A0845
Title: Self-normalising change-point detection procedures for high-dimensional data
Authors: Charl Pretorius - North-West University (South Africa) [presenting]
Heinrich Roodt - North-West University (South Africa)
Abstract: Nonparametric CUSUM-based test criteria are presented for detecting changes in the means of high-dimensional panel data. The test statistics are shown to be self-normalising in the sense that their null distributions are asymptotically pivotal, even in the presence of weak serial dependence. Hence, unlike many existing procedures, the new tests have the practical advantage of not requiring estimation of the long-run variance of the error component in each panel and therefore eliminate the reliance on choosing bandwidth parameters. This allows for the tabulation of general asymptotic critical values, which may readily be used in applications, including in situations where the true underlying data-generating process is unknown. The results are further generalised to the case where the panels are allowed to depend on unobserved common factors. Numerical results are presented, which show that the new tests are level-respecting under the null hypothesis for large enough samples and under weak time-dependence. An application to real data is discussed, and it is shown that the tests perform well when the innovations follow popular financial time series models such as ARMA and GARCH models.