Abstract
The method of mortality forecasting proposed by Lee and Carter describes a time series of age‐specific log‐death rates as a sum of an independent of time age‐specific component and a bilinear term in which one of the component is a time‐varying factor reflecting general change in mortality and the second one is an age‐specific parameter. Such a rigid model structure implies that on average the mortality improvements for different age groups should be proportional, regardless of the calendar period: a single time factor drives the future death rates. This paper investigates the use of multivariate time series techniques for forecasting age‐specific death rates. This approach allows for relative speed of decline in the log death rates specific to the different ages. The dynamic factor analysis and the Johansen cointegration methodology are successfully applied to project mortality. The inclusion of several time factors allows the model to capture the imperfect correlations in death rates from 1 year to the next. The benchmark Lee–Carter model appears as a special case of these approaches. An empirical study is conducted with the help of the Johansen cointegration methodology. A vector‐error correction model is fitted to Belgian general population death rates. A comparison is performed with the forecast of life expectancies obtained from the classical Lee–Carter model. Copyright © 2009 John Wiley & Sons, Ltd.