A clustering method for online shopping search data based on Continuous Wavelet Transformation (CWT) and its inverse transformation was proposed for variable selection in predictive model. The method decomposed original series into different periodic components by taking full account of special characteristics of search data and reconstructed such periodic components into input vectors. Clustering was implemented through weighted Fuzzy C-Means (FCM) algorithm. The variables (keywords) were selected according to their membership function values in each group. Variable selection effectiveness was then evaluated through a prediction model for Chinese monthly Consumer Price Index (CPI). The experimental results indicate that search volume series have different periodic components and the keywords within the same group are highly consistent in commodity type. Compared to other variable selection methods, the prediction model based on the wavelet clustering can achieve better prediction accuracy, the one-step and three-step relative prediction errors are 0.3891% and 0.5437% respectively, and the selected variables also have clearly economic meaning. The proposed method is particularly suitable to address variable selection problem of high-dimensional predictive model in the big data era.

In order to take the multi-fractal properties of financial time series into consideration, a clustering method based on measuring similarities among CSI 300 index stocks through scaling curve was proposed. The algorithm firstly fitted scaling curve at different autocorrelation order through Multi-Scale Detrend Fluctuation Analysis (MSDFA). Then it abstracted the distribution or shape features of scaling curve for the construction of pattern vector. Clustering was implemented by weighted K-means algorithm and the optimal number of categories was determined by Davis-Bouldin Index (DBI). The result shows the clustering based on scaling curve can discover industry aggregation and strong linkage of different plates within the stock market. The portfolio built from different clusters can reduce the risk greatly and the proposed method outperforms clustering method based on linear trend features of original series.