Forecasting Evolving Curves
This dissertation describes methodologies for forecasting and testing integer valued time series that consist of Poisson counts. In Chapter 2 we look at univariate time series in which the counts are large and can be approximated using Gaussian models. These models are motivated by data gathered from the call center of a large financial institute. The goal of the models in this application is to predict accurately the one-day-ahead arrival process of incoming calls to the call center. A secondary goal is to provide a Bayesian algorithm to dynamically update these forecasts as more data becomes available. To test the underlying Gaussian assumption we develop in Chapter 1 a simple new graphical procedure which provides confidence bands for a normal quantile-quantile plot. These bands define a test of normality which is both a powerful and visually insightful tool compared to the common used testing procedures. In Chapter 3 we develop a Bayesian nonparametric model for multivariate Poisson time series which have small counts. We demonstrate this model using simulations and by forecasting violent crimes pattern across Washington D.C.^
"Forecasting Evolving Curves"
(January 1, 2012).
Dissertations available from ProQuest.