Embracing the power of ensemble forecasting: A novel hybrid approach for advanced predictive modeling

Abstract

Amidst the persistent threat of epidemics, effectively managing their complexities requires accurate forecasting to anticipate their trajectory, thus enabling the preparation and implementation of effective mitigation strategies. With a special emphasis on COVID-19, the present work focuses on the Omicron variant, recognizing its significance in the global context of infectious diseases. The proposed research evaluates the effectiveness of both univariate and multivariate frameworks utilizing statistical and deep learning approaches to forecast the spread of the epidemic. Forecasting robustness is boosted by effectively correlating linear and non-linear components with the original series. To improve the performance, correlation is facilitated using correlation-driven weights within the statistically enforced deep learning model (WD-ensemble framework). The modeling process utilizes 493 data points and multivariate time-series records, including infected cases, vaccinated cases, and stringency index. The training dataset spans from November 1, 2021, to January 17, 2023, while the testing dataset covers the period from January 18, 2023, to March 8, 2023. The proposed WD-ensemble framework, incorporating stochasticity, outperforms all other state-of-the-art models, yielding highly reliable forecasts with remarkably low RMSE of 907.54, MAPE of 0.0008, and MAE of 670.78. It demonstrates a reduction in error percentages compared to the top-performing existing model, with decreases of 30.0267% in RMSE, 20% in MAPE, and 24.9411% in MAE. A pivotal revelation in this research is the robust negative correlation (-0.86) between vaccinated and confirmed cases as compared to the stringency index, implying that widespread vaccination could warrant the relaxation of stringent measures, including business and school closures.