t=1}^{20} v^t 1{T≥t}; its variance is E[Y²]-A(q)². The identity is correct because a person receives both payment i and payment j exactly when they survive through the later of the two payment dates.
| Assumed annual q at ages 110–129 | Survival 5 years | Survival 10 years | Expected discounted tail payments per person, A(q) | Variance of one person's discounted payments |
|---|---|---|---|---|
| 0.40 | 0.077760 | 0.006046618 | 1.395321 | 2.986918 |
| 0.50 | 0.031250 | 0.000976563 | 0.943396 | 1.683346 |
| 0.60 | 0.010240 | 0.000104858 | 0.634921 | 0.970664 |
Every displayed number follows from the two finite sums with 20 terms, 3% discounting, and the stated q. The values are expressed in units of an annual payment of one; multiply by B to obtain a conditional value for any fixed annual benefit B in the same monetary unit. The difference A(0.40)-A(0.60)=0.760400 is economically meaningful conditional on already being alive at 110. Its effect on a 65-year-old's entire annuity is attenuated by survival and discounting to age 110.
For an equally hypothetical person aged 100, assume annual q=0.20 in each of the next ten years and q=0.50 thereafter for 20 annual payments beginning at age 111. The expected discounted value at age 100 of only those later payments is (0.8/1.03)^10 A(0.50)=0.075374. If the tail q after 110 is instead 0.40 or 0.60, keeping the first ten years identical, the tail values are 0.111481 and 0.050728, respectively. This computation quantifies how uncertainty in a conditional tail is filtered by the probability of reaching it and the time value of money. The first-ten-year payments, if any, are deliberately excluded from these tail-only amounts.
Portfolio variance and common mortality uncertainty
If N conditionally independent annuitants each have the same payment unit and death probability q, portfolio average discounted payments have expected value A(q) and variance Var(Y|q)/N. This is the familiar reduction of idiosyncratic longevity volatility through pooling. The variance of total benefits is N Var(Y|q), so the absolute standard deviation still grows as the square root of N. Pooling does not remove a shared change in future mortality assumptions.
To expose the distinction, suppose N=1,000 and a single common scenario is selected with equal prior probability: q=0.40 for all lives, or q=0.60 for all lives. Conditional on that scenario, lives are independent; unconditional on the scenario, future payment outcomes are correlated. The expected per-person discounted payment is (1.39532059+0.63492063)/2=1.01512061. The law of total variance gives
which here equals 0.00197879+0.14455202=0.14653082, giving a standard deviation of approximately 0.382793 per person in the portfolio average. For comparison, if q=0.50 is fixed with certainty and the 1,000 deaths are independent, the standard deviation of the average is sqrt(1.68334643/1000)=0.041029. The first figure is a deliberately extreme two-point shared uncertainty exercise and not an estimated predictive interval. Its scientific role is to show that uncertainty about a common mortality regime survives even in a large portfolio.
This decomposition corresponds to an established insight in longevity-risk finance: stochastic mortality includes aggregate factors as well as individual deaths. Cairns, Blake, and Dowd's two-factor mortality research emphasizes stochastic changes and parameter uncertainty over long horizons. The IFoA Longevity Risk Framework distinguishes systemic population risk from portfolio-specific risks. A reserve based on average life-contingent cash flows and a capital measure for adverse aggregate survival answer related but different questions.
Contract design and household consequences
For an annuitant, unexpectedly long life increases the value of dependable life-contingent income if the payer performs. For a life-only annuity issuer, longer survival tends to increase the number of payments. A joint-life annuity depends on joint survival and survivor benefits, requiring a joint mortality specification and, where relevant, dependence between lives. A guarantee period diminishes immediate death contingency for guaranteed installments, and a deferred income annuity concentrates valuation sensitivity on the probability of reaching commencement and surviving thereafter.
The sign and magnitude of tail exposure depend on the actual benefit. A contract paying a death benefit may have the opposite primary exposure to lower mortality. An escalating annuity, especially one indexed to prices or a fixed increasing schedule, can assign relatively high nominal or real weight to very late payments. Indexed benefits also require consistency between benefit escalation, inflation assumptions, and discounting. Household welfare analysis adds liquidity needs, possible bequest preferences, health expenditures, and contract counterparty risk, none of which can be derived solely from a mortality plateau.
The OECD 2025 pension indicators show that average remaining life expectancy at age 65 differs materially across countries and sexes and is forecast to change. These headline period/cohort summaries inform the international setting, but they do not identify the conditional hazard at 110 for insured purchasers. Retirement-income policy must pay attention to broad survival variation as well as statistical uncertainty in the remote tail. Extreme age is a specialized part of the general longevity-financing problem.
Valuation, capital, and financial reporting
The valuation task depends on the institutional framework. Under the IFRS 17 accounting overview, qualifying insurance groups are measured using fulfillment cash flows, risk adjustment for non-financial risk, and, where applicable, contractual service margin. IFRS 17 accounting is an integrated standard with recognition and presentation requirements; the hypothetical discounted payout A(q) above is one input to thinking about benefit cash flows and is not an IFRS 17 liability calculation. Expenses, premiums, contract boundaries, discount-rate requirements, grouping, reinsurance, risk adjustment, and changes in assumptions must be treated according to the applicable standard.
The IAIS Insurance Core Principles provide an international supervisory framework for valuation and insurer risk management. Their supervisory function differs from a national statutory formula or the contractual obligation to the beneficiary. Jurisdictional implementations may make different technical choices about margin, best estimate, capital, and permitted data. The same population mortality forecast therefore need not produce identical statutory reserves, economic liabilities, and solvency capital amounts across countries.
For a specific regulatory illustration, Solvency II Article 138 describes the life longevity-risk standard formula shock as an instantaneous permanent 20% decrease in mortality rates used in calculating technical provisions, for policies where such a decrease increases the relevant technical provisions. The 20% is a defined regulatory stress of mortality rates, not a universal increase of 20% in annuity cash flows or a statistical confidence bound for human lifespan. Its portfolio effect requires revaluation using the actual payment and financial assumptions. The health longevity provision has its own specified scope and should not be confused with this life sub-module.
U.S. professional pension practice likewise distinguishes assumption selection from an insurer's contract reserve basis. The current ASOP No. 27, effective for the stated qualifying reports and measurement dates from January 1, 2025, governs the selection of assumptions for specified pension and retiree group benefit obligations. Its discussion of demographic credibility, mortality improvement, and disclosure is valuable for research methodology, but it is a scoped standard and does not automatically prescribe the reserving basis for every commercial life annuity worldwide. ASOP No. 35 was repealed for covered work from 2025; reliance on its superseded status would confuse the current U.S. pension reference.
Scientific Validity and Sensitivity
Reconciliation of evidence and mortality assumptions
Before an extreme-age estimate enters a liability model, its provenance should be documented in a structured chain: source population; birth/death registration and age verification; period of observation; selection and censoring rules; unit of exposure; statistic reported (hazard, central death rate, or annual death probability); model family and estimation interval; confidence or uncertainty characterization; proposed extrapolation; and the annuity population to which the estimate will be applied. The chain makes visible the assumptions that separate external demographic evidence from contract experience.
The inference should be subjected to competing explanations. If a fitted hazard flattens after 105, ask whether the same pattern survives excluding poorly validated cases, altering observation windows, accounting for left truncation, changing the age threshold, and fitting competing models. If the plateau parameter cannot be estimated precisely from observable exposures, it should be a stress dimension. Uncertainty quantification should not present narrow within-model confidence intervals as though the model family and underlying measurement process were certain.
Temporal relevance also matters. A 2012 annuitant table, a 2018 Italian observation window, 2025 mortality-improvement commentary, and a 2026 valuation date are not interchangeable time stamps. The RPEC 2025 mortality update identifies the limitations of available post-pandemic trend evidence. A model reviewer must establish a consistent valuation date, data cutoff, cohort projection, and scenario definitions. Reusing an old table may be warranted by a binding statutory basis; reusing it as a scientific prediction requires independent justification.
Tail-closure conventions and stress testing
Every mortality table has a computational endpoint, open-ended terminal interval, or extrapolation convention. An age at which q is set to 1 enforces death within a year in that model; its purpose may be table closure for a designated use. The 2012 SOA period-table metadata explicitly documents its mortality cap and age-120 closure. These are modeling and reserving conventions. They do not establish a biological maximum or a directly validated death frequency at that age.
Stress analysis should vary the upper-age mortality level, the approach to that level, the forecast improvement rate, the terminal treatment, and dependence across lives. It should also change valuation yield curves and benefit profiles. A low-sensitivity result at age 65 can conceal substantial model uncertainty among the people already alive at 105. Conversely, a large conditional tail variation need not be material at the total-company level if exposures are tiny. Quantify both conditional extreme-age outcomes and unconditional total liability impacts; report absolute currency values and percentages against appropriate denominators.
The following sensitivity architecture follows established actuarial reasoning and is not a new theoretical model. First calculate a base case using the governed mortality assumptions. Next replace only the upper-age tail with several plausible conditional shapes, recording the age of transition. Then alter improvement, selection, and common risk assumptions, followed by discount and product scenarios. Finally assess financial materiality and evidence quality for the resulting range. Different assumptions should be changed separately where identification matters and jointly where interactions could create additional exposure.
Numerical replication and model audit
The numerical illustration above uses an annual time step and horizon H=20, p=1-q, v=1/(1.03), and a single benefit unit at each year-end conditional on survival. The following pseudocode is sufficient to replicate every conditional present value, survival figure, and single-life variance without proprietary data:
H, i = 20, 0.03
v = 1/(1+i)
for q in (0.4, 0.5, 0.6):
p = 1-q
mean = sum(v**t * p**t for t in range(1,H+1))
second = sum(v**(a+b) * p**max(a,b)
for a in range(1,H+1) for b in range(1,H+1))
variance = second - mean**2
print(q, p**5, p**10, mean, variance)
For the shared-risk example, compute μ0.4 and μ0.6 and their conditional variances, then use E[Var(Y|q)]/1000 + Var(E[Y|q]) with the two q scenarios equally weighted. This is a two-scenario illustration of uncertainty propagation, not a predictive credibility interval, risk margin, or an internal capital model. The conditional 1,000-person calculation assumes identical benefit amounts and no deaths correlated beyond the shared q state. More realistic work requires heterogeneous ages, policy characteristics, time-varying mortality, benefit types, and dependence structures.
A formal implementation audit should additionally test boundary cases q=0 and q=1, zero discounting, agreement between analytical sums and deterministic simulations, and invariance to regrouping of identical lives. Validate the definition of survival on the payability date and the exclusion or inclusion of the last year. Compare discounted expected payments from a cohort projection with an independent implementation using event-time survival simulation. Track code, input table versions, scenario metadata, checksum of the exact source file, and numeric tolerance. For any real portfolio, reconcile the number of contracts in the in-force extract to administration records and payments to ledgers before attributing discrepancies to mortality.
Jurisdictional and temporal boundaries
The international argument is that contingent payments are valued using appropriate survival distributions and financial assumptions, subject to actual contract obligations and governing rules. Its detailed legal, accounting, and capital realization is jurisdiction-specific. IFRS 17 applies where relevant accounting requirements mandate it; Solvency II capital provisions apply in their defined supervisory area and regulated entities; U.S. ASOP 27 applies to the professional work and dates within its scope. None of these authorities, independently, determines the law governing a particular annuity purchase in every jurisdiction.
The source review is not a systematic worldwide survey of pension law or insurer filings. Official HTML summaries of standards were compared with their published institutional descriptions, but subscription-only full standard texts were not assumed read. One Human Mortality Database methods PDF was consulted through its indexed text, while independent page-image confirmation was unavailable in the research interface; detailed PDF-only claims consequently remain constrained to the official text scope and are not presented as freshly audited empirical results. No proprietary experience data, individual IDL extracts, or external peer-review decision is represented as obtained.
Integrated Results and Future Research Agenda
Synthesized findings
The first result is a measurement result. Extreme-age mortality is a conditional distribution constructed from verified events and complete risk exposure within a defined observation frame. Its apparent behavior depends on survival selection and data procedures. The empirical findings of Gampe, Barbi and colleagues, Rootzén and Zholud, and the critical work of Gavrilov and Gavrilova together establish the need to examine the competing mechanisms behind a fitted high-age plateau. The scientific evidence supports explicit conditional assumptions and sensitivity, with the validity of each inference confined to its sample and design.
The second result is a valuation result. An annuity's late-horizon payment value factorizes into probability of reaching the tail, discounting to the tail, and conditional cash flows within the tail. This factorization supplies a transparent explanation for why extreme-age estimates can matter strongly to existing centenarian contracts but much less to some newly retired portfolios. The full liability also depends on the payment schedule, benefits after death, survival of other lives, expenses, and discounting; it is never identified by a single high-age mortality statistic.
The third result concerns risk allocation. Pooling reduces random deviations caused by independent deaths, but aggregate mortality improvement or parameter uncertainty can shift many payouts together. The hypothetical 1,000-life exercise makes the contrast calculable: a common two-state uncertainty dominates the portfolio average variance even though conditional idiosyncratic risk diversifies. This is an application of standard variance decomposition and established longevity-risk research, not the proposal of a new stochastic model.
The fourth result is institutional. A demographically plausible hazard is only one input to a governed valuation. Approved reserving tables, supervisory capital shocks, economic valuation assumptions, financial-reporting measurement, and actual contractual promises may require different but explicitly reconciled bases. The relevant standards show that source traceability, chosen assumptions, risk adjustment, capital, and disclosure belong to distinct professional and legal decisions.
The fifth result is methodological. Reproducible mortality valuation requires the demographic evidence chain and the contract cash-flow chain to join at an explicit prospective, population-specific mortality assumption. Neither highly detailed demographic record linkage nor sophisticated discounted cash-flow calculation alone establishes that join. The bridge is tested through population comparability, forecast assumptions, contract interpretation, exposure reconciliation, and sensitivity of the liability.
Research design for an empirical annuity portfolio
An empirical extension would begin with authorized anonymized policy records containing verified date of birth, issue date, coverage period, paid annuity amounts, payment frequency, sex or other lawfully relevant characteristics, geographic basis, and benefit form, together with independently audited dates of death and exposure. It would document the legal and data-governance basis for access. A population comparator could come from a transparent mortality source whose calendar and age definitions can be reconciled. Where exposure at age 105 or more is negligible, estimation would explicitly treat the terminal hazard as partially identified rather than silently fitting many high-age parameters.
The primary outcomes would include conditional survival rates, expected and distributional present values of cash flows, changes in reserves across upper-age model variants, and dispersion under systematic mortality scenarios. A credible design would use pre-specified model alternatives, out-of-sample validation at supported ages, competing selection specifications, and independent actuarial implementation. Sensitivity results would state both conditional values among survivors to advanced ages and whole-portfolio currency effects. This is a reproducible research agenda awaiting data, not a report of analyses already conducted on insurer microdata.
Limits, authorial contribution, and further questions
The original contribution of this monograph is an integrated, explicit, cross-disciplinary analytical synthesis of the source literature and institutional requirements, together with complete derivation and replication of a tail-sensitive cash-flow illustration. The work consolidates distinctions often dispersed across longevity demography, actuarial modeling, annuity economics, and solvency supervision. Its calculations are elementary consequences of accepted survival and probability theory. No priority claim is made for the variance decomposition, the annuity formula, a mortality-plateau hypothesis, or a newly invented actuarial framework.
Important unresolved research questions remain: how to estimate high-age annuitant-specific hazard with sparse exposures; how to distinguish genuine deceleration from record errors in multinational samples; how changes in longevity selection after purchasing an annuity affect the far tail; what method best incorporates uncertain mortality improvement beyond observed ages; and how to make population-to-portfolio basis risk visible in risk margins and capital while preserving comparability across jurisdictions. Addressing these questions requires new longitudinal data, explicit model evaluation, and replication, not further assertion from the existing summary statistics.
Conclusion
Extreme old-age mortality has a precise financial significance: it governs the conditional duration of payments to the rare annuitants who reach very advanced ages and contributes to the ultimate tail of larger retirement-income liabilities. The observational evidence is challenging because survival is rare, age errors can be influential, available records differ in selection, and several plausible hazard specifications can fit the limited upper-age data. The strongest available studies contribute genuine information about particular populations, while methodological critiques define the checks necessary before the results are transferred to insurance portfolios.
Life-contingent valuation turns these demographic issues into a sequence of survival-weighted payments. The expected tail value is determined jointly by access to the tail, conditional survival after entry, payment design, and discounting. Random individual longevity diversifies under pooling; uncertainty about the common future mortality regime remains. These propositions have been derived and numerically replicated in this dossier using fully declared assumptions, without attributing empirical status to the illustrative rates.
For an insurer, pension institution, or scientific analyst, the required practice is traceable. Establish the observed population and age-validation evidence, choose an explicit prospective mortality basis, reconcile the applicable payment terms, test competing extreme-age scenarios, quantify whole-portfolio materiality and systemic covariance, and interpret the results within the applicable accounting, solvency, and professional framework. This provides a rigorous research basis for annuity payment uncertainty at exceptional ages and a defined agenda for further empirical investigation.
References
Actuarial Standards Board. (2023). Actuarial Standard of Practice No. 27: Selection of Assumptions for Measuring Pension Obligations (adopted December 2023; applicable from January 1, 2025 under specified criteria). https://www.actuarialstandardsboard.org/asops/adopted-asop-no-27-selection-of-assumptions-for-measuring-pension-obligations/
Barbi, E., Lagona, F., Marsili, M., Vaupel, J. W., & Wachter, K. W. (2018). The plateau of human mortality: Demography of longevity pioneers. Science, 360(6396), 1459–1461. https://pmc.ncbi.nlm.nih.gov/articles/PMC6457902/ DOI: 10.1126/science.aat3119
Belzile, L. R., Davison, A. C., Rootzén, H., & Zholud, D. (2021). Human mortality at extreme age. Royal Society Open Science, 8(9), 202097. https://pubmed.ncbi.nlm.nih.gov/34631116/ DOI: 10.1098/rsos.202097
Cairns, A. J. G., Blake, D., & Dowd, K. (2006). A two-factor model for stochastic mortality with parameter uncertainty: Theory and calibration. Journal of Risk and Insurance, 73(4), 687–718. https://onlinelibrary.wiley.com/doi/10.1111/j.1539-6975.2006.00195.x
European Insurance and Occupational Pensions Authority. (2015/2024 online text). Commission Delegated Regulation (EU) 2015/35, Article 138: Longevity risk sub-module. Solvency II Single Rulebook. https://www.eiopa.europa.eu/rulebook/solvency-ii-single-rulebook/article-5758_en?prefLang=lv
Gampe, J. (2010). Human mortality beyond age 110. In H. Maier, J. Gampe, B. Jeune, J.-M. Robine, & J. W. Vaupel (Eds.), Supercentenarians (pp. 219–230). Springer. https://www.demogr.mpg.de/en/publications_databases_6118/publications_1904/book_chapters/human_mortality_beyond_age_110_3980/
Gavrilov, L. A., & Gavrilova, N. S. (2019). Late-life mortality is underestimated because of data errors. PLOS Biology, 17(2), e3000148. https://doi.org/10.1371/journal.pbio.3000148
Human Mortality Database. (2025). Methods Protocol for the Human Mortality Database, version 6, August 5, 2025 (J. R. Wilmoth, K. Andreev, D. Jdanov, D. A. Glei, and T. Riffe, with collaborators). https://mortality.org/File/GetDocument/Public/Docs/MethodsProtocolV6.pdf
Institute and Faculty of Actuaries. (n.d.). Longevity Risk Framework: Mortality Research Hub. https://www.actuaries.org.uk/learn-and-develop/research-and-knowledge/mortality-research-hub/longevity-risk-framework
International Association of Insurance Supervisors. (2026 online access). Insurance Core Principles and ComFrame Online Tool, particularly valuation principles. https://www.iais.org/icp-online-tool/
International Database on Longevity. (2026 online access). International Database on Longevity: Methods, validation, coverage, and access. Institut national d'études démographiques and contributing institutions. https://supercentenarians.org/en
International Financial Reporting Standards Foundation. (2026 online access). IFRS 17 Insurance Contracts: Issued standard overview and effective-date information. https://www.ifrs.org/issued-standards/list-of-standards/ifrs-17-insurance-contracts/
Lee, R. D., & Carter, L. R. (1992). Modeling and forecasting U.S. mortality. Journal of the American Statistical Association, 87(419), 659–671. https://www.tandfonline.com/doi/abs/10.1080/01621459.1992.10475265
Mitchell, O. S., Poterba, J. M., Warshawsky, M. J., & Brown, J. R. (1999). New evidence on the money's worth of individual annuities. American Economic Review, 89(5), 1299–1318. https://www.aeaweb.org/articles?id=10.1257%2Faer.89.5.1299
OECD. (2025). Pensions at a Glance 2025: OECD and G20 Indicators, chapter: Life expectancy. OECD Publishing, Paris. https://www.oecd.org/en/publications/pensions-at-a-glance-2025_e40274c1-en/full-report/life-expectancy_02c88784.html
Poterba, J. M., & Solomon, A. (2021; revised October 2025). New Evidence on the Money's Worth of Immediate and Deferred Individual Annuities. NBER Working Paper No. 28557. https://www.nber.org/papers/w28557
Rootzén, H., & Zholud, D. (2017). Human life is unlimited—but short. Extremes, 20, 713–728. https://link.springer.com/article/10.1007/s10687-017-0305-5
Society of Actuaries. (2011–2012). 2012 Individual Annuity Reserving Report & Table. Society of Actuaries Research and Payout Annuity Table Team. https://www.soa.org/resources/experience-studies/2011/2012-ind-annuity-reserving-rpt/
Society of Actuaries. (2012 period basis; certified December 2013). 2012 IAM Period Table—Male, Age Nearest Birthday (table ID 2585), Mortality and Other Rate Tables. https://mort.soa.org/ViewTable.aspx?TableIdentity=2585
Society of Actuaries Research Institute, Retirement Plans Experience Committee. (2025). RPEC 2025 Mortality Improvement Update (October 2025). https://www.soa.org/resources/research-reports/2025/rpec-mort-improvement-update/