Testing for Dependence in Non-gaussian Time Series Data

Abstract

This paper provides a general methodology for testing for dependence in time series data, with particular emphasis given to non-Gaussian data. A dynamic model is postulated for a continuous latent variable and the dynamic structure transferred to the non-Gaussian, possibly discrete, observations. Locally most powerful tests for various forms of dependence are derived, based on an approximate likelihood function. Invariance to the distribution adopted for the data, conditional on the latent process, is shown to hold in certain cases. The tests are applied to various financial data sets, and Monte Carlo experiments used to gauge their finite sample properties.