All Models Are Useful: Bayesian Ensembling for Robust High Resolution COVID-19 Forecasting.

Area of Science:

• Epidemiology
• Computational Biology
• Data Science

Background:

• Accurate infectious disease forecasting is crucial for public health policy and resource management.
• Existing COVID-19 forecasting models show variable performance due to data noise and pandemic dynamics.
• High-resolution, timely forecasts are needed for effective intervention strategies.

Approach:

• Developed a real-time forecasting pipeline integrating statistical, machine learning, and mechanistic models.
• Employed a Bayesian ensembling scheme to combine probabilistic forecasts from diverse methods.
• Operationalized the pipeline for nearly six months, serving US policymakers at local, state, and federal levels.

Key Points:

• The Bayesian ensemble demonstrated performance comparable to or better than individual models.
• Each contributing model offered unique strengths across different spatial regions and time points.
• The ensemble approach improved forecast accuracy, particularly at longer forecast horizons compared to other models in the CDC COVID-19 Forecast Hub.

Conclusions:

• A performance-based ensemble of multiple forecasting methods can yield robust, high-resolution predictions for infectious diseases.
• Such integrated pipelines are essential for real-time pandemic response and informed decision-making.
• The developed system enhances lead time for mechanistic scenario projections, supporting proactive public health planning.