We propose a new splitting approach to extend the decision trees to temporal data. The proposed split aims to determine for each daughter node the representative time series, the observation period best discriminating the output variable, and the optimal contribution of the values and of the behavior for the proximity evaluation. A new extension of the Dynamic time warping is also proposed. The high efficiency and interpretability of the proposition is illustrated through many public datasets and compared to two important alternative algorithms. Future work will focus on the stability evaluation through real datasets.