Metodi per la previsione delle probabilità di sopravvivenza con un'applicazione alla popolazione italiana dal 1872 al 2018

The objective is the creation of algorithms that can be used to forecast survival probabilities. A set of mortality laws is briefly presented, then it is fitted to period life tables for both female and male Italian populations from 1872 to 2018. After the fitting procedure, 1-Year and 20-year forecasts of survival probabilities are computed via time series forecasting. In particular, the forecasting methods employed are: ETS, Bagged ETS, ARIMA, NNAR (Neural Network Autoregression) models. Forecasts are computed in two different ways: time series forecasts applied directly to the time series of survival probabilities for each age and forecasts based on the time series of parameters of models. After these forecasts are obtained, some algorithms that combine the different estimates are applied. These Forecasting Composite Algorithms (FCAs) take a weighted average of all the forecasts to compute the definitive forecast. Weights are assigned on the base of previous forecasting errors. Based on the 20-Year predictions made at the end of 1998, cohort survival functions for 35, 65 and 80 years old people are estimated.