Adaptive Public Health Forecasting with ML and ARIMA Averaging

Public health forecasting requires models that can quickly adapt to sudden shifts in surveillance data without overreacting to noise or temporary trends. This research compares the performance of traditional autoregressive integrated moving average (ARIMA) models with machine learning approaches like random forest and XGBoost for forecasting public health outcomes. The study used 190 weekly observations of Ontario COVID-19 case counts from 2020 to 2023, employing rolling-origin time-series cross-validation to maintain temporal order. The evaluation focused on three dimensions: responsiveness after turning points, forecast horizons from one to six weeks, and the amount of historical training data. ARIMA models demonstrated rapid adaptation post-turning points but showed increased error at longer horizons. Conversely, random forest and XGBoost were less immediately responsive but maintained more stable error over extended periods. For shorter forecasts, training on recent data proved more effective, especially for XGBoost. A key contribution is the development of Machine Learning and ARIMA Model Averaging (MLAMA), a non-negative performance-weighted ensemble. MLAMA's weights dynamically adjust based on forecast horizon and responsiveness settings. This ensemble consistently achieved the lowest normalized mean absolute percentage error across most forecast horizons, highlighting its practical utility in combining the complementary strengths of statistical and machine learning forecasts.