State-Space Methodology


Thesis Abstract:

We make use of the state-space model in three separate statistical problems: time series discrimination, signal comparison testing, and model selection. We propose and investigate two discriminant procedures for classifying a state-space process of unknown origin into one of several known populations. Both approaches employ a nearest neighbor classification rule (Cover and Hart, 1967) and utilize discriminant functions based on Kullback (1968) information measures.  Our first procedure classifies state-space processes based on the mechanism generating the states, and our second procedure classifies state-space processes based on the latent trajectory of the states.  The first discriminant function assesses the discrepancy between the Kalman smoothed values (Kalman, 1960; Jazwinski, 1970) produced under two fitted models: the model associated with the unclassified series and a model associated with one of the classified series.  The second discriminant function compares the smoothed version of the unclassified series to that of a classified series.  We provide interpretable representations for the discriminants, and develop formulae for easy computation of the functions in the context of the EM algorithm. Both methods are evaluated in extensive simulations and the second procedure is illustrated with an application to the stock market. We derive a procedure for testing if the unobserved states of two state-space processes follow the same latent trajectory.  Our test statistic, which considers the point-wise progression of two state-space series, has an approximate chi-square distribution under the null hypothesis and the assumption of Gaussian measurement noise.  The test is shown to perform well in simulations, and is illustrated with data from climatology.  We develop a model selection criterion useful in the state-space framework.  Our criterion is based on estimating Kullback's directed divergence between the data-generating density and an estimated model density.  To obtain an approximately unbiased estimate of the directed divergence, we employ Monte Carlo methods proposed by Hurvich, Shumway & Tsai (1990).  Our criterion is shown to clearly outperform the currently available model selection criteria in simulations.

Keywords:  Kullback-Leibler information, nearest-neighbor discrimination, time series analysis, time series discrimination, state-space model, signal extraction, model


Publications: