A stock index prediction framework: Integrating technical and topological mesoscale indicators

With its growing importance in predicting future stock trends, nearly everyone watches the Chinese financial market. Traditional approaches typically employ a variety of statistical techniques or machine learning methods for stock index predicting, and often rely on analysis of technical indicators. In the existing literature, researchers rarely attempt to predict the stock index by using the topological features of temporal stock correlation networks. Keeping this in mind, we first calculate the correlation coefficient of any two stocks using the classic Visibility Graph Model (VGM). Then, by using the Planar Maximally Filtered Graph (PMFG) method, we generate temporal stock correlation networks from historical stock quantitative data. Next, we choose fourteen frequently adopted Technical Indicators (TIs) and five Topological Mesoscale Indicators (TMIs, extracted from the temporal stock correlation networks) as predictive variables of six machine learning classifiers. To improve forecast accuracy and to address potential overfitting problems, we modify the classic Sequential Backward Selection (SBS) algorithm to learn the most significant predictive variables for each classifier. We then conduct a series of comprehensive experiments on three Chinese stock indices to validate our prediction framework's performance. Experimental results show that using a combination of TIs and TMIs significantly improves forecast accuracy over conventional methods that use either TIs or TMIs exclusively.