A1302
Title: Dynamic matrix factor model for counts data
Authors: Han Xiao - Rutgers University (United States) [presenting]
Abstract: A dynamic factor model is considered for matrix time series, where the observations are counts data. The model is formulated as Poisson observations conditional on the rate matrices, which have log-normal distributions. The logarithm of the rate matrices has the form of a dynamic Gaussian factor model of matrix time series, where the dynamics are captured by the factor process. A log moment method is proposed to estimate the loading matrices, and use the variational inference to estimate the factor process. An autoregressive model is imposed on the factor process to enable predictions. It is also considered to include a trend component in the log rate matrices to capture possible nonstationarity. Theoretical and numerical analyses are conducted for the proposed model.
