A scale space multiresolution method for extraction of time series features

A scale space multiresolution feature extraction method is proposed for time series data. The method detects intervals where time series features differ from their surroundings, and it produces a multiresolution analysis of the series as a sum of scale‐dependent components. These components are obtained from differences of smooths. The relevant sequence of smoothing levels is determined using derivatives of smooths with respect to the logarithm of the smoothing parameter. As time series are usually noisy, the method uses Bayesian inference to establish the credibility of the components. © The Authors. Stat published by John Wiley & Sons Ltd.