Short-term forecasting of COVID-19 using support vector regression: An application using Zimbabwean data

Area of Science:

• Statistical modeling
• Machine learning applications
• Predictive analytics

Background:

• Accurate forecasting is crucial across various scientific disciplines.
• Traditional models often struggle with complex covariate relationships.
• Improving prediction accuracy enhances decision-making and resource allocation.

Purpose of the Study:

• To evaluate the impact of pairwise hierarchical interactions on prediction accuracy.
• To assess the benefits of combining forecasts from individual predictive models.
• To identify the optimal modeling strategy for enhanced forecasting.

Main Methods:

• Utilized the least absolute shrinkage and selection operator (LASSO) for variable selection.
• Employed Gradient Boosting Method (GBM), Generalized Additive Models (GAMs), and Support Vector Regression (SVR) for individual forecasts.
• Combined forecasts using linear quantile regression averaging (LQRA) and evaluated performance using Mean Absolute Error (MAE).

Main Results:

• Incorporating pairwise interactions improved forecast accuracy in single models (GBM, GAMs, SVR).
• The SVR model with radial basis kernel function and interactions yielded the lowest MAE among single models.
• Combining forecasts via LQRA further reduced MAE, indicating ensemble benefits.

Conclusions:

• Ensemble methods, by combining predictions from multiple models, demonstrably improve forecast accuracy.
• The Support Vector Regression model incorporating pairwise hierarchical interactions proved to be the most effective overall.
• The study validates the utility of interaction terms and model combination for robust predictive modeling.