An empirical evaluation of mathematical, machine learning, and hybrid modeling approaches for COVID-19 forecasting

Abstract

Accurate forecasting of infectious disease spread is essential for public health decision-making. This study compares mechanistic susceptible-exposed-infectious-removed (SEIR) models, machine-learning methods, and hybrid artificial intelligence (AI)-driven epidemiological frameworks using coronavirus disease 2019 (COVID-19) case data from India, the United States, Italy, and Japan. After preprocessing and exploratory analysis, we estimate time-varying transmission rates and develop two hybrid extensions: one predicts beta(t) using machine learning, and the other learns residual errors from the SEIR model. A baseline machine-learning model using lag-based and effective reproduction number (Rt)-driven features is also evaluated. Performance is assessed using mean absolute error (MAE), root mean squared error (RMSE), mean absolute percentage error (MAPE), high-incidence MAPE (on days with observed incidence ge 100 cases), and peak-timing error. With the SEIR transmission-rate spline fitted strictly to training data and all models evaluated on an identical 223-day held-out test window, a simple persistence baseline outperforms the mechanistic SEIR model, both hybrid AI extensions (Tracks A and B), and the pure machine-learning Random Forest model across all four countries (Diebold--Mariano test, p<0.001 in every comparison). The hybrid AI components do not consistently improve forecasting accuracy relative to the SEIR baseline. None of the mechanistic or machine-learning approaches anticipates new epidemic waves beyond the training period, underscoring the difficulty of medium-range epidemic forecasting under genuinely held-out conditions. The comparative evaluation highlights the complementary strengths and limitations of these approaches and provides guidance for future epidemic forecasting models.