Clustering (financial) time series according to their predictability by automatically chosen predictors

Siemens Austria

The project tries to find a small number of observed predictors (leading indicators) for a large number of time series. Therefore a measure for the degree of predictability for each series by the potential predictors is constructed. Both a time domain (based on the cross correlation function) and a frequency domain (based on the cross spectrum) version is given.


The result is a possibly rectangular (depending on the choice of the potential predictors) and essentially asymmetric predictability matrix. A heuristic clustering method is developed to cope with this type of problem. It is a generalization of the PAM algorithm of Kaufman and Rousseeuw (1990), a k-medoids method, for symmetric distance matrices. Alternatively, integer programming solutions could be used to find the clusters.


The approach is applied to a set of 298 daily financial return series for the period January 1998 to November 2000. It is possible to predict 236 (of 298) series reasonably well by 5 automatically chosen series.