A Strategy Integrating Iterative Filtering and Convolution Neural Network for Time Series Feature Extraction

Abstract

Time series processing is a vital issue that is encountered in various fields. However, such data are mostly non-stationary on account of the fact that they are affected by a variety of factors. In this paper, we present a supervised strategy by integrating the iterative filtering (IF) method and convolution neural network (CNN) to automatically extract features of time series, where the IF technique can decompose the raw time series into intrinsic mode functions (IMFs), and then the CNN aims to extract the features from the images constructed by the IMFs under specific task. To illustrate the performance of the proposed strategy, we apply it in one-step and multi-step predictive tasks on the national association of securities dealers automated quotations (NASDAQ) data. Furthermore, we compute the importance of the extracted and raw features by the combined decision trees, such as random forest (RF) and gradient boosted decision trees (GBDT). The results indicate the significant improvement of the proposed strategy.