Mortality forecasting at age 65 and above: an age-specific evaluation of the Lee-Carter model

The Lee-Carter (LC) model has been recurrently used to forecast mortality, including by national statistical offices. This model has many advantages, but tends to underpredict life expectancy due to its assumption of constant age-specific response to the time index. Does this bias emerge from all ages or from specific ages only? In this paper, we aim to provide a more detailed evaluation of the model, by evaluating its accuracy, bias and robustness by age. Our analysis is based on 806 out-of-sample forecasts using various different fitting periods and forecast horizons. We focus on age 65 and above, as most deaths occur at older ages nowadays. We showed that the LC model is not fundamentally biased, but its main assumption of constant age-specific response to the time index, often leading to constant rates of mortality improvement, is not appropriate in all populations and at all ages.

Automated summary

This study provides a detailed evaluation of the Lee-Carter model in terms of accuracy, bias, and robustness by age, showing that its assumption of constant age-specific response to the time index may not be appropriate in all populations and at all ages.