A Comparative Analysis of Forecasting Models on COVID-19

ABSTRACT

The COVID-19 spread all around the world, causing more than a million deaths and reaching over 50 million confirmed cases. A forecast of these numbers is vital for the adequate preparations of health care capacities and for the governments to take the necessary decisions. In this study, it is aimed to predict the evolution of COVID-19 figures, employing alternative statistical models such as the Holt-Winters, ARIMA, and ARIMAX while using the time series corresponding to different parameters of this disease such as daily cases, daily deaths, and the stringency index. Considered are the John Hopkins University epidemiological world data and the top ten countries with the highest cases, along with China. The fitting of the time series and the upcoming 10 days projections resulted in a high level of accuracy, presented with alternative error metrics and comparisons between the situations of countries. Holt-Winters is the best performing model, while ARIMAX gives the worst accuracy results. Moreover, it was found that the use of coefficient determination and Bayesian information criterion alone are not suitable, and scale independent metrics should be employed when the data ranges differ. The results of this study would be useful to set up benchmark results for other studies and the projections may be used for medical, economic, and social precaution and preparation. © 2022, The Author(s), under exclusive license to Springer Nature Switzerland AG.