Previsão de mortalidade no Brasil com séries temporais curtas: uma abordagem com combinação de preditores e transferência de aprendizado

Abstract

This study addresses the challenge of mortality rate forecasting in Brazil using short time series, exploring and comparing statistical, machine learning, and hybrid methodologies. The central objective was to develop and evaluate a forecasting system that enhances the accuracy of demographic projections, considering the specific characteristics of Brazilian data. Statistical

models (Lee-Carter, FDM), classical time series models (ARIMA, ETS), deep learning architec- tures with transfer learning (CNN, GRU, and a hybrid CNN-GRU), and forecast combination

techniques were applied. Deep learning models were pre-trained on the Human Mortality Database and fine-tuned with Brazilian data from 2000 to 2015 to forecast the period 2016 to 2019. The results indicate a notable complementarity between approaches. The classical time series models, ETS (and ARIMA, closely), emerged as robust benchmarks, consistently leading in aggregate metrics (RMSE, MAE, and sMAPE), particularly benefiting from the smoothed nature of official data. Deep learning models with transfer learning did not surpass classical models in overall performance but showed localized gains in critical age groups, such as age 0 and the male accident hump. Regarding combinations, a nuanced picture emerged: inverse error weighting reduced RMSE/MAE for males and the total group, while simple averaging remained competitive in sMAPE, especially for females and in aggregate. The study concludes that while established time series models are highly effective for Brazilian mortality data, transfer learning and combination strategies offer complementary value in capturing complex dynamics within specific segments.