An Ensemble Classifier for Stock Trend Prediction Using Sentence-Level Chinese News Sentiment and Technical Indicators

Abstract

In the financial market, predicting stock trends based on stock market news is a challenging task, and researchers are devoted to developing forecasting models. From the existing literature, the performance of the forecasting model is better when news sentiment and technical analysis are considered than when only one of them is used. However, analyzing news sentiment for trend forecasting is a difficult task, especially for Chinese news, because it is unstructured data and extracting the most important features is difficult. Moreover, positive or negative news does not always affect stock prices in a certain way. Therefore, in this paper, we propose an approach to build an ensemble classifier using sentiment in Chinese news at sentence level and technical indicators to predict stock trends. In the training stages, we first divide each news item into a set of sentences. TextRank and word2vec are then used to generate a predefined number of key sentences. The sentiment scores of these key sentences are computed using the given financial lexicon. The sentiment values of the key phrases, the three values of the technical indicators and the stock trend label are merged as a training instance. Based on the sentiment values of the key sets, the corpora are divided into positive and negative news datasets. The two datasets formed are then used to build positive and negative stock trend prediction models using the support vector machine. To increase the reliability of the prediction model, a third classifier is created using the Bollinger Bands. These three classifiers are combined to form an ensemble classifier. In the testing phase, a voting mechanism is used with the trained ensemble classifier to make the final decision based on the trading signals generated by the three classifiers. Finally, experiments were conducted on five years of news and stock prices of one company to show the effectiveness of the proposed approach, and results show that the accuracy and P / L ratio of the proposed approach are 61% and 4.0821 are better than the existing approach.