Jie Jian, University of Chicago

Title: Bayesian Poisson-Randomized Gamma Tensor Factorization with Application to International Trade Flows

Date: Thursday, May 7th, 2026
Time: 1:30PM (PDT)
Location: ASB 10900

Abstract: We develop a Bayesian tensor factorization for nonnegative, semi-continuous multiway data exhibiting excess zeros, heavy right tails, and heterogeneous dispersion across slices. Such features are pervasive in monetary-valued panel data, including international trade, where most cells of the exporter × importer × product × year tensor are zero while positive entries span several orders of magnitude. We propose a hierarchical model that places a low-rank nonnegative CP decomposition on a latent Poisson rate and couples it with a conditional Gamma likelihood, a construction related to the Tweedie family. We develop a hybrid variational--Monte Carlo algorithm that scales posterior inference to tensors with tens of millions of entries. Applied to international trade data, the model outperforms competing tensor-based baselines in out-of-sample prediction and recovers interpretable multiway dependence across exporters, importers, products, and time that is difficult to recover from gravity-type or pairwise network analyses operating on fewer dimensions.