Quantifying Uncertainty in Deep Spatiotemporal Forecasting

Quantifying uncertainty is critical to risk assessment and decision making in high stakes domains. However, prior works for deep neural network uncertainty estimation have mostly focused on point prediction. A systematic study of uncertainty quantification methods for spatiotemporal forecasting has been missing in the community. In this paper, we analyze forecasting uncertainty in spatiotemporal sequences from both the Bayesian and Frequentist point of view via statistical decision theory. We further conduct case studies to provide an empirical comparison of Bayesian and Frequentist uncertainty estimation techniques. Through experiments on traffic and COVID-19 forecasts, we conclude that, with a limited amount of computation, Bayesian methods are typically more robust in mean prediction, while Frequentist methods are more effective in estimating the confidence levels.