Predicting high increases in stock prices using text mining and data resampling techniques

Abstract

Text mining techniques have been demonstrated their effectiveness in developing stock prediction models, in which most of them focus on predicting whether the future stock price will rise or fall as a binary classification problem. However, in practice, existing prediction models cannot fulfill a well-defined investment portfolio composed of high-, medium-, and low-risk target stocks for different levels of return on investment (ROI). In order to achieve this practical demand, the prediction models should be able to predict different stock rise ratios for the investment portfolio. To construct this kind of prediction models, the class imbalance problem occurs in the training datasets that the number of data examples in the high-rise class is much less than the ones in the nonhigh-rise class. Therefore, the aim of this paper is to examine the performances of text mining-based stock prediction models by different machine learning and deep learning techniques in predicting different high-stock-price ratios, including 3 %, 5 %, 7 %, and 9 %. In addition, different data resampling techniques are employed to rebalance the class imbalanced training datasets to construct the prediction models for performance comparison. The experimental results indicate that one-class classifiers, such as one-class support vector machine and isolation forest, perform very well over the class imbalanced datasets in terms of AUC rates and the type I error of misclassifying high-rise cases into the nonhigh-rise class. Furthermore, after rebalancing the training datasets using over- and hybrid sampling algorithms, most classifiers show certain performance improvement, where hybrid sampling is the better choice than oversampling.