When in Doubt: Neural Non-Parametric Uncertainty Quantification for Epidemic Forecasting

Accurate and trustworthy epidemic forecasting is an important problem that\nhas impact on public health planning and disease mitigation. Most existing\nepidemic forecasting models disregard uncertainty quantification, resulting in\nmis-calibrated predictions. Recent works in deep neural models for\nuncertainty-aware time-series forecasting also have several limitations; e.g.\nit is difficult to specify meaningful priors in Bayesian NNs, while methods\nlike deep ensembling are computationally expensive in practice. In this paper,\nwe fill this important gap. We model the forecasting task as a probabilistic\ngenerative process and propose a functional neural process model called EPIFNP,\nwhich directly models the probability density of the forecast value. EPIFNP\nleverages a dynamic stochastic correlation graph to model the correlations\nbetween sequences in a non-parametric way, and designs different stochastic\nlatent variables to capture functional uncertainty from different perspectives.\nOur extensive experiments in a real-time flu forecasting setting show that\nEPIFNP significantly outperforms previous state-of-the-art models in both\naccuracy and calibration metrics, up to 2.5x in accuracy and 2.4x in\ncalibration. Additionally, due to properties of its generative process,EPIFNP\nlearns the relations between the current season and similar patterns of\nhistorical seasons,enabling interpretable forecasts. Beyond epidemic\nforecasting, the EPIFNP can be of independent interest for advancing principled\nuncertainty quantification in deep sequential models for predictive analytics\n