George Kapetanios ,
Queen Mary, University of London
October 1, 2007
Testing serial dependence is central to much of time series econometrics. A number of tests that have been developed and used to explore the dependence properties of various processes. This paper builds on recent work on nonparametric tests of independence. We consider a fact that characterises serially dependent processes using a generalisation of the autocorrelation function. Using this fact we build dependence tests that make use of neural network based approximations. We derive the theoretical properties of our tests and show that they have superior power properties. Our Monte Carlo evaluation supports the theoretical findings. An application to a large dataset of stock returns illustrates the usefulness of the proposed tests.
J.E.L classification codes: C32, C33, G12
Keywords:Independence, Neural networks, Strict stationarity, Bootstrap, S&P500