Forecasting of COVID-19 Fatality in the USA: Comparison of Artificial Neural Network-Based Models

Abstract

The first death caused by the novel coronavirus in the USA was declared on February 29, 2020, in the Seattle area in Washington state. Forecasting the number of deaths has great importance in terms of public psychology and strategic decisions to be taken by the government. There are several data-driven models in the literature to predict the deaths in the USA caused by COVID-19. However, most of them are based on a few variables of the data for forecasting. From this point of view, this study provides an artificial neural network (ANN)-based approach by considering 12 different variables for forecasting the cumulative deaths caused by COVID-19 in the USA. The proposed ANN structure was trained with three algorithms, namely scaled conjugate gradient algorithm, Levenberg-Marquardt algorithm and Bayesian regularization algorithm. These three forecasting models were constructed on 13 parameters such as 12 inputs and one output. The sensitivity and performance of the proposed forecasting models were analyzed and compared by using indices mean absolute error, mean absolute percentage error, correlation coefficient (R-value), sum square error, variance account for, mean square error and root-mean-square error. Results show that the forecasting model with Bayesian regularization performs better than other models for forecasting the cumulative deaths due to COVID-19 in the USA.