Enhancement of Non-stationary Time-Series Clustering Analysis Using Projection Pursuit Regression Method

The theory and application of time-series clustering analysis is an effective explanatory technique in various research fields. To overcome the limitations and many assumptions in conventional model-based clustering, this study utilizes the projection pursuit regression method as explanatory tool for formulating, identifying and estimating nonlinear models to approach the complex regression surface and then applied the projection pursuit method and an agglomerative scheme to cluster time series based on a similarity measure. This clustering can be applied to nonlinear, non-stationary, non-Gaussian models, and models involving interactions in predictor variables. Simulation results and real data analysis for categorizing the collection of average personal income of 25 states in the US demonstrate that this scheme compares favorably with other methods for similar clustering tasks.