Stock price movement prediction using representative prototypes of financial reports

Stock price movement prediction is an appealing topic not only for research but also for commercial applications. Most of prior research separately analyzes the meanings of the qualitative or quantitative features, and does not consider the categorical information when clustering financial reports. Since quantitative or qualitative features contain only partial information, there may be no synergy by considering them individually. It is more appropriate to predict stock price movements by simultaneously taking both quantitative and qualitative features into account. Therefore, in this study, we utilize a weighting scheme to combine both qualitative and quantitative features of financial reports together, and propose a method to predict short-term stock price movements. The proposed method employs the categorical information to localize the clusters and improve the purity of each resultant cluster. We gathered 26,255 reports of companies listed in the S&P 500 index from the EDGAR database and conducted the GICS (Global Industrial Classification System) experiments based on the industry sectors. The empirical evaluation results show that the proposed method outperforms the SVM, naïve Bayes, and PFHC methods in terms of accuracy and average profit.