Prediction of machine tool wear is highly dependent on the quality of the measured data and the ability to extract information from such raw data. These data are presented in the form of time series, which cannot be used directly by conventional machine learning algorithms, such as the one used in this work. To link the raw data and the learning algorithm, it is first necessary to extract a feature set from the time series. An important but little analyzed aspect is the size of the window required for feature extraction. If this window is too small, not much information will be obtained, on the other hand, if the window is too large, there will be more chance of outliers and other irregularities of the data being introduced. In the present work, we use a novel database corresponding to machine tool wear to demonstrate the impact of window size. An optimally chosen window size, plus an adequate feature extraction, allows us to obtain results comparable to the state of the art, i.e., median scores of 89 %, which are comparable to that obtained by the first place of the recently held data challenge.