Forecasting Daily COVID-19 Cases in Kenya: A Comparative Study of Three Time Series Models

TLDR

These models were least unsuitable for real-time daily forecasting of epidemics in resource-constrained settings with sparse data and irregular surveillance reporting and recommend incorporating these models into routine disease surveillance in resource-scarce settings such as Kenya.

Abstract

Background: For health system preparedness, understanding when to expect a dramatic change in the trend of an infectious disease is critical, especially when the data are readily available but limited. Kenya's COVID-19 surveillance was further constrained by inadequate infrastructure support for testing and weekend reporting, as well as for reporting on the Alpha, Delta, and Omicron variants. Methods: This study aimed to compare the predictive performance of the three univariate time series models. These are ARIMA, Holt-Winters Exponential Smoothing, and Facebook Prophet. The study used 1,387 daily confirmed COVID-19 cases collected in Kenya from March 15, 2020, to December 31, 2023. A quantitative comparative design with an 80/20 train-test split was used (training period: March 15, 2020, to March 28, 2023; test period: March 29, 2023, to December 31, 2023). The Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and coefficient of determination (R2) were used to compare the model accuracy. Results: The R2 statistics for the three models showed that they explained little of the variance in daily case counts, with all values ranging from .01 to .16. This was largely due to the high noise and the series's zero-heavy nature. Through this limited scope, Holt-Winters ETS(A,AD,A) resulted in the lowest point errors (RMSE = 23.48, MAE = 4.00, R2 = 0.16) followed by ARIMA(5,1,4) (RMSE = 24.70, MAE = 5.09, R2= 0.01). Facebook Prophet returned substantially larger errors across all three metrics (RMSE = 82.80, MAE = 38.10, R² = 0.03), with RMSE nearly 3.5 times that of Holt-Winters. Conclusion: These models were least unsuitable for real-time daily forecasting of epidemics in resource-constrained settings with sparse data and irregular surveillance reporting. We recommend incorporating these models into routine disease surveillance in resource-scarce settings such as Kenya, with recalibration adjusted for changes in epidemic seasonality and timing that depend on the epidemic phase.