We consider the problem of determining whether the community structure found by a clustering algorithm applied to nancial time series is statistically signi cant, or is due to pure chance, when no other information than the observed values and a similarity measure among time series are available. As a subsidiary problem we also analyse the in uence of the choice of similarity measure in the accuracy of the clustering method. We propose two raw-data based methods for assessing robustness of clustering algorithms on time-dependent data linked by a relation of similarity: One based on community scoring functions that quantify some topological property that characterises ground-truth communities, and another based on random perturbations and quanti cation of the variation in the community structure. These methodologies are well-established in the realm of unweighted networks; our contribution are versions of these methodologies properly adapted to complete weighted networks.

Peer Reviewed