This article deals with the cluster analysis of panels of short time series which are gathered by observing a fixed set of units at regular intervals of time. Given the reduced length of time series, we follow a non-model-based strategy in which the priority is to assign a value to the dissimilarity between the observed time series themselves rather than to the processes that generate those time series. Our main contribution is a new method for clustering short panel data that can be used to explore whether the various cross-sectional time series move in a similar way and whether there is substantial variation in each panel over time. It is also important to identify outliers and other anomalies because researchers may be especially interested in studying panels that behave unusually, at least with respect to the rest of the data.Although finding group structures within short panels remains challenging, the results acquired for real and simulated data are valid and encouraging for further study.
Dynamic Time Warping, K-medoid, Multiple Distance Matrices, Sequence Analysis.