Introduction. When dealing with time series, the application of a smoothing lter (to get rid of random uctuations and better recognise the relevant structure) is usually one of the rst steps. In the literature on time series similarity measures, however, the impact of smoothing is not explicitly or systematically considered { despite extensive experiments in, e.g., [2]. Instead, complex similarity measures are frequently applied (e.g. dynamic time warping (DTW)), which implicitly deal with noise, but mainly with temporal dilation and translation e ects. So up to now it is unclear, to what extent the good performance of DTW is due to its smoothing or warping capabilities. Optimal Filter. In this work we consider a simple Euclidean distance applied to preprocessed (smoothed) time series. It is unlikely that one similarity measure ts all problem types (or data sets), so by choosing an appropriate lter, we may adopt to the problem at hand. The lter is automatically determined given a training set of classi ed series, such that distances between series of the same (di erent) class are minimised (maximised). The obtained similarity measure is then tested in cross-validated 1-NN classi cation for various data sets (as in [2]) and compared against the DTW performance. Starting from Euclidean distance (without any preprocessing) as a baseline, it turns out that for many data sets a substantial fraction of the performance improvement obtained with DTW is also obtained by choosing the appropriate lter. In some cases, the performance is even better than with DTW, which is due to the fact that a lter is a versatile tool: for some problems it may be advantageous to distinguish time series by their derivative rather than the original series and in such cases a lter that estimates the derivative may be retrieved. For further details the reader is referred to [1].