Stock Data Clustering and Multiscale Trend Detection

Generally, trend detection algorithms over the data stream require expert assistance in some form. We present an unsupervised multiscale data stream algorithm which detects trends for evolving time series based on a data driver data stream. The raw stream data clustering algorithm is incremental, space dilating and has linear time complexity. The evolving stream is incrementally explored on a number of windows. Whenever a change occurs, we switch the analysis on this driver data stream in order to detect the new aggregated patterns and the new best selection of window widths among an exponential base set. The window widths are detected using a slightly modified version of an incremental SVD procedure. We apply this clustering algorithm to a free public NYSE stock exchange financial data feed, investigating incrementally the developing trends during the current crisis data from 2007 to 2009. The algorithm detected the changing widths of the trends as well as the trends themselves in the market.