CNN-LSTM and GNN-LSTM models added to a Lee-Carter baseline reduce test MSE by about 24% versus MortFCNet on French regional mortality data from 1990-2019, with largest gains at oldest ages.
A penalized distributed lag non-linear Lee-Carter framework for regional weekly mortality forecasting
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
abstract
Accurate forecasts of weekly mortality are essential for public health and the insurance industry. We develop a forecasting framework that extends the Lee-Carter model with age- and region-specific seasonal effects and penalized distributed lag non-linear components that capture the delayed and non-linear effects of heat, cold, and influenza on mortality. The model accommodates overdispersed mortality rates via a negative binomial distribution. We model the temporal dynamics of the latent factors in the model using SARIMA processes and capture cross-regional dependencies through a copula-based approach. Using regional French mortality data (1990-2019), we demonstrate that the proposed framework yields well-calibrated forecast distributions and improves predictive accuracy relative to benchmark models. The results further show substantial heterogeneity in temperature- and influenza-related relative risks between ages and regions. These findings underscore the importance of incorporating exogenous drivers and dependence structures into a weekly mortality forecasting framework.
fields
stat.AP 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Climate-Driven Mortality Forecasting Using Deep Learning
CNN-LSTM and GNN-LSTM models added to a Lee-Carter baseline reduce test MSE by about 24% versus MortFCNet on French regional mortality data from 1990-2019, with largest gains at oldest ages.