What to Forecast? An Exploratory Study of the Implication of Using Different Indicators to Forecast Mortality

Researchers and private and public institutions have suggested different ways to forecast mortality over the years. Many of these forecasting models are based on extrapolative methods of the age-specific death rates. However, more recent studies have looked into forecasting models based on other mortality indicators, such as life expectancy or the life table deaths. We here ask what are the implication of choosing one specific indicator over another to forecast mortality? We compare five models based on singular value decomposition of the matrix by age and time of different indicators: logged death rates, logit of death probabilities, logit of survival probabilities, log-ratio transformation of the life table deaths and life expectancy at birth. The results show that the time-indexes of all these indicators are very similar, but that forecasting using death rates and probabilities of death leads to more pessimistic forecasts than the other models.