Point forecasts of mortality rates in Malaysia: a comparison of principal component methods

Abstract

The problem of forecasting mortality has been widely discussed and is being considered as an important issue among actuaries and demographers. Using age and sex-specific mortality data in Malaysia, we compare and evaluate the point forecast accuracy of six principal component methods.  These six methods are the lee-carter (LC) method, the unadjusted LC method (LCnone), the lee-miller (LM) variant, the booth-mandonald-smith variant (BMS) and the two versions of the hyndman-ullah (HU). The overall performance of all methods was evaluated using out-sample forecasts based on the six error measures of three different data spanning with three different step-ahead forecasts. The data set consists of annual age-specific death rates (ASDR) for 17 age groups for both genders for the years 1980 to 2011.  Results show that when the fitted period is longer, the LM method exhibits a better fit in Malaysia male ASDR and the HU6 method attains a better fit in Malaysia female ASDR. When the fitted period becomes shorter, the BMS method outperforms for male and HU1 for female ASDR. However, for overall performance, the LM method consistently outperforms others for Malaysia male ASDR and the LM and HU methods are the best good of fit for Malaysia ASDR.