Infectious disease forecasting using multi-model ensembles

Abstract

Accurate short-term forecasting of infectious disease hospitalisations is critical for health- care resource allocation and timely public health interventions. Multi-model ensemble forecasts, which combine predictions from several independent models into a single ag- gregate, have consistently outperformed individual models in real-world forecasting hubs, yet the theoretical reasons for this remain incompletely understood. This thesis develops a probabilistic framework for multi-model ensemble aggregation that makes explicit the role of bias, error variance, model correlation, and outlying components in determining ensemble accuracy. The framework is used to characterise and compare mean, median, and weighted aggregation strategies, with particular attention to their robustness under realistic error structures. The theoretical predictions are evaluated empirically using two simulated datasets of COVID-like hospitalisation trajectories: a synthetic pandemic dataset of 324 epidemic scenarios generated with CovaSim using Swedish demographic parameters, and a set of 96 counterfactual COVID-19 scenarios for Norway. Fifteen forecasting models spanning statistical, autoregressive, and mechanistic approaches are used as ensemble components. The results confirm that the unweighted median ensemble consistently produces accu- rate and stable forecasts, outperforming both individual models and weighted ensemble alternativesontheCovaSimdatasetandperformingcompetitivelyontheNorwegiancoun- terfactuals. This robustness is explained theoretically by the median’s bounded sensitivity to outlying components and favourable bias properties, at the cost of a efficiency penalty increasing MSE, a trade-off that favours the median given the heavy-tailed error distri- butions observed empirically. Weighted ensembles show modest advantages only where component model performance is stable across scenarios, indicating that the median is a robust and reliable aggregation strategy in novel disease settings where no historical per- formance data is available. Ensemble performance is found to be fundamentally limited by pairwise correlation among component model errors, with diversity and individual model skill acting as complementary rather than substitutable requirements.