Tests for serial independence and linearity based on correlation integrals

We propose information theoretic tests for serial independence and linearity in time series. The test statistics are based on the conditional mutual information, a general measure of dependence between lagged variables. In case of rejecting the null hypothesis, this readily provides insights into the lags through which the dependence arises. The conditional mutual information is estimated using the correlation integral from chaos theory. The significance of the test statistic is determined with a permutation procedure and a parametric bootstrap in the tests for independence and linearity, respectively. The size and power properties of the tests are examined numerically and illustrated with applications to some benchmark time series.