PANK-A financial time series prediction model integrating principal component analysis, affinity propagation clustering and nested k-nearest neighbor regression

This paper proposes an integrated computational intelligence model called PANK for financial time series prediction. A PANK model consists of three parts: 1) Principal Component Analysis (PCA) for reducing redundancy information, 2) Affinity Propagation Clustering (AP) for generating exemplars and corresponding clusters as feature extraction, and 3) a nested reformulation of k-Nearest Neighbor regression (Nested KNN) for prediction modeling. The model captures training and testing data with a sliding window, uses PCA to reduce the redundancy information of historical data set and generates information-rich principal components which are input to AP for clustering, and applies Nested KNN to transform the clusters into output as prediction. In this paper, we advance the original KNN to a new Nested KNN which can tackle the large amount of computation and disequilibrium samples problem of original KNN. A specific PANK model is constructed and tested on Chinese stock index with 15-year historical data set, achieving best hit rate of 0.80.