Short-Range Forecasting of Coronavirus Disease 2019 (COVID-19) During Early Onset at Various Levels of Geographic Granularity Using Seven Different Methods

Abstract

Background: The coronavirus disease 2019 (COVID-19) spread rapidly across the world since its appearance in Wuhan city, Hubei province, China in December 2019. We explore the effectiveness of forecasting methods during early onset for COVID-19 case counts over the upcoming seven-days at the county, health district, and state levels. Our goal is to assess which forecasting methods can aid decision makers about anticipated one-, three-, and seven-day case growth.Methods: We apply forecasting methods across VA’s counties and independent cities, its health districts, and the state using data reported by the The New York Times. We forecast one-, three-, and seven-days forward using: (1) a naïve approach; (2) Holt-Winters exponential smoothing (HW); (3) growth rate (Growth); (4) moving average (MA); (5) autoregressive (AR); (6) autoregressive moving average (ARMA); and (7) autoregressive integrated moving average (ARIMA). Using Median Absolute Error (MdAE) and Median Absolute Percentage Error (MdAPE), we compare the resulting 216698 forecasts to VA’s historical data through April 22, 2020.Findings: Single-day MA forecast with three-day lookback obtains the lowest MdAE and is statistically significantly different than 66 percent of evaluated alternatives at each geographic level. Furthermore, methods assuming stationary means of prior days’ case counts outperform methods with assumptions of weak- or non-stationarity. Finally, for an individual forecasting method, MdAPE evaluation reveals in most cases statistically significant differences across geographic levels demonstrating that there is not a one-size-fits-all solution to short-term COVID-19 forecasting.Interpretation: For decision makers applying short-term forecasts of COVID-19: (1) MA is an effective option for forecasting cumulative case counts one-, three-, and seven-days forward; (2) methods assuming stationarity of means in prior-observations are more effective than methods assuming weak- or non-stationary means; and (3) the intended level of geographic resolution should be a factor in selecting the appropriate forecasting method.Funding Statement: This work received no external funding.Declaration of Interests: All authors declare no competing interests.