Previsão de séries temporais para os óbitos no Brasil causados pela COVID-19 no âmbito da pandemia

• Autor(a) principal: Provenza, Marcello Montillo
• Data de Publicação: 2022
• Outros Autores: [e-mail]
• Tipo de documento: Tese
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações da UERJ
• Texto Completo: http://www.bdtd.uerj.br/handle/1/18970
• topic: Time series; Deaths; Statistical models; Machine learning; Engenharia química; Processamento de dados; COVID-19 (Doença); Séries temporais; Óbitos; Modelos estatísticos; Aprendizado de máquina; Econometria; COVID-19; COVID 19; ENGENHARIAS::ENGENHARIA QUIMICA

The predictions of deaths from COVID-19 are useful for the formulation of public policies, allowing the use of more effective social isolation strategies with less economic and social impact, in addition to promoting indicators of how the population adheres to vaccines. The objective of this work is to explore a broad set of prediction methods to identify the best models without vaccination coverage (Case 1) and with vaccination coverage (Case 2) in Brazil. The methods of Artificial Intelligence and the classical methods of econometrics were considered. The cross-validation technique for time series was implemented, thus providing an accurate estimate to assess the predictive capacity of the models. Each model was adjusted considering an initial training base of 30 values. In Case 1, daily and cumulative deaths from the Oxford COVID-19 Government Response Tracker database were used. In Case 2, the dataset comes from Our World in Data, where the seven-day moving average was adopted as a reference to improve data quality. In Case 1 the models were trained and tested with 266 samples considering a forecast horizon of 7 days. In Case 2 the models were trained and tested 494 times considering a forecast horizon of seven days, 486 times considering a horizon of 15 days, and 471 times assuming a horizon of 30 days. Models of different classes were adopted: ETS algorithms, ARIMA, regression, and machine learning. The forecasts were compared using the average results of the forecast metrics: R2, RMSE, MAE, and MAPE. In Case 1, the accumulated forecasts offered better results than the daily ones, as the models are less influenced by the components of the time series: cycle and seasonality. The best results for the prediction of daily deaths were obtained by the Ridge regression method (R2 = 0.772, RMSE = 136, and MAE = 113). The best results for predicting cumulative deaths were obtained by the Cubist regression method (R2 = 0.993, RMSE = 468, and MAE = 409). In Case 2, the ARIMA model with one differentiation showed the best results for a seven-day horizon (RMSE = 74 and MAE = 64).