Hierarchical Clustering of Time Series using Gaussian Mixture Models and Variational Autoencoders

Abstract: This thesis proposes a hierarchical clustering algorithm for time series, comprised of a variational autoencoder to compress the series and a Gaussian mixture model to merge them into an appropriate cluster hierarchy. This approach is motivated by the autoencoders good results in dimensionality reduction tasks and by the likelihood framework given by the Gaussian mixture model. In contrast to similar clustering algorithms, this algorithm tries to answer the question of the true number of clusters in the data set and gives superior visualisation possibilities. Furthermore, the thesis shows how cluster analysis in general can be applied to several interesting problems in finance, and specifically how this algorithm is engineered to outperform standard clustering algorithms on these problems.