FORECASTING MORTALITY AND LIFE EXPECTANCY: EVIDENCE FROM DYNAMIC CONDITIONAL SCORE (DCS) MODELS.

Abstract

This thesis modifies the Lee-Carter model by adopting DCS models to describe the mortality index kt , and aims to examine if this modification can overcome some limitations of the original Lee-Carter model, which is based on several fundamental assumptions that have been violated empirically. Using DCS class models is advantageous as it allows the specification of skewed and/or heavy-tailed distributions and also, it offers flexibility by letting distribution parameters vary over time. We select an appropriate conditional distribution for ?kt by examining whether the PITs are i.i.d. Uniform[0, 1], and compare the goodness-of-fit of different DCS models using AIC and BIC. Forecast accuracy is evaluated in a pseudo-out-of-sample forecasting exercise. We find that the modified Lee-Carter has an edge in forecasting demographic variables that are more volatile, and although shorter-term point forecasts from both models are similar, the implied trends are different. Thus, choice of model becomes important in long-term forecasting.