Activity Number:
|
43
|
Type:
|
Contributed
|
Date/Time:
|
Sunday, July 31, 2016 : 2:00 PM to 3:50 PM
|
Sponsor:
|
Section on Bayesian Statistical Science
|
Abstract #319569
|
|
Title:
|
A Class of Bayesian Multivariate Time Series Models for Counts
|
Author(s):
|
Refik Soyer* and Tevfik Aktekin and Nicholas Polson
|
Companies:
|
The George Washington University and University of New Hampshire and The University of Chicago
|
Keywords:
|
Multivariate counts ;
Particle learning ;
Particle filtering ;
NonGaussian time series
|
Abstract:
|
We consider modeling of multivariate time-series of correlated counts which often arise in finance, operations and marketing applications. Dependence among series arises as a result of sharing a common environment. We consider a class of multivariate Poisson time series models by assuming a common environmental process modulating the rates of the individual series. This setup gives us a class of dynamic multivariate negative binomial time series. We develop Bayesian inference for these models using particle filtering and Markov chain Monte Carlo methods. A by-product of particle filtering in our set up is predictive likelihoods which we refer to as multivariate confluent hyper-geometric negative binomial distribution. We discuss issues of sequential filtering, smoothing and prediction and illustrate the proposed models using a simulated data set as well as actual data on weekly household shopping trips.
|
Authors who are presenting talks have a * after their name.