Forecasting the COVID-19 pandemic in the United States and Peru using ARIMA, LSTM, GRU, CNN, and a hybrid approach

Abstract

The COVID-19 outbreak spread swiftly and infected many individuals resulting in overwhelmed and overfilled hospitals causing an immense loss of life globally. Identifying the number of infected individuals preemptively provides critical time for governmental and health officials to implement a strategy to respond to the pandemic such as requiring masking, reducing public gatherings, closing restaurants, as well as additional time to prepare hospitals and medical staff for surges in infections. The work explores implementing convolutional neural network models (CNN), long short-term network models (LSTM), gated recurrent unit models (GRU), the combination of encoding CNN layers and decoding LSTM and/or GRU layers in a hybrid model, and Auto-Regressive Integrated Moving Average (ARIMA) models to predict COVID-19 case count in the United States and Peru for 7, 15 or 30 days in the future using 30 days of case counts. The study evaluates predictions from January 23, 2020 through March 9, 2023 for the United States and March 6, 2020 through April 2, 2023 for Peru. For each model, the forecasting results are displayed visually and presented statistically using RMSE and MAPE. The hybrid model performed as well as or better than any other model when predicting 7 days, 15 days, or 30 days into the future. These results demonstrate models that potentially assist healthcare providers and policymakers’ response to the spread of COVID-19.