Automated detection of time-dependent cross-correlation clusters in nonstationary time series

A novel method for the detection of cross-correlation clusters in multivariate time series is suggested. It is based on linear combinations of the eigenvectors corresponding to the largest eigenvalues of the equal-time cross-correlation matrix. The linear combinations are found in a systematic way by maximizing an appropriate distance measure. The performance of the algorithm is evaluated with a flexible time-series–based test framework for cluster algorithms. Attribution errors are investigated quantitatively in model data and a comparison with three alternative approaches is made. As the algorithm is suitable for unsupervised online application we demonstrate its time-resolved use in the example of cluster detection in time series from human electroencephalogram.

[1]  J. Bouchaud,et al.  Noise Dressing of Financial Correlation Matrices , 1998, cond-mat/9810255.

[2]  Anil K. Jain,et al.  Algorithms for Clustering Data , 1988 .

[3]  I. Jolliffe.  Principal Component Analysis , 2002 .

[4]  V. Plerou,et al.  Random matrix approach to cross correlations in financial data. , 2001, Physical review. E, Statistical, nonlinear, and soft matter physics.

[5]  V. Plerou,et al.  Universal and Nonuniversal Properties of Cross Correlations in Financial Time Series , 1999, cond-mat/9902283.

[6]  U. Stephani,et al.  Detection and characterization of changes of the correlation structure in multivariate time series. , 2005, Physical review. E, Statistical, nonlinear, and soft matter physics.