Comparative Analysis Of Predictive Performance Of Holt-Winters And Facebook Prophet On Kenyan Covid-19 Data

Abstract

Comparative time-series forecasting is essential for identifying and predicting the trajectories of infections such as COVID-19. This study offers a thorough comparison of two time series forecasting methodologies: The Holt-Winters (HW) exponential smoothing technique and Meta’s (Facebook’s) Prophet model, as applied to COVID-19 case data. We assessed each model’s capacity to capture trend dynamics, seasonal variations, and sudden structural shifts linked to pandemic waves and policy measures, using publicly accessible epidemiological time-series data. The Holt-Winters model, which focuses on level, trend, and seasonality components, offers a clear foundation for short-term forecasting but has shortcomings when faced with irregular shocks and non-linear patterns. Conversely, Prophet’s decomposable additive structure, which includes automatic changepoint detection and variable seasonality, exhibits superior flexibility to sudden changes in the transmission patterns. Forecast accuracy was evaluated using conventional error metrics (Root Mean Square Error; RMSE, Mean Absolute Error; MAE, R-Squared; R2), indicating that Holt-Winters (HW) typically surpassed Facebook Prophet in both daily and cumulative confirmed cases of COVID-19. The comparative analysis emphasizes the significance of model selection according to the epidemic environment and illustrates the advantages of conventional time-series techniques for reliable public health forecasting. This study provides methodological insights for academics and decision-makers in pursuit of efficient methods for monitoring and forecasting the dynamics of infectious diseases.