Clustering of Time Series via Evidence Accumulation

When clustering time series, relevant information resides in their pairwise associations. Therefore, a measure of dissimilarity among time series is required. When the phenomenon of interest exhibits extreme dependence, it is possible to base such metrics on tail dependence coefficients. Identifying the best measure is far from trivial and largely impacts the resulting partition. In this contribution, we propose a novel approach, which accumulates evidence from a multiplicity of partitions obtained from several dissimilarities, potentially resulting in more robust final clusters. We finally illustrate the proposed approach by applying it to the analysis of financial time series where the emphasis is on possible lower tail dependence.