Forecasting mortality rates and life expectancy in the year of Covid-19

Forecasting mortality rates and life expectancy is an issue of critical importance made arguably more difficult by the effects the current Covid-19 pandemic. In this paper we compare the performances of a simple random walk model (benchmark), three variants of the standard Lee-Carter model (Lee-Carter, Lee-Miller, Booth-Maindonald-Smith), the Hyndman-Ullah functional data analysys model, and a general factor model. We use both symmetric and asymmetric loss functions, as the latter are arguably more suitable to capture preferences of forecast users such as insurance companies and pension and health system planners. In a counterfactual study, designed exploiting the particular features of Italian data, we reproduce the likely impact of Covid-19 on forecasts using 2020 as a jump-off year. To put the results in perspective, we also carry out out a general assessment on 1950-2016 data for three countries with very diverse demographic profiles, France, Italy and USA. While the results with these latter datasets suggest that in normal conditions the Lee-Miller and Hyndman-Ullah models are somehow superior,from the counterfactual study the best option appears to be the Booth-Maindonald- Smith model.