Sequence Clustering for Regime Change Detection

点击翻译按钮获取中文摘要

This Engage grant application aims to build a research partnership between our research group andDimensional Research Canada. The research project proposes to investigate the use of advanced data miningtechniques for detecting regime changes from financial time series data. We will design, implement and test anew method in order to obtain a set of rich features characterizing financial regimes. The method will be basedon our recent Model-based Categorical Sequence Clustering (MCSC) algorithm. It contains the followingsteps: 1) representing financial time series via candlesticks and performing cluster analysis to convertcandlestick representation to categorical representation; 2) performing hierarchical sequence clustering usingthe MCSC algorithm; 3) exploring the variable-order Markov model of the MCSC to extract signature patterns;and 4) investigating the usefulness of these patterns to characterize regimes through a comparative study forpredicting market returns. The innovative nature of the proposed methodology lies in its unsupervised learningapproach to analyzing regimes and regime changes. It has the potential to overcome the major limitations of theconventional approaches. It will allow the discovery of long and statistically significant patterns, yielding morepredictive and easier-to-interpret features. The project is expected to provide direct economic benefits to ourindustry partner, by expanding the scope of its existing platform and its internal knowledge of how to developfuture products for complex time series analysis. It also has the potential to benefit the industry, public serviceand scientific research sectors in general.