To accurately detect anomalies and without prior knowledge in a time series, is it better to build the detectors from the initial temporal representation, or to compute a new (tabular) representation using an existing automatic variable construction library? In this article, we answer this question by conducting an in-depth experimental study for two popular detectors (Isolation Forest and Local Outlier Factor). The results obtained, for 5 different datasets, show that the new representation, calculated using the tsfresh library, allows Isolation Forest to significantly improve its performance.