Assessing Statistical Methods for Generating Forecasts for COVID-19

Abstract

The COVID-19 pandemic, a persistent global health emergency that has affected almost all facets of daily life, was initially discovered in Wuhan, China, in December 2019. Since that time, the virus has rapidly spread over the globe, causing serious social and economic upheavals necessitating the need for reliable forecasting methods. This study compares ten distinct models to predict the number of confirmed COVID-19 cases in Sri Lanka, aiming to assess the performance of statistical models using limited and volatile real-world data characterized by trends, random peaks, and autocorrelations. In addition to the classical ARIMA model, various smoothing and filtering techniques were explored to capture the unique characteristics of the data. The model consistencies in multiple-day predictions were demonstrated, and robust evaluation criteria, along with non-robust measures, were utilized to enhance the effectiveness of the evaluation process. The results highlight the effectiveness of traditional smoothing and filtering strategies such as Simple Exponential Smoothing, Holt’s Exponential Smoothing, and the Smoothing Splines technique coupled with the ARIMA model. This study also discovered that the ARIMA model, when applied directly to the original data without using any smoothing or filtering approaches, failed to forecast adequately, thereby demonstrating the insufficiency of the ARIMA model on its own to provide credible forecasts when given a volatile set of data.