A Comparison of Time-Series Models in Predicting COVID-19 Cases

Abstract

Objective: As the world is striving to control the COVID- 19 pandemic, one aspect of this strife is to project how many new cases will be experienced in the coming days so that health services can better be planned and real time health policies can be developed depending on the spread of the pandemic, its speed and direction. To address this, we compared time-series modeling approaches as to which one more accurately projects how many new cases to expect within 5 days. Material and Methods: In this research, we used the accumulating COVID-19 cases for all countries since the beginning of the pandemic in China in December 31, 2019, and aimed at identifying best time-series model to project COVID-19 cases and deaths. Results: We showed that Conditional Lest Square modeling with AR(1) auto-correlation structure should be the model to be chosen for case projections. For death projections, Conditional Lest Square modeling with AR(2) auto-correlation structure showed slightly beter performance than its compatibles, and should be the model of choice. We also observed that the observed level of confidence interval is lower than its expected level. Conclusion: Future cases and dates due to COVID-19 can be projected successfully with time-series models with Conditional Lest Square modeling using AR(1) auto- correlation structure for cases and AR(2) auto-correlation structure deaths.