Evaluating How the Social Restriction, the Government Response, the Health, and Economic Indices Affected the Prediction of the Number of Deaths Provoked by COVID-19 in Brazil Using Classical Statistical and Machine Learning Models

Abstract

The COVID-19 death predictions are helpful for the formulation of public policies, allowing the use of more effective social isolation strategies with less economic and social impact. This article evaluates a wide range of forecasting methods to identify the best models for predicting cumulative and daily deaths caused by COVID-19 in Brazil, considering a forecast for a seven-day horizon. With the seven-day horizon, the predictions have more accuracy. The dataset is from Oxford Covid-19 Government Response Tracker. The jackknife resampling technique was implemented, thus providing an accurate estimate for evaluating the predictive capacity of the models. Each model was fitted with 266 jackknife samples considering 30-day training bases. The comparison between predictions was made using the average results, considering R², MAPE, RMSE, and MAE. Models from different classes were adopted: 1 ETS, 4 ARIMA, 18 regression models, and 7 machine learning algorithms. The cumulative death models produce better results than daily deaths, as the cumulative death models are less influenced by time series components: cycle and seasonality. The best results for predicting daily deaths were attained by the Ridge regression method. The best results for predicting cumulative deaths were obtained by the Cubist regression method.

Keywords:
COVID-19 Deaths; Jackknife; Statistical models; Regression models; Machine learning.

HIGHLIGHTS

• Thirty statistical and machine learning models have been used to predict COVID-19 in Brazil.

• The time series of accumulated deaths produces better estimates than daily deaths.

• The Cubist nonlinear regression model provides better predictions of accumulated deaths.