Lee–Carter Mortality Forecasting

• Lee–Carter is a mortality modeling framework that decomposes log mortality rates into fixed age effects and a stochastic period effect, facilitating probabilistic forecasts.
• It employs statistical methods such as maximum likelihood and Bayesian techniques (e.g., Kalman filtering, HMC) to quantify uncertainty and improve projection accuracy.
• Extensions include multi-population modeling, incorporation of cohort effects, and state-space formulations to address challenges at extreme ages and structural shifts.

The Lee–Carter framework is a foundational and widely adopted approach for modeling and forecasting mortality rates in demographic and actuarial contexts. It is based on the decomposition of the logarithm of age-specific mortality rates into fixed age effects and a time-varying period effect, with stochastic dynamics applied to capture systematic mortality improvement over time. The framework offers extensions for joint multi-population modeling, incorporation of cohort and volatility effects, and various statistical estimation strategies.

1. Mathematical Formulation of the Lee–Carter Model

The classical Lee–Carter (LC) model expresses the log central death rate at age $x$ and year $t$ as

$\log m_{x,t} = \alpha_x + \beta_x\,\kappa_t,$

where:

• $\alpha_x$ is the average log mortality rate at age $x$ across time,
• $\beta_x$ quantifies the sensitivity of age $x$ to the period factor (i.e., how improvements at age $x$ respond to changes in the period index),
• $\kappa_t$ is a time-varying mortality index reflecting period-specific mortality improvement.

For stochastic projection, $\kappa_t$ is assumed to follow a time series process, typically a random walk with drift:

$t$0

with identifiability constraints $t$1 and $t$2. Extensions can incorporate cohort terms (e.g., Renshaw–Haberman: $t$3) and generalizations to Age–Period–Cohort (APC) models (Fung et al., 2016).

2. Estimation and Inference

Parameter estimation in the LC class is conducted via maximum likelihood, typically under a Poisson or negative binomial framework for observed death counts, using exposures $t$4 as an offset:

$t$5

or, for overdispersed settings,

$t$6

For fully Bayesian variants and for the incorporation of smooth age, period, and cohort effects, this framework is extended using generalized additive models (GAM) with P-splines and Hamiltonian Monte Carlo for joint posterior sampling (Hilton et al., 2018). Model selection among competing cut-offs (e.g., age thresholds for model transition) is achieved using Pareto-smoothed importance sampling leave-one-out cross-validation (PSIS-LOOIC) and Bayesian stacking to form predictive mixtures of models.

In joint multi-population extensions, as in the Li–Lee model, the period index is decomposed into a "common" (global) trend and population-specific deviations, both modeled as stochastic processes subject to identifiability constraints. Cross-population period innovations may be modeled jointly as correlated random walks (Robben et al., 2021).

3. Extensions: Old-Age, Volatility, and State-Space Generalizations

The Lee–Carter framework struggles at extreme old ages, where data are sparser and the log-linear structure becomes inadequate. To address this, parametric “Kannisto-like” formulae for $t$7 can be used, imposing a logistic asymptotic behavior:

$t$8

with $t$9 governing the old-age mortality plateau. Multi-population models may share $\log m_{x,t} = \alpha_x + \beta_x\,\kappa_t,$0 across populations to borrow statistical strength in old-age extrapolation (Hilton et al., 2018).

For full dynamic modeling, the framework is often cast in a state-space or hidden Markov model setting, which enables:

• Unified likelihood or Bayesian inference,
• Flexible identification constraints suited to filtering,
• Inclusion of observation error heteroscedasticity,
• Stochastic volatility in period effects (via time-varying process innovations),
• Cohort effects with ARIMA dynamics (Fung et al., 2016).

Bayesian estimation is facilitated by Kalman filtering and smoothing, particle MCMC, or Hamiltonian Monte Carlo.

4. Forecasting and Uncertainty Quantification

Forecasts are generated by simulating the time series of the period index (and cohort effects, if included), then combining with age-$\log m_{x,t} = \alpha_x + \beta_x\,\kappa_t,$1 coefficients to yield projected log-mortality, and converting to death probabilities:

$\log m_{x,t} = \alpha_x + \beta_x\,\kappa_t,$2

Credible bands (e.g., 90%) are obtained by summarizing the empirical quantiles of projected mortality rates or derived quantities (e.g., period/cohort life expectancy) across posterior draws (Hilton et al., 2018). This inherently probabilistic approach enables comprehensive quantification of both parameter and process uncertainty, which is increasingly required in pension, insurance, and regulatory forecasting contexts.

Model averaging across competing transition-ages or structural choices is handled via Bayesian stacking, forming a weighted ensemble whose predictive distribution is:

$\log m_{x,t} = \alpha_x + \beta_x\,\kappa_t,$3

with $\log m_{x,t} = \alpha_x + \beta_x\,\kappa_t,$4 selected to maximize expected log predictive density.

5. Performance, Empirical Results, and Limitations

Empirical validation on UK data (1961–2013) demonstrates that stacking across models with different old-age cut-offs provides superior robust forecasts, with credible bands that largely conform to observed log-mortality in out-of-sample tests (88–92% of observations covered at 90% level) (Hilton et al., 2018). Some undercoverage can occur in rapidly changing subpopulations or ages where recent trend acceleration exceeds stochastic volatility captured in the model.

Joint-sex fitting with shared old-age asymptote and highly correlated period innovations substantially increases sampling efficiency and produces biologically plausible co-movement in male and female projections.

Life expectancy intervals under the Bayesian framework are wider than scenario-based official bands on short horizons and become narrower beyond 30 years, reflecting the proper propagation of parameter uncertainty.

Key limitations of the basic Lee–Carter model include:

• Inadequacy at very young and very old ages without explicit model extensions,
• Inability to accommodate one-off shocks or regime changes (e.g., pandemic years) except by scenario analysis,
• Implicit assumption of log-linearity, which may not always hold after structural breaks,
• Sensitivity to the handling of cohort effects or heteroscedasticity, addressed in state-space and Bayesian extended frameworks.

6. Practical and Regulatory Impact

The Lee–Carter class and its Bayesian GAM or state-space generalizations are core tools for life-insurance, social-security, and longevity-risk management. Their flexibility in handling joint-population inference, old-age smoothing, uncertainty quantification, and direct model averaging make them especially suited for modern stochastic projections required by Solvency II and similar regulatory regimes.

Key features enabling regulatory adoption include:

• Full probabilistic forecasting with credible bands,
• Direct scenario analysis for stress testing,
• Modular implementation in computational frameworks (e.g., Stan, R, Python) exploiting fast HMC or Kalman filtering,
• Empirical validation against holdout data and coverage properties.

Models in this class offer transparent, technically rigorous, and extensible methods for long-range mortality risk assessment, supporting actuarial pricing, reserving, and policy setting at scale (Hilton et al., 2018, Fung et al., 2016).