Time series forecasting of the COVID-19 pandemic: a critical assessment in retrospect

Abstract

The COVID-19 pandemic is perceived by many to have run its course, and forecasting its progress is no longer a topic of much interest to policymakers and researchers as it once was. Nevertheless, in order to take lessons from this extraordinary two and a half years, it still makes sense to have a critical look at the vast body of literature formed thereon, and perform comprehensive analyses in retrospect. The present study is directed towards that goal. It is distinguished from others by encompassing all of the following features simultaneously: (i) time series of 10 of the most affected countries are considered; (ii) forecasting for two types of periods, namely days and weeks, are analyzed; (iii) a wide range of exponential smoothing, autoregressive integrated moving average, and neural network autoregression models are compared by means of automatic selection procedures; (iv) basic methods for benchmarking purposes as well as mathematical transformations for data adjustment are taken into account; and (v) several test and training data sizes are examined. Our experiments show that the performance of common time series forecasting methods is highly sensitive to parameter selection, bound to deteriorate dramatically as the forecasting horizon extends, and sometimes fails to be better than that of even the simplest alternatives. We contend that the reliableness of time series forecasting of COVID-19, even for a few weeks ahead, is open to debate. Policymakers must exercise extreme caution before they make their decisions utilizing a time series forecast of such pandemics.