A Quantitative Comparison of the Lee-Carter Model under Different Types of Non-Gaussian Innovations

Abstract

In the classical Lee-Carter model, the mortality indices that are assumed to be a random walk model with drift are normally distributed. However, for the long-term mortality data, the error terms of the Lee-Carter model and the mortality indices have tails thicker than those of a normal distribution and appear to be skewed. This study therefore adopts five non-Gaussian distributions—Student’s t-distribution and its skew extension (i.e., generalised hyperbolic skew Student’s t-distribution), one finite-activity Lévy model (jump diffusion distribution), and two infinite-activity or pure jump models (variance gamma and normal inverse Gaussian)—to model the error terms of the Lee-Carter model. With mortality data from six countries over the period 1900–2007, both in-sample model selection criteria (e.g., Bayesian information criterion, Kolmogorov–Smirnov test, Anderson–Darling test, Cramér–von-Mises test) and out-of-sample projection errors indicate a preference for modelling the Lee-Carter model with non-Gaussian innovations.