RESEARCH DOSSIER
Mortality Improvement and Longevity Assumptions
Abstract. Mortality improvement assumptions determine how an observed age-specific mortality surface is carried forward into future calendar years and, consequently, how the survival of insured lives, annuitants, and pensioners is valued. This monograph develops the actuarial distinction between an observed decline in mortality, a prospective assumption about that decline, and the economic sensitivity of a liability to the assumption. It derives discrete and continuous improvement measures; explains the roles of period, birth cohort, attained age, and exposure; examines statistical models and their identifiability; and compares the empirical and institutional choices underlying contemporary mortality projection practices. The analysis incorporates current evidence through November 14, 2026, including the Human Mortality Database methodology, the Continuous Mortality Investigation's CMI2025 framework and CMI2026 consultation, and the Society of Actuaries' 2026 retirement and insured-life assessments. It gives a fully specified, reproducible forty-year hypothetical annuity and life-insurance calculation to quantify the direction, magnitude, and nonlinear accumulation of improvement assumptions. The findings establish that calibration of a base mortality table and selection of a future improvement trajectory are inseparable valuation decisions; observed period recovery following an extraordinary mortality shock does not, by itself, identify a durable future cohort trend; and transfer of population projections into insured and annuity portfolios requires explicit treatment of selection, socioeconomic composition, credibility, and exposure. The resulting synthesis describes a rigorous assumption-setting practice grounded in traceable data, clear valuation conventions, model comparison, numerical validation, and transparent sensitivity reporting. The monograph presents original analytical interpretation and calculations; it introduces no proprietary theory, doctrine, or forecast for any real insurance portfolio.
Keywords: Mortality improvement; longevity assumptions; actuarial science; survival analysis; period mortality; cohort mortality; Lee–Carter; age–period–cohort models; mortality projection; annuity valuation; life insurance; longevity risk; mortality selection; model risk; Solvency II; IFRS 17; CMI; Society of Actuaries.
Introduction
Mortality improvement is the temporal change in mortality intensity, rates, or probabilities for a specified age and population. Longevity assumption setting is the act of selecting a prospective age-and-calendar-time mortality surface for contracts whose cash flows depend on survival or death. The two tasks meet in the mapping of observed experience into future liabilities. They have different evidentiary burdens: an improvement statistic can be measured from exposure-adjusted past counts, while a future improvement assumption rests on a forecast, a judgment about structural change, or an explicit blend of the two.
A projected death probability for a person age 65 in a future year cannot be inferred from a life expectancy reported for the population in a single calendar year. The Human Mortality Database overview distinguishes original death and exposure data, derived rates, and life-table probabilities. The 2024 U.S. mortality report reports period life expectancy and death rates; those observations do not constitute the complete mortality pathway of a living cohort. The 2025 OECD longevity comparison further demonstrates why country, sex, age, and period matter to the description of a seemingly universal development.
A sound mortality study distinguishes four objects. The historical data surface consists of observed deaths and exposed lives. The graduated base table describes a chosen valuation population at a specified base date. The improvement surface advances that base table across future time. The valuation functional converts the resulting death and survival probabilities into cash flows under a product-specific financial basis. Differences at any of these stages can explain divergent results, even where two institutions use the same name for a mortality model.
The research question is how these objects should be related when mortality falls over long periods, stalls or reverses in shorter periods, and evolves differently across cohorts. The method combines exact actuarial derivations, a comparative reading of source publications, a reproducible deterministic numerical experiment, and institutional analysis. Model equations are presented for interpretive and assumption-setting purposes; estimation of an actual country-specific coefficient or fitting of a live portfolio is outside the empirical sample available to this monograph. Contemporary institutional statements are dated and identified as current at November 14, 2026. All original numerical examples are explicitly hypothetical and were computed independently of the cited historical series.
This research examines longevity risk through the construction and consequences of improvement assumptions. Specialized calibration algorithms, insurer-specific experience studies, reinsurance structures, and product regulation are discussed only where they change the interpretation of improvement. The authorial contribution is the integrated explanation of where demographic measurement, projection, and insurance valuation meet, together with the mathematical sensitivities and governance implications developed below.
Mortality Improvement Metrics
Mortality measures and the exposure denominator
Let D(x,t) be deaths at completed age x during calendar year t, and E(x,t) the corresponding central exposure in person-years. The central death rate is m(x,t)=D(x,t)/E(x,t). A death probability q(x,t) represents the chance that an individual alive at exact age x dies before age x+1 under the applicable period or projected cohort mortality conditions. The survival probability is p(x,t)=1−q(x,t). A continuous force of mortality μ(a,s) is a hazard at exact attained age a and calendar time s. These are related, but numerically distinct, objects.
Under a constant force over the one-year age interval, q(x,t)=1−exp[−m(x,t)] when m also represents that force. A more general life-table conversion depends on the distribution of deaths within the year: q=m/[1+(1−a_x)m], where a_x is the expected fraction of the year lived by those dying in the interval, under the usual life-table approximation and unit interval. Thus, for m=0.020, constant-force q is approximately 0.0198013, while applying the simplistic identity q=m introduces a small but cumulative difference. The HMD Methods Protocol, version 6 details the reconstruction of exposure, deaths, and period/cohort life tables; its August 2025 revision also addresses infant-death timing and monthly birth exposure.
The denominator matters as much as the numerator. A death count can decline while the death rate rises if exposed lives fall faster. Open-ended oldest-age cells, migration, revised population estimates, mortality registration delay, and changes in the composition of the exposed population can alter apparent improvement. An A/E ratio, calculated as observed deaths divided by expected deaths for the same exposure and specification, is a measure of current fit to a baseline; its change is not automatically a direct measure of the structural future improvement rate.
Arithmetic, geometric, and logarithmic improvement
For the same age x in consecutive years, define an arithmetic mortality improvement rate on a consistently specified death-rate basis by
[ I_{x,t}=1-\frac{m(x,t+1)}{m(x,t)}. ]
A positive I represents a reduction in mortality; a negative value represents deterioration. Over n years, the geometric annualized rate is
[ \bar I_{x;t,n}=1-\left[\frac{m(x,t+n)}{m(x,t)}\right]^{1/n}. ]
The corresponding continuously compounded improvement is r=(1/n)log[m(x,t)/m(x,t+n)] and its discrete equivalent is I=1−exp(−r). If mortality falls by 20% over ten years, the annual compounded reduction is 1−0.8^(1/10), approximately 2.2067%, rather than 2.0000%. The logarithmic quantity supports additive time-series decompositions; the geometric rate supports direct multiplicative application to mortality rates. The units, population, rate measure, sex, calendar-year convention, and treatment of open age groups must be disclosed whenever an annual improvement is stated.
For a fixed x, the ratio m(x,t+1)/m(x,t) compares different birth cohorts observed at the same attained age in successive periods. For a fixed birth cohort c=t−x, the diagonal comparison m(x+1,t+1)/m(x,t) combines aging and mortality change. Treating that diagonal ratio as improvement confounds the substantial increase in mortality with age and the calendar-time effect. The distinction is central to every projected survivor path.
Mortality-level displacement and trend
One temporary shock in t may multiply a base mortality rate by k_t while leaving its long-run growth structure unchanged. If m(x,t)=m^(x,t)k_t, then observed annual improvement incorporates both the structural movement in m^ and the change in shock multiplier k_t. A mortality spike followed by reversion can produce an extremely high subsequent observed improvement even where the underlying trend has not accelerated. A fitted linear trend over such years can therefore exaggerate the durability of recovery. Robust methods, selective weighting, and explicitly specified shock overlays aim to distinguish the components while preserving the evidence of lives lost.
The HMD short-term mortality series documents the use of high-frequency death information during the pandemic; its weekly counts offer valuable shock diagnostics but require adjustment and interpretation before becoming annual improvements for a particular insured cohort. The Hyndman–Ullah robust forecasting study addresses unusual years within a functional mortality forecasting framework. Neither observation supplies a unique assumption for future abnormal mortality episodes.
Age standardization and portfolio mixes
The crude mortality rate for portfolio t equals Σ_x w(x,t)m(x,t), where w is an exposure share. Its movement contains a change in age-specific mortality and a change in portfolio age weights. If a pension portfolio ages, the crude rate may increase despite improvements at every age. An age-standardized comparison instead fixes w(x) to a reference distribution and compares Σ_x w(x)m(x,t). Even then, the comparison can be affected by socioeconomic, underwriting, duration, regional, and product composition within age cells.
A useful attribution decomposes observed mortality change into within-cell rate change and changing mixture weights. With m̄_t=Σ_g w(g,t)m(g,t), the exact two-date change can be written as Σ_g w(g,t)[m(g,t+1)−m(g,t)]+Σ_g [w(g,t+1)−w(g,t)]m(g,t+1). This formula assigns the first component to rate movement under starting weights and the second to composition movement under ending rates. The order is a convention; symmetric decompositions are possible. Its value is diagnostic: two apparently different national or insured improvement trends may become more comparable after removing the changing population mixture.
Base tables, future scales, and internal consistency
A mortality scale is a map of prospective improvement rates indexed by age and projection calendar year, or by attained age and duration. If q(x,t_0) denotes a base table at t_0, a common convention is
[ q(x,t_0+k)=q(x,t_0)\prod_{j=0}^{k-1}[1-I(x,t_0+j)] ]
for a fixed attained age x. In a real cohort valuation, x itself advances in each successive year, and the relevant cells are q(x_0+j,t_0+j). The base date and the calendar years already reflected in an experience table must be aligned before multiplying improvements. Updating a base table to a new experience period while also reapplying historical improvements from that period double counts the same decline. Conversely, using a pandemic-elevated base without allowing for an anticipated normalization can overstate future mortality for a lifetime valuation. The CMI's base-table interaction analysis specifically emphasizes such consistency.
A final conceptual distinction follows. The “best estimate” of mortality is conditional on an objective, information date, population, and judgmental horizon. It is neither a single timeless biological number nor necessarily identical across a life insurer, an annuity provider, a pension plan, and a regulator. The subsequent chapters derive the consequences of this conditionality without treating each context as a separate mortality science.
Period and Cohort Effects
The age–period–cohort geometry
Let c=t−x denote approximate year of birth. The mortality surface m(x,t) has an age direction x, a calendar-period direction t, and diagonal birth cohorts c. A period effect changes the mortality environment in a calendar year, potentially across several ages. A cohort effect represents a persistent difference associated with conditions experienced by a birth generation, potentially moving across attained ages. A change in diagnostic practices, exposure to smoking, early-life public health, vaccination, medical access, or socioeconomic selection may contribute to these patterns; their causality requires research beyond statistical correlation.
The identity c=t−x creates a formal identification problem. In additive age–period–cohort specifications, linear trends can be transferred among the age, period, and cohort components without changing fitted rates unless constraints are imposed. Consequently, a fitted “cohort effect” is conditional on identification restrictions, calendar window, data quality, and model class. This explains why cohort parameters should be evaluated through predictive behavior and substantive plausibility rather than interpreted mechanically as causal properties of birth-year membership.
Historical period tables hold the age-specific rates of a chosen year constant as a hypothetical mortality environment. A cohort table follows one birth cohort through actual or projected future rates. The 2024 U.S. life-table methodology explicitly distinguishes these two constructions. For an individual currently age x in t_0, projected t-year survival is
[ {}{k}p{x,t_0}=\prod_{j=0}^{k-1}{1-q(x+j,t_0+j)}. ]
A single-year period life expectancy does not incorporate the same future improvement path. If mortality improves over future years in all relevant ages, cohort survival under that path generally exceeds survival under unchanged period rates, holding the initial base and other conventions fixed.
The Lee–Carter period decomposition
Lee and Carter (1992) formulate an influential statistical model with an age profile, a common period index, and age-dependent period sensitivity:
[ \log m(x,t)=a_x+b_x k_t+\varepsilon_{x,t}. ]
The usual identifiability constraints normalize the b_x coefficients and center the k_t sequence; specific conventions vary. The estimated k_t is then forecast using a time-series process. The original Lee–Carter mortality forecast applied the framework to U.S. age-specific mortality. A decline in k_t translates into mortality improvements at ages where b_x is positive, but its fitted history and forecast uncertainty are sensitive to calendar range, smoothing, and how future periods are modeled.
Taking first differences shows the link to improvement: log m(x,t+1)−log m(x,t)≈b_x[k_(t+1)−k_t], if the fitted age pattern is fixed. For small annual changes, the discrete improvement rate is approximately −b_x Δk_t. This relation is interpretable only after the model's normalization and population have been specified. It does not imply that different ages must share the same improvement percentage.
A weakness of extrapolating a single period factor is the potential failure of stable age sensitivities or a stable linear trend through structural changes. Booth, Maindonald, and Smith's variable-decline Lee–Carter analysis illustrates the empirical importance of fitted periods and changing mortality decline, using Australian experience. The relevant practical question is whether the extrapolation encodes a defensible future mortality pattern for the liabilities being measured.
Counts and exposure: Poisson formulations
For death counts D(x,t) with exposure E(x,t), a canonical conditional likelihood specification is D(x,t)~Poisson[E(x,t)m(x,t)]. Brouhns, Denuit, and Vermunt (2002) embed the log-bilinear age-period structure in a Poisson regression framework, relating counts to their exposure offsets. The approach reflects the fact that a rate estimated from a small exposure has different statistical credibility from the same rate estimated using a very large exposure. Overdispersion, dependence, cohort effects, data revisions, and reporting differences may nevertheless require generalized distributions or additional treatment.
A complete projection should separate estimation uncertainty about parameters from future innovation uncertainty. In a Poisson model the likelihood reflects counting variability conditional on modeled rates. When the period process is simulated, future period innovations add process uncertainty. Repeated estimation over sampled histories adds parameter uncertainty. These components overlap in some computational procedures, so separately reported intervals must state precisely what their simulation includes.
Cohort extensions and higher-age dynamics
Renshaw and Haberman (2006) extend the Lee–Carter family with age-specific cohort mortality reduction factors. A schematic log-rate representation adds a term γ_(t−x) with an age-related cohort loading. Such models can capture systematic diagonal residuals that a period-only projection misses. Fitting richer age–period–cohort models increases demands for identification, boundary behavior, and out-of-sample validation, particularly when recent cohorts are incompletely observed at old ages.
Cairns, Blake, and Dowd (2006) model the logit of post-age-60 one-year death probabilities using two stochastic period factors, one acting on level and another on the age slope. Their two-factor mortality study directly connects high-age mortality dynamics to longevity risk. Additional stochastic mortality models differ in age response, cohort terms, trend dynamics, and forecast tails. The generalized frameworks and reproducible estimation tools documented by Villegas, Kaishev, and Millossovich (2018) in StMoMo facilitate comparison of these model structures rather than establishing a universally best specification.
At extreme ages, limited exposure, age misstatement, upper-age closure, and selection can dominate apparent pattern differences. A model may fit ages 60–90 while producing materially implausible survival between 100 and 115. Tail behavior must be inspected separately from ordinary in-sample residuals. Extrapolation should be tested against the oldest-age data that are sufficiently credible for the valuation population; unknown post-110 outcomes must not silently inherit a model's functional convenience.
Coherence across related populations
If male and female mortality, or national and insured mortality, are forecast independently, relative rates may diverge without plausible justification. Li and Lee's coherent mortality forecast explicitly investigates correlated long-run forecasts for groups of populations. In actuarial application, coherence may be useful for related sexes, regions, and reference populations; it is not a requirement that all groups converge to identical rates. Group-specific persistent differences can coexist with shared dynamics. Any constraint has to reflect the actual transfer problem: a selected annuity group cannot be assumed automatically to inherit the population mortality level of its country.
Taken together, period and cohort models identify statistical regularities in the past and structured projections for future mortality. They cannot establish their own long-term demographic scenario independently of interpretation. The next chapter examines what recent empirical and institutional evidence supports and where it remains contested.
Historical Trend Comparison
Reconstructing internationally comparable trends
The observational foundation of a mortality projection is a series of numerators and denominators that can be compared across ages, sexes, places, and years. Death registration completeness, age validation, the demographic census base, migration estimation, and changes in institutional reporting each affect a time trend. The HMD data construction process standardizes death counts, population estimates, exposures, death rates, and life tables across selected high-quality national populations. Its methods are exceptionally useful for reproducibility, yet its eligibility criteria also restrict the countries for which long comparable series are available.
The 2025 HMD Methods Protocol notes that a variation in the timing of births may create apparent cohort features if person-year exposures are reconstructed from a uniform-birth approximation during historical discontinuities. This is a striking example of an apparent substantive cohort pattern arising from measurement architecture. A credible cohort hypothesis must first withstand tests for numerator and denominator artifacts, age-heaping, denominator revisions, and the choice of population at risk. Observed diagonal structures are evidence to investigate, not automatic explanations.
For countries with less complete death registration or shorter historical series, direct replication of an HMD-style fitting window may be inappropriate. The UN World Population Prospects 2024 assembles national estimates and projections for a much wider range of populations, using a broad mixture of censuses, registration records, surveys, and estimation methods. A WPP projection is therefore a valuable international demographic benchmark, but it has different observation and modeling provenance from an insurance company's policyholder experience study. The choice between a population series and insured experience must be documented at the interface between external evidence and contractual exposure.
The recent record and the meaning of recovery
The pandemic era shows why a single recent rate of improvement can mislead. A large excess-mortality shock may be followed by recovery and a low observed death-rate year; a one-year improvement from the shock peak includes the unwinding of the temporary elevation. At the same time, surviving populations may undergo selection, long-term health impacts may persist, and the timing of deaths may alter subsequent observations. Those mechanisms can operate simultaneously. Their net effect is not identifiable merely from successive annual life-expectancy values.
The U.S. National Center for Health Statistics reported life expectancy at birth of 79.0 years for 2024, up 0.6 year from 2023, and a 3.8% decline in the age-adjusted death rate from 750.5 to 722.1 per 100,000 reference-population members. These are observed period indicators from the final 2024 mortality statistics. They demonstrate material recovery in the period surface. They do not measure the prospective annual improvement of U.S. annuitants aged 80 or the future trend of any insured-life portfolio. The NCHS separately reports that life expectancy at age 65 was 19.7 years in 2024; interpreting that number as cohort expectancy would improperly assume future mortality followed the same 2024 period pattern indefinitely.
The OECD's Pensions at a Glance 2025 reports average remaining life expectancy at age 65 in 2024 across OECD countries of 18.5 years for men and 21.6 for women. Its long-run projections include further longevity gains, while noting that the pace of old-age improvement has slowed. These regional summaries highlight variation across jurisdictions without establishing an annual trend that may be applied directly to every retirement annuity. Similarly, the United Nations 2024 projections contain country- and age-specific scenarios to 2100 rather than a single global insurance assumption.
Britain: CMI_2025 and a pending CMI_2026
The Continuous Mortality Investigation's CMI_2025 was released on March 10, 2026, with Working Paper 211, calibrating to England and Wales population mortality through December 31, 2025. The CMI_2025 working paper states that its core output produces higher cohort life expectancies than CMI_2024 when other parameters are held constant. The official March 2026 CMI announcement quantifies illustrative age-65 changes of about eight weeks for males and six weeks for females. These are changes between versions under specified settings, not the total cohort life expectancy of a person and not a universal projection of future improvement.
The CMI_2025 explanatory guidance describes a deterministic framework blending estimates of current short-term improvement with a user-chosen long-term rate. The long-term improvement is an actuarial assumption rather than a number supplied by historical data alone. Thus a quoted “CMI result” requires the long-term rate, initial adjustment, convergence/weighting settings, base table, and valuation date to be reproducible. Two organizations using the same CMI release can still have different prospective survival curves for legitimate, documented reasons.
Working Paper 214, issued September 16, 2026, addresses the interaction of base tables and improvements. It cautions against a mismatch between pandemic-era base rates and how their excess is assumed to run off. This reinforces the mathematical requirement that historical selection and prospective improvement be specified jointly. The later CMI_2026 consultation, Working Paper 216 of October 15, 2026, proposes a future model calibrated through 2026 and discusses high-age population estimates and mortality selection. As of the present research cutoff, CMI_2026 is a consultation proposal, not an adopted production model. Confusing the version number in a future consultation with a published model would distort the 2026 evidence record.
Registration date and occurrence date also require attention. Working Paper 211 observes that its calibration uses when deaths were registered; a reconstruction by occurrence date can differ where registration delays change. For high-frequency mortality monitoring, neither calendar alignment nor completeness is a trivial technicality. They shape a purported acceleration or deceleration of short-term improvement.
United States: population pension scales and insured-life scales
The Society of Actuaries' Retirement Plans Experience Committee reported in its October 2026 mortality update that it would not issue a new mortality-improvement scale for pension valuations in 2026. It continues to regard MP-2021 as a reasonable best-estimate foundation for retirement plan applications, especially ages 65 and above. The same report estimated an age-standardized actual-to-expected ratio for ages 65+ of 98.7% for July 2025–June 2026, against a benchmark generated from 2019 rates with MP-2021; its calendar-2025 comparison was 100.8%. These ratios are descriptive checks against that benchmark, with data and exposure revisions and reporting completeness affecting their precision. They do not show a statistically resolved replacement long-term slope.
The MIM-2026 technical overview extends its historical data through 2023 and retains the MP-2021 prospective scale. Its modeling practice excludes the post-2019 observations from the forward trend-estimation component while making those years available for retrospective analysis. This is an explicit treatment of an abnormal period, not an assertion that no mortality outcomes occurred during the pandemic. The original MP-2021 materials document the prior rate scale based on experience through 2019. The distinction among tool version, historical data update, and prospective adopted improvement scale is essential when citing a 2026 model name.
A separate 2026 Society of Actuaries recommendation concerns individual life insurance under AG38/VM20. The insured-life improvement scale recommendation uses insured-population observations for most adult ages, a historical component informed by 2013–2023 experience, and future assumptions related to the 2026 Social Security Administration Trustees Report. Its age-dependent treatment reflects the different insured lives, exposures, and valuation setting. The existence of a 2026 scale here does not contradict continued use of MP-2021 in U.S. pension applications. The two scale programs govern different kinds of business, draw on different experience, and use different specified methodologies.
The 2025 mortality improvement survey released in October 2026 reports heterogeneity across companies, products, functions, attained ages, and projection horizons. Its results do not support a claim of a single post-pandemic industry consensus. This is evidence of practice variation; it is not itself proof that all choices in the survey are equally supported by mortality data.
An evidence-comparison table
| Source or framework | Principal population | Observational window highlighted | Main prospective role | Limitation for another portfolio |
|---|---|---|---|---|
| HMD annual tables | Selected national populations | Country-specific historical series; Methods Protocol v6 (2025) | Comparable historical rates and cohort/period construction | Different age/sex/selection mix from insured lives |
| UN WPP 2024 | Countries and areas worldwide | Estimates from 1950 onward; projection to 2100 | International population scenarios | National estimation assumptions vary in quality and source |
| NCHS 2024 final data | U.S. population | Calendar 2024 | Recent period mortality anchor | Period life expectancy is not prospective cohort survival |
| CMI_2025 | England and Wales general population | Calibrated through 2025 | Short/long-term improvement framework with user inputs | Long-term rate and portfolio adjustment remain choices |
| CMI_2026 consultation | Proposed England and Wales series | Proposed calibration through 2026 | Future model design discussion | Proposal, not current released model |
| RPEC 2026 / MP-2021 | U.S. population for retirement assumptions | Post-pandemic monitoring against pre-2020 scale | Retention and testing of pension improvement basis | Not a selected insured-life table |
| 2026 AG38/VM20 recommendation | U.S. insured-life experience and other inputs | Adult insured rates 2013–2023 for HMI | Individual life insurance valuation | Not a general pension projection scale |
The comparison demonstrates that disagreement in a label such as “current mortality improvement” may reflect population membership, period of observation, or intended projection function. A valid comparison requires these dimensions to be fixed before the outcome is interpreted. It is also possible for a newer dataset to support retaining an earlier forward scale; more recent input data do not logically force a new assumption.
Longevity Assumption Construction
The valuation population and its base date
A projection begins by identifying who is exposed: population residents, medically underwritten lives, employees earning pensions, retirees in payment, deferred annuitants, voluntary annuity buyers, or insured beneficiaries. These groups differ in entry selection, socioeconomic position, benefits, cause-of-death mix, and mortality by duration since selection. A base table should reflect the target group, attained age, sex where relevant to the basis, calendar date, contractual eligibility, and experience credibility. The base table is a conditional estimate at the selected origin rather than a statement that every member has the same mortality risk.
A generic population-to-portfolio mapping might be expressed as m_P(x,t_0)=ρ(x,d,g)m_R(x,t_0), where R is a reference population and ρ is a relative mortality factor incorporating age x, duration d, and relevant group attributes g. If ρ subsequently evolves, improvement experienced by the portfolio need not equal improvement in its reference group. In differential form, ∂_t log m_P = ∂_t log m_R + ∂_t log ρ. This relation proves the importance of a dynamic selection assumption: the future trend of the selected portfolio includes both reference mortality improvement and the evolution of selection itself.
Direct estimates from a portfolio are preferable in principle when death counts and exposures are credible, but small cells can produce unstable estimates. Pooling ages, years, or populations can reduce variance at the cost of bias when their mortality dynamics differ. The data strategy should describe which sources were directly estimated, which cells were borrowed from reference tables, how portfolios were standardized, and what assumptions carry estimates beyond observed experience.
The short-to-long-term trajectory
Historical improvement estimates exhibit short-term variability, while very long horizons rely on structural and judgmental assumptions. A transparent deterministic projection may interpolate between current improvement I_short(x,t) and a long-run assumption I_long(x):
[ I(x,t_0+k)=w_x(k)I_{short}(x,t_0)+[1-w_x(k)]I_{long}(x), ]
where w_x(0)=1 and w_x(k) approaches zero at a disclosed rate. A geometric or exponential convergence weight, or a spline with conditions on slope and endpoint, has different short-horizon effects. This schematic expression illustrates a family of conventional blends; it is not presented as a new model. The actual CMI algorithm contains specific graduation and smoothing components, and its published CMI explanation must be consulted for that implementation.
For a forty-year annuity, a small difference in a long-term rate compounds through many survival probabilities. The long-run assumption deserves separate approval from a short-term statistical fit because it carries much of the economic exposure after the data-informed horizon. Structural drivers such as medical treatment, smoking patterns, public health, environmental conditions, inequalities, and future shocks can motivate scenarios, but a justified scenario requires evidence and explicit pathways. An announced technological possibility does not constitute a measurable adjustment to the mortality curve without corroborating survival experience or a controlled causal bridge.
Constraints and coherence of a projected surface
For each future age and year, a projected one-year death probability must satisfy 0≤q≤1, and implied survival along a cohort must be non-increasing with duration. Hazard estimates must be nonnegative; achieved death probabilities should remain compatible with the chosen conversion formula and any maximum-age convention. Across attained ages, monotonicity can be a useful check for selected old-age mortality but should not be imposed as an inflexible rule where measured young-adult or other mortality shapes are non-monotone. Model fitting must differentiate actual information from formal smoothness assumptions.
Projection coherence extends across multiple insured subpopulations. If group A initially has lower mortality than group B, an unconstrained fit might reverse their ordering decades later. A reversal could be realistic if their improvement rates differ, but it requires a plausible explanation and sensitivity analysis. Li and Lee's coherence argument supplies a statistical motivation for linking related projections. It does not automatically set the long-run ratio for annuity buyers relative to a country's total population.
Distinguishing target, model, and governance assumptions
An assumption file should identify each substantive choice, not merely state “mortality follows model X.” Distinct choices include the reference population, data vintage, exposure treatment, base year, cause-of-death scope, amount- or life-weighting, experience graduation, projected trend model, cohort inclusion, tail extrapolation, convergence horizon, ultimate rate, handling of exceptional years, and any management overlay. Financial assumptions such as discount rates, benefit indexation, expense loadings, and policyholder options affect the liability even though they are not themselves mortality improvements. Keeping the assumptions separate makes sensitivity attribution possible.
Model governance entails an accountable scientific record. The International Actuarial Association's ISAP 1 standards catalogue situates general actuarial practice guidance and model governance within internationally recognized model standards. A documented assumption process should preserve source versions and dates, define the intended decision, evaluate adequacy of data, compare alternative methodologies, communicate limitations, and record actual approvals. Model standards describe professional conduct and are not a license to infer that any unpublished scenario is empirically settled.
Numerical construction: a specified synthetic cohort
The following example is deliberately synthetic. A single life is age 65 at time 0. For projection years j=0,…,39, choose the illustrative base death probabilities along the future age path as q_base(j)=0.012(1.09)^j. Let i be a constant annual improvement rate applied by calendar duration to each future attained-age cell:
[ q_j(i)=0.012(1.09)^j(1-i)^j,\qquad S_0=1,\qquad S_{j+1}=S_j[1-q_j(i)]. ]
Here j indexes both years since valuation and age 65+j. The base curve's 9% multiplicative age-gradient is an invented pedagogical input rather than an observed national mortality parameter. The assumption i is a fractional reduction per subsequent calendar year and applies multiplicatively; it does not change q_0 at the valuation date. All q_j over the 40-year horizon remain below one under the scenarios shown. Payments occur at the end of each year while the life is then alive, discounting at a flat effective 3% per year. There is no payment after year 40, so the valuation is a term-certain projection horizon, not a complete whole-life annuity to extinction.
The forty-year annual life-annuity factor is a_40(i)=Σ_(j=0)^39 S_(j+1)/(1.03)^(j+1). For a unit death benefit payable at each year-end if the life dies in that year, d_40(i)=Σ_(j=0)^39 S_j q_j(i)/(1.03)^(j+1). The expected number of survival-contingent year-end payments is e_40(i)=Σ_(j=0)^39 S_(j+1). These equations completely specify the experiment; all displayed results are independently recalculated below rather than taken from an outside model output.
| Annual improvement i | Forty-year annuity factor a_40 | Unit death-benefit PV d_40 | Expected payments e_40 | Survival through year 40 |
|---|---|---|---|---|
| −1.0% (deterioration) | 13.505829 | 0.577027 | 18.760262 | 0.001591 |
| 0.0% | 14.088191 | 0.557259 | 19.979891 | 0.011016 |
| 0.5% | 14.399967 | 0.544920 | 20.654047 | 0.021965 |
| 1.0% | 14.723783 | 0.530520 | 21.367948 | 0.038658 |
| 1.5% | 15.057318 | 0.513941 | 22.115953 | 0.061723 |
| 2.0% | 15.397586 | 0.495257 | 22.890241 | 0.091197 |
Moving from zero improvement to 1.0% per annum raises the synthetic annuity factor by 0.635592, or approximately 4.51%, while lowering the term death-benefit value by 0.026739, or approximately 4.80%. A further move from 1.0% to 2.0% raises the annuity factor by another 0.673803. These changes are not linear in the assumption because improvements compound across age-year cells and survival products. This nonlinear accumulation explains why a modest percentage-point assumption change can become financially important over long durations, especially for payments at the far end of the horizon.
The example cannot be quoted as evidence that any particular insurer's annuity reserves would rise by 4.51% for a one-percentage-point change in improvement. Its exact outcome reflects the invented age curve, fixed 40-year horizon, payment timing, and 3% discount. A longer contractual horizon, joint-life benefits, spouse-contingent payments, guarantee periods, or a different asset yield will change the sensitivity. The next chapter formalizes the uncertainty and stress response.
Shock and Trend Sensitivities
The conceptual separation of uncertainty
A mortality projection involves several distinguishable uncertainty sources. Observation uncertainty arises because death counts, exposures, and reported characteristics are imperfect and may be revised. Parameter uncertainty arises from estimating a finite number of fitted coefficients and their interactions. Process uncertainty describes the variability of future mortality conditional on a parameterized model. Structural uncertainty concerns the model family, the fit window, the assumed long-term regime, and the tail behavior. Population basis risk arises when a reference population and an actual benefit population respond differently to the same historical and future drivers. Scenario uncertainty captures specified departures, including extraordinary public-health, socioeconomic, or medical developments, that may be outside the fitted distribution.
These components have different evidentiary status. Death-count noise can often be quantified from an estimated sampling model. Parameter uncertainty can be approximated by resampling, covariance propagation, or a posterior distribution. Structural uncertainty cannot be removed merely by generating more Monte Carlo paths within one chosen model. Reference-to-portfolio differences require independent empirical evidence, not additional simulation from a population model. A scenario set is a collection of conditional calculations and can inform resilience even where scenario probabilities are unknown.
The two-factor Cairns–Blake–Dowd study explicitly examines parameter uncertainty in high-age mortality modeling. The StMoMo modeling framework enables alternative age–period–cohort specifications to be fitted and compared. The central actuarial judgment is how much of the uncertainty distribution is credible at relevant ages and horizons, not how many mathematically possible simulations can be produced.
Exact sensitivity in the synthetic cohort
For the constructed q_j(i), differentiating produces dq_j/di=−j q_j(i)/(1−i). Consequently, for k≥1,
[ \frac{\partial S_k}{\partial i}=S_k\sum_{j=0}^{k-1}\frac{j q_j(i)}{(1-i)[1-q_j(i)]}. ]
All terms are nonnegative for j≥0 where 0≤q_j<1 and i<1. The derivative proves that higher improvement increases survival in this particular model. The corresponding annuity derivative is the discounted sum of the survivor derivatives: da_40/di=Σ_(k=1)^40 v^k ∂S_k/∂i. This formula identifies why later-year benefits have more cumulative improvement exposure: the sensitivity of survival at k depends on the entire preceding path. It also explains why the derivative with respect to improvement is not simply an age-specific mortality elasticity multiplied by duration.
A finite-difference sensitivity should be calculated with the same base rates, financial basis, cap on the projection, and benefit design in each run. Changing the base mortality table and the improvement scale simultaneously without a factorial decomposition masks which assumption generated the movement. Conversely, treating them as statistically independent may be equally misleading where the base table and trend are fitted to overlapping experience. Both the mathematical derivative and the governance distinction point to joint but separable reporting of the inputs.
One-time mortality reductions versus improvement-rate changes
A permanent instantaneous level reduction is mathematically distinct from a rise in the annual improvement rate. For a uniform level shock s, q_j^shock=(1−s)q_j(i) for all projection years j; the reduction applies immediately to q_0. A one-percentage-point change in i, by construction, leaves q_0 unchanged and compounds increasingly across later years. The economic effects therefore depend on very different timing.
In the hypothetical cohort with i=1.0%, a permanent 20% decrease in every projected q_j changes the forty-year annuity factor from 14.723783 to 15.766808, a rise of 1.043025, approximately 7.08%. The unit death-benefit PV falls from 0.530520 to 0.488734. This experiment is an illustration of a uniform mortality-rate level shock; it is not a calculation of an insurer's Solvency Capital Requirement. It ignores liabilities outside the specified life, changes to own funds, diversification, contractual options, tax, expenses, management actions, and valuation-specific conditions.
The distinction matters in the Solvency II longevity sub-module, whose delegated-regulation Article 138 specifies an instantaneous permanent 20% reduction in mortality rates for contracts where lower mortality increases relevant technical provisions. It is a prescribed capital stress, not a historical estimate that mortality will necessarily improve by 20% each year and not an estimate of annual mortality improvement. A proper regulatory calculation determines the change in basic own funds under the stress in the relevant product set and applies the jurisdiction's full rules.
Structural breaks and post-pandemic model choice
An explicit shock treatment may take one of several forms: estimating the long-run trend excluding a shock interval; retaining all years with reduced statistical weight; specifying a transient mortality-level overlay; adding a permanent level shift; or modeling several future regimes. Each specification answers a different question. Excluding years from the fitted long-run process does not remove the information they convey about current mortality levels. Retaining every year as an ordinary realization does not prove the historical epidemic is a recurrent component of a linear trend.
The SOA RPEC 2026 assessment illustrates this tension. It observes recent age-65-plus mortality near its MP-2021 projection while finding insufficient evidence of a stable changed improvement trend. The SOA MIM-2026 treatment separates extended retrospective data through 2023 from the pre-2020 period used for forward estimation. Conversely, the CMI updates the fitted short-term mortality experience to 2025 while retaining explicit user decisions about long-term improvement. These practices are defensible as different assumption frameworks; comparing their outputs without controlling populations and model parameters would confound statistical and contractual questions.
A model review should test whether observed deviations reflect death registration timing, exposure changes, population composition, statistical noise, model misspecification, or true trend evolution. That is particularly important at old ages, where denominator revisions may be comparable in effect to the change attributed to mortality improvement. The CMI_2026 consultation explicitly raises high-age population estimates as a potential methodological issue; it remains a consultation rather than an adopted result at the research date.
Validation should be predictive and economically relevant
An adequate mortality projection can be tested along several dimensions. Historical reconstruction checks whether the stated exposure and death counts recover the published rates and whether the age and calendar axes are aligned. In-sample diagnostics identify systematic residuals by age, period, and cohort. Rolling-origin backtests fit each model using information available at an earlier cutoff and evaluate realized mortality at subsequent dates. Forecast comparisons should include both statistical fit and how errors affect the annuity or death-benefit cash flows. A model that improves mean squared error at ages rarely represented in a portfolio need not improve its benefit valuation.
Suitable forecast criteria include weighted log-rate error, calibration of death-count distributions, coverage of predictive intervals, cohort survival error, and present-value-weighted bias. Scoring a full period/cohort surface against the future experience actually observed at the verification date is preferable to checking a single life-expectancy number. However, no rolling-origin exercise can directly observe the forty-year future of a cohort whose remaining life is not complete. Long-horizon validation consequently requires historical cohorts, proxy comparisons, parameter stability diagnostics, and transparent scenario bounds.
A concise validation protocol can be specified as follows. First freeze the training data vintage, exposure definition, and projection cutoff. Fit the base table and improvement model using only information available at that date. Generate annual rate surfaces, surviving cohort distributions, and cash-flow metrics for predefined forecast years. Compare with subsequently observed data under a fixed measurement convention, quantify the effect of data revisions, and repeat over multiple cutoffs. Record failures of plausible mortality ordering, long-run divergence, or oldest-age behavior even when numerical scoring improves. This is a methodology for assessment, not a claim that the present monograph has run a multi-country historical backtest.
Scenario interpretation without spurious precision
Point forecasts support a central valuation, while scenarios reveal asymmetric exposures. A retirement annuity liability generally benefits economically from higher mortality, other things equal, because fewer survivor payments are made; a pure death benefit often has the opposite sensitivity. The combined position may contain partial offsets but timing, benefit amounts, age distribution, risk selection, and discounting prevent a universal hedge ratio. Apparent offsets from adding actuarial products can reverse if the portfolios have different future improvement rates.
An assumption report should present both absolute financial amounts and relative changes, identify which scenario dimensions are changed, and disclose the horizon. Uncertainty ranges should be labeled according to how they were constructed: statistically calibrated predictive intervals, parameter confidence intervals, deterministic stress bands, or management scenarios. A numerical band without a stated construction method implies more knowledge than the evidence provides.
Pricing and Reserving Consequences
Survival-contingent cash flow valuation
Consider a life age x at t_0 receiving B_k at the end of year k if alive. On a deterministic discount curve with discount factor v_k, the expected present value of the payments through K is
[ V_A=\sum_{k=1}^{K} v_k B_k,{}{k}p{x,t_0}. ]
When all benefits and discounts are nonnegative and mortality improvements increase the cohort survival terms, V_A increases. The magnitude is concentrated at durations where discounted payments are material and the probability of survival is sensitive to the projected death rates. A life annuity whose major payments occur near age 95 may be particularly sensitive to improvement at older attained ages even if its initial exposure is at age 65. A deferred annuity can be more sensitive to the initial projection interval because the payment stream begins only after deferral.
For a unit term life insurance benefit paid at the end of the year of death,
[ V_D=\sum_{k=1}^{K} v_k B_k,{}{k-1}p{x,t_0}q(x+k-1,t_0+k-1). ]
The direct sensitivity of each death probability and the indirect effect of survival to that year act jointly. Increasing longevity often reduces term death-benefit value, but the direction of every contract's total liability cannot be established without its detailed benefits, premiums, surrender options, guarantees, and expenses. In a whole-life setting, postponing death also changes the timing of eventual payment. A mix of annuity and death-contingent obligations requires contract-level valuation rather than applying a single mortality shock sign to the entire balance sheet.
The exact synthetic calculation above quantifies both directions under one transparent convention. At 1.0% annual improvement, the annuity factor is 14.723783 and the term death-benefit factor is 0.530520. At 2.0% annual improvement, the values become 15.397586 and 0.495257 respectively. These quantities isolate the mortality assumption because the same invented base table, flat discount rate, forty-year horizon, and year-end timing are retained. Actual insurer technical provisions incorporate wider coverage and accounting requirements.
Age, discount rate, and liability duration
The survival derivative of a payment at k is accumulated from every preceding mortality cell. Discounting dampens exposure to late survivor payments, while long benefit duration amplifies the influence of future improvements. A low discount rate increases the relative weight of distant cash flows; indexed annuities and survivor pensions can magnify it further. A high starting mortality level does not automatically imply high improvement sensitivity when very few members survive to the durations affected by the change. The relevant product is the discounted benefit amount times the derivative of its survival probability.
This interpretation also explains why pricing assumptions and capital assumptions can differ. The pricing basis may include margins for uncertainty, expenses, contract options, and commercial strategy; a current best-estimate liability typically aims at expected cash flows under its specified accounting or regulatory valuation framework. A margin should have an explicit purpose rather than being embedded invisibly into both mortality level and improvement rate. Otherwise, a practitioner cannot distinguish conservative population selection from deliberate risk protection.
Financial reporting and regulatory valuation
The IFRS 17 insurance-contract standard establishes the measurement framework for qualifying insurance contracts and combines current estimates of future cash flows with the relevant insurance-service and financial reporting components. A change in prospective mortality can alter expected policy cash flows and, depending on the contract group and measurement model, affect accounting components. The standard is not a mortality improvement table; it supplies an accounting context in which transparently justified assumptions must operate.
The European Solvency II Article 77 defines the best estimate of technical provisions as a probability-weighted present value of future cash flows based on credible, up-to-date information and realistic assumptions, with a separately specified risk margin in the ordinary case. The calculation depends on the applicable risk-free term structure and the entire policy cash-flow projection. These requirements make the valuation population and projected longevity path material, not decorative, assumptions. They also demonstrate why a scenario calculation using a flat 3% rate is pedagogical and cannot be substituted for a regulatory technical provision.
The Solvency II 20% longevity stress is a standardized capital scenario for the applicable liabilities, distinct from their best-estimate improvement scale. Its permanence assumption emphasizes level sensitivity and provides a stress benchmark. The capital framework may not fully capture structural model risk or the total cost of prolonged deviations from a best estimate. Supplementary sensitivity analysis can therefore examine higher or lower convergence rates, changed cohort patterns, sustained deviations at oldest ages, and portfolio-to-population divergence.
Retiree portfolios, individual life contracts, and product asymmetry
A uniform population trend can produce different financial results across annuities and life insurance. Annuitants are selected by willingness and eligibility to buy an income product, and often differ in mortality from the general population. Pensioner mortality varies with the employer, benefit amount, socioeconomic profile, survivor design, and whether the portfolio contains active, deferred, or in-payment members. Insured-life mortality depends on underwriting, contract duration, lapse behavior, sum insured, and changing entry selection. These differences influence both the starting mortality rate and how a future improvement should be transferred.
The separate U.S. 2026 individual-life scale and 2026 retirement plan review illustrate this product-specific methodology. One is an insured-life improvement recommendation for a particular valuation application; the other reviews a population-based pension scale. Their numbers cannot be interchanged without explicit reconciliation of policy context and exposure. The distinction is especially consequential when constructing a portfolio with annuity and life-insurance obligations that appear to offset mortality risk. Shared macro-demographic shocks may coexist with different selection dynamics and liability durations.
Assumption-setting as a controlled scientific process
The fundamental decision process is a succession of evidence-bearing choices. First, select and document the data universe and valuation population. Second, estimate or select the base mortality level with a traceable exposure denominator. Third, isolate apparent period anomalies and investigate coherent cohort residuals. Fourth, choose a prospective improvement mechanism and disclose long-run assumptions. Fifth, translate period improvements into the diagonal attained-age and calendar-year path used in each contract's survivor probabilities. Sixth, evaluate valuation sensitivity against plausible alternatives and prescribed capital stresses. Finally, freeze the data and assumptions for reproducibility, while assigning an update procedure to genuinely new evidence.
The international sources converge on several of these requirements without implying identical rates. HMD provides measurement provenance; statistical literature identifies model behavior; CMI provides a user-parameterized projection framework; SOA reports distinguish retirement and insured populations; accounting and solvency rules specify the valuation setting. Their proper integration is more defensible than selecting the newest published projection label and treating it as a universal longevity forecast.
Original analytical synthesis
The economic sensitivity of a liability is determined by its exposure to the complete projected survival path. Accordingly, the adequacy of a mortality improvement assumption is evaluated by the consistency among observation, population selection, cohort projection, and cash-flow timing. A model can perform well in reconstructing observed annual rates and still misstate an annuity valuation if the base date is misaligned, a cohort term is extrapolated unrealistically, or an insured-population selection factor is frozen despite credible evidence that it changes.
The synthesis yields three linked conclusions. First, the mortality level and future improvement are identifiable as separate recorded assumptions but must be evaluated as a coherent pair, especially when a recent shock changes the observed base. Second, projection uncertainty and portfolio basis risk have different empirical remedies: more simulations address conditional process variability, while better portfolio data and independent validation address transfer errors. Third, long-run assumption governance has an economic function because small differences in annual improvement compound into nontrivial survivor-duration and present-value effects. These conclusions follow from established models and the calculations here; they are analytical deductions rather than claims to invent a new theoretical system.
Conclusion
Mortality improvement and longevity assumptions connect observed population change with the long-term promises of insurers, pension arrangements, and annuity providers. Their scientific treatment begins with rigorous definitions of rates, probabilities, exposures, age, period, and cohort, and extends to explicit choices about the future. The identity of the population is integral to the assumption. An age-standardized national trend is evidence about a population; its adoption for a selected insured portfolio requires a reasoned and testable transfer.
The comparison of historical, academic, and institutional evidence shows why contemporary mortality projections remain plural. The latest available CMI model at the research date is CMI_2025, calibrated through the end of 2025, while CMI_2026 remains in consultation. U.S. pension research in 2026 maintains MP-2021 as a reasonable retirement projection foundation amid insufficient evidence for replacing the long-term slope, even as the Society of Actuaries has issued a different 2026 recommendation for individual insured-life valuation. These developments are compatible because their data, populations, and uses differ.
Period recovery after mortality shocks does not independently determine the shape of a future cohort trajectory. The same mortality history can support distinct conditional projections if the selected base date, normalization of exceptional years, convergence horizon, ultimate improvement assumption, and population selection differ. Coherent actuarial work makes these choices inspectable. A projection's validity is assessed through source provenance, mathematical consistency, out-of-sample testing where possible, longevity-tail diagnostics, and financially relevant sensitivities.
The independently calculated synthetic cohort shows how a change from zero to 1.0% annual improvement raises a forty-year annuity factor from 14.088191 to 14.723783 and lowers a forty-year death-benefit factor from 0.557259 to 0.530520 under the fixed assumptions. A permanent 20% mortality-level reduction is economically and mathematically different from increasing the annual improvement rate. Their distinction is necessary when interpreting best-estimate liabilities and prescribed longevity capital stresses.
The resulting research conclusion is operational and scientific: mortality-level calibration, improvement projection, cohort survival, and benefit valuation are parts of one traceable calculation. Each requires a population, a time convention, a validated measurement basis, and a declared degree of uncertainty. The long-term scale is a forecast assumption supported and bounded by evidence, and its financial effects are measurable only through the specific liabilities it governs.
Reproducibility and Editorial Method
Source review and research limitations
This dossier is based on first-party statistical and institutional documentation and identifiable peer-reviewed research. Sources were examined through official publisher, institutional, repository, or scholarly index pages available by November 14, 2026. Some institutional working-paper full texts require subscriber access; for these, only the public official descriptions and release statements are relied on. The HMD and NCHS PDF sources were located and their accessible textual records inspected; complete image-level PDF verification was not available in the web retrieval interface. There was no direct download and estimation of HMD country-level death/exposure microdata, nor a live annuity portfolio data set. Accordingly, the manuscript makes no fitted-country forecast, parameter uncertainty interval, causal effect estimate, or real-company reserve measurement claim.
The numerical example is fully specified by its formulas and parameter values and is independently calculated from those assumptions. Values are rounded to six decimals only for display, with arithmetic performed at full floating-point precision. The forty-year tail truncation is an intentional modeling convention, not an estimate of complete remaining life expectancy. It can be reproduced using the following pseudocode, which is also executable Python 3:
def evaluate(i, level=1.0):
survival = 1.0
annuity = death_pv = expected_payments = 0.0
for j in range(40):
q = level * 0.012 * (1.09 ** j) * ((1.0 - i) ** j)
assert 0.0 <= q < 1.0
v = 1.03 ** -(j + 1)
death_pv += survival * q * v
survival *= 1.0 - q
annuity += survival * v
expected_payments += survival
return annuity, death_pv, expected_payments, survival
for i in [-0.01, 0.0, 0.005, 0.01, 0.015, 0.02]:
print(i, evaluate(i))
print('20 percent level stress', evaluate(0.01, 0.8))
An independent discounted-cash-flow identity provides an exact reconciliation. Put v=1/1.03, let S_40(i) be survival to the end of year 40, and use the stated end-of-year annuity and death-benefit timing. Then d_40(i)+(1−v)a_40(i)+v^41 S_40(i)=v. This follows by summing death probabilities S_j q_j=S_j−S_(j+1) after discounting, with the terminal survival term retained. The values computed above satisfy the identity to floating-point precision. Positivity, survivor monotonicity, and increasing annuity factors were checked numerically for the stated scenario set.
Editorial provenance
The research was written as a complete standalone Research Dossier with an approved canonical title, source-specific contextual hyperlinks, full References, and a separate machine-readable source-to-claim ledger. Author: Angela Bogdanova (Artificial Sapiens). Publisher: Aisentica Development. All research and mathematical interpretations were prepared for the Insurance by Bogdanova encyclopedia. No external peer review, public publication date, DOI assignment, or live site release is claimed by this manuscript. The accompanying PASSPORT.json and EVIDENCE.json are the authoritative internal research package metadata and factual-check scope.
References
Arias, Elizabeth, Jiaquan Xu, and Betzaida Tejada-Vera. 2026. United States Life Tables, 2024. National Vital Statistics Reports 75 (5). National Center for Health Statistics, August 25. https://www.cdc.gov/nchs/data/nvsr/nvsr75/nvsr75-05.pdf.
Booth, Heather, John Maindonald, and Len Smith. 2002. “Applying Lee-Carter under Conditions of Variable Mortality Decline.” Population Studies 56 (3): 325–336. https://doi.org/10.1080/00324720215935.
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. https://www.sciencedirect.com/science/article/pii/S0167668702001853.
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. https://doi.org/10.1111/j.1539-6975.2006.00195.x.
Continuous Mortality Investigation. 2026a. “CMI Working Paper 211: CMI Mortality Projections Model, CMI_2025.” Institute and Faculty of Actuaries, March 10. https://www.actuaries.org.uk/learn-and-develop/continuous-mortality-investigation/cmi-working-papers/mortality-projections/cmi-working-paper-211.
Continuous Mortality Investigation. 2026b. “Mortality Improvements and CMI_2025: Frequently Asked Questions.” Institute and Faculty of Actuaries. https://www.actuaries.org.uk/learn-and-develop/continuous-mortality-investigation/cmi-working-papers/mortality-projections/cmi-working-paper-211/mortality-improvements-and-cmi2025-frequently-asked-questions-faqs.
Continuous Mortality Investigation. 2026c. “CMI Working Paper 214: Interaction between Mortality Base Tables and Improvement Assumptions.” Institute and Faculty of Actuaries, September 16. https://www.actuaries.org.uk/learn-and-develop/continuous-mortality-investigation/cmi-working-papers/mortality-projections/cmi-working-paper-214.
Continuous Mortality Investigation. 2026d. “CMI Working Paper 216: CMI_2026 Consultation.” Institute and Faculty of Actuaries, October 15. https://www.actuaries.org.uk/learn-and-develop/continuous-mortality-investigation/cmi-working-papers/mortality-projections/cmi-working-paper-216.
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. https://doi.org/10.1016/j.csda.2006.07.028.
Human Mortality Database. 2026a. “Overview: Scope and Basic Principles; Computing All-Cause Death Rates and Life Tables.” University of California, Berkeley, Max Planck Institute for Demographic Research, and associated HMD team. Accessed November 14, 2026. https://www.mortality.org/Project/Overview.
Human Mortality Database. 2026b. “Short-Term Mortality Fluctuations: Data and Documentation.” HMD; latest displayed update October 26, 2026. https://www.mortality.org/Data/STMF.
International Actuarial Association. 2026. “International Standards of Actuarial Practices.” Includes ISAP 1 General Actuarial Practice and associated model standards. Accessed November 14, 2026. https://actuaries.org/publications/international-standards-of-actuarial-practices/.
International Financial Reporting Standards Foundation. 2026. “IFRS 17 Insurance Contracts.” Effective for annual periods beginning January 1, 2023, subject to the standard's transition provisions. https://www.ifrs.org/issued-standards/list-of-standards/ifrs-17-insurance-contracts/.
Institute and Faculty of Actuaries. 2026. “CMI Model Shows Further Rise in Cohort Life Expectancy.” News release, March 10. https://actuaries.org.uk/news/cmi-publishes-cmi_2025-mortality-projections-model/.
Lee, Ronald D., and Lawrence R. Carter. 1992. “Modeling and Forecasting U.S. Mortality.” Journal of the American Statistical Association 87 (419): 659–671. https://doi.org/10.1080/01621459.1992.10475265.
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. https://pmc.ncbi.nlm.nih.gov/articles/PMC1356525/.
National Center for Health Statistics. 2026. “Mortality in the United States, 2024.” NCHS Data Brief 548, January 29; by Jiaquan Xu, Sherry L. Murphy, Kenneth D. Kochanek, and Elizabeth Arias. https://www.cdc.gov/nchs/products/databriefs/db548.htm.
Organisation for Economic Co-operation and Development. 2025. Pensions at a Glance 2025: OECD and G20 Indicators. Paris: OECD Publishing; “Life Expectancy” chapter. https://www.oecd.org/en/publications/pensions-at-a-glance-2025_e40274c1-en/full-report/life-expectancy_02c88784.html.
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. https://doi.org/10.1016/j.insmatheco.2005.12.001.
Society of Actuaries Research Institute, MIM Advisory Group. 2026. Overview of the 2026 Updates to the Mortality Improvement Model (MIM). June 23. https://www.soa.org/resources/research-reports/2026/2026-updates-mim/.
Society of Actuaries Research Institute, Mortality Improvement Life Working Group. 2026. “Individual Life Insurance Mortality Improvement Scale Recommendation for AG38/VM20.” September 22. https://www.soa.org/resources/research-reports/ag-38-mortality-improvement/.
Society of Actuaries Research Institute, Retirement Plans Experience Committee. 2021. Mortality Improvement Scale MP-2021. https://www.soa.org/resources/experience-studies/2021/mortality-improvement-scale-mp-2021.
Society of Actuaries Research Institute, Retirement Plans Experience Committee. 2026. RPEC 2026 Mortality Improvement Update. October 22. https://www.soa.org/resources/research-reports/2026/rpec-mortality-improvement-update/.
Society of Actuaries Research Institute. 2026. “2025 Mortality Improvement Survey Report.” Research report authored by Kara Clark, October 15. https://www.soa.org/resources/research-reports/mort-improvement-survey-series/.
United Nations, Department of Economic and Social Affairs, Population Division. 2024. World Population Prospects 2024. New York: United Nations. https://population.un.org/wpp/.
Villegas, Andrés M., Vladimir K. Kaishev, and Pietro Millossovich. 2018. “StMoMo: An R Package for Stochastic Mortality Modeling.” Journal of Statistical Software 84 (3): 1–38. https://doi.org/10.18637/jss.v084.i03.
Wilmoth, John R., Kirill Andreev, Dmitri Jdanov, Dana A. Glei, and Timothy Riffe, with collaborators. 2025. Methods Protocol for the Human Mortality Database, version 6, August 5. Human Mortality Database. https://mortality.org/File/GetDocument/Public/Docs/MethodsProtocolV6.pdf.
European Insurance and Occupational Pensions Authority. 2021. “Solvency II Single Rulebook: Article 77, Calculation of Technical Provisions.” Directive 2009/138/EC; rulebook entry updated May 3, 2021. https://www.eiopa.europa.eu/rulebook/solvency-ii-single-rulebook/article-2160_en.
European Insurance and Occupational Pensions Authority. 2024. “Solvency II Single Rulebook: Article 138, Longevity Risk Sub-Module.” Commission Delegated Regulation (EU) 2015/35; rulebook entry updated March 26, 2024. https://www.eiopa.europa.eu/rulebook/solvency-ii-single-rulebook/article-5758_en?prefLang=lv.