This paper proposes an extension of the classification trees to time series input variables. A new split criterion based on time series proximities is introduced. First, it relies on an adaptive (i.e., parametrized) time series metric to cover both behavior and values proximities. The metric's parameters may change from one internal node to another to best bisect the set of time series. Second, it involves the automatic extraction of the most discriminating sub-sequences. The proposed time series classification tree is applied to a wide range of datasets: public and new, real and synthetic, univariate and multivariate data. We show, through the carried out experiments, that the proposed tree outperforms temporal trees using standard time series distances, and leads to good performances compared to other competitive time series classifiers.