In cross country skiing there are different skiing techniques, known as gears. For professional skiers it is useful to analyze a ski session with respect to the gears used in different parts of the track. We have developed a statistical machine learning method that uses data from accelerometers in e.g. a mobile phone placed on the chest of a skier to classify the gears used. The statistical model used is based on a Markov chain of multivariate Gaussian distributions. The same model can in addition to classification be used for anomaly detection and unsupervised clustering of skiing movements. The method is evaluated on real data from elite skiers collected during a training race.