MODELING COVID-19 MORTALITY IN BRAZIL USING STATISTICAL AND MACHINE LEARNING METHODS WITH DATA ON VACCINATION, VIRUS REPRODUCTION RATE, AND LOCKDOWN

The COVID-19 death predictions help formulate public policies, allow the use of more effective vaccination and social restriction strategies. This article evaluates a wide range of forecasting methods to identify the best models for predicting the deaths caused by COVID-19 in Brazil, considering a horizon of seven, 15, and 30 days. The dataset is from Our World in Data. The moving average of seven days was adopted as a reference for this time series to improve the data quality of the deaths time series. Models from different classes were adopted: one ETS, two ARIMA, 12 regression models, and 14 machine learning algorithms. Each model was adjusted considering a starting training base of 30 values. The models have been trained and tested 212 times considering a 7-day horizon, 204 times considering a 15-day horizon, and 189 times assuming a 30-day horizon. The comparison between predictions was made using the average results, considering RMSE and MAE. The methodology applied here could be used by other researchers, making it possible to practice more effective public policies in future epidemics. The ARIMA and ARIMAX models showed the best results for short-term and medium-term forecasting, and the cubist model showed the best results for long-term forecasting. The ARIMAX was the best fit to predict a short horizon, and the Cubist regression was the best fit for longer horizons.