Predicting spatio-temporal flows is a challenging problem as dynamic spatio-temporal data possess underlying complex interactions and nonlinearities. Traditional statistical modeling approaches use a data generating process, generally motivated by physical laws or constraints. Deep learning (DL) is a form of machine learning for nonlinear high dimensional data reduction and prediction. With its theoretical roots in the Kolmogorov-Arnold representation theorem [Kolmogorov, 1963], it applies layers of hierarchical hidden variables to capture these interactions and nonlinearities without using a data generating process.
Building on joint work with co-authors Vadim Sokolov (GMU) and Nick Polson (Booth), this talk uses a Bayesian perspective of DL to explain its application to the prediction and uncertainty quantification of spatio-temporal flows. To illustrate our methodology, we predict the sharp discontinuities in traffic flow data, and secondly, we predict short-term futures market prices as a function of the order book depth. Finally, we highlight collaborations with various researchers addressing practical applications, including spatio-temporal models for actuarial risk (e.g., climate insurance, flood insurance) and a recent Google Summer of Code project.