TiDEFormer—a heterogenous stacking ensemble approach for time series forecasting of COVID-19 prevalence

Forecasting time series data over extended periods remains a formidable task in practical scenarios, such as the ongoing COVID-19 epidemic. The current variant of concern, JN.1, has increased transmissibility and reduced susceptibility to vaccinations in comparison to previous strains. As a result, there is an urgent requirement to forecast the daily incidence of COVID-19 in the near future. While deep learning models have demonstrated potential in predicting time series, they lack effectiveness in forecasting over long durations. This study seeks to fill the current gap by implementing a novel ensemble-based approach that incorporates two highly promising deep learning models: Time series Dense Encoder (TiDE) and Self attention-based Transformer model. The TiDEFormer, which combines TiDE and Transformer models using a heterogenous stacking ensemble technique, has exhibited greater accuracy in comparison to other proficient algorithms. The work employs the Blocked Time Series Cross validation technique to build distinct accurate models. In addition, the models are subjected to hyper-parameter tuning using the Grid Search Algorithm. The test results of TiDEFormer on the COVID-19 Dataset show a significant improvement in the Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) by around 22% and 17% respectively, compared to the TiDE model. The percentage improvement with respect to the Transformer model is approximately 9% and 6% respectively. Furthermore, an accurate prediction of the COVID-19 situation on daily basis is available for all countries, spanning a period of 200 days forecasting horizon covering approximately these months (From Dec 2023 to July 2024).