Trend Analysis and Intervention Effect Starting Point Detection of COVID-19 Epidemics Using Recalibrated Time Series Models — Worldwide, 2020

Objective: This study aimed to identify a model for short-term coronavirus disease 2019 (COVID-19) trend prediction and intervention evaluation. Methods: We compared the autoregressive integrated moving average (ARIMA) model and Holt exponential smoothing (Holt) model on predicting the number of cumulative COVID-19 cases in China. Based on the mean absolute percentage error (MAPE) value, the optimal model was selected and further tested using data from the United States, Italy and Republic of Korea. The intervention effect starting time points and abnormal trend changes were detected by observing the pattern of differences between the predicted and real trends. Results: The recalibrated ARIMA model with a 5-day prediction time span has the best model performance with MAPEs ranged between 2% and 5%. The intervention effects started to show on February 7 in the mainland of China, March 5 in Republic of Korea and April 27 in Italy, but have not been detected in the US as of May 19. Temporary abnormal trends were detected in Korea and Italy, but the overall epidemic trends were stable since the effect starting points. Conclusion: The recalibrated ARIMA model can detect the intervention effects starting points and abnormal trend changes; thus to provide valuable information support for epidemic trend analysis and intervention evaluation.