Learning Sparse Representation by K-SVD for Stock Classification

Selecting good stocks for investment is an essential problem in finance. However, accessing stocks is very difficult because many factors may affect the stocks and their relationship is also very complicated to analyze. Applying probabilistic statistical classification model, such as logistic regression, for stock analysis is promising. Logistic regression is an efficient classifier for predicting the outcome of a categorical dependent variable (class label) based on some predictor variables (features). Also, there has been a growing interest in the study of sparse representation of data. Using an over-complete dictionary that contains prototype bases, data are described by sparse linear combinations of these bases. Such a representation has biological interpretation and is very reasonable. Motivated by these reasons, in this paper, we propose a framework that combines sparse representation and logistic regression for stock classification. We use K-SVD algorithm to unsupervised learn sparse representation for the training data. The classifier is trained on these new representations of the data. We test our method in a real-world stock classification task, and our empirical results demonstrate the advantage of our method by comparing with other state of art representation learning methods.