Proximal support vector machine based hybrid prediction models for trend forecasting in financial markets

•Proximal support vector machine (PSVM) have been used for forecasting of daily stock trends.•Input vectors for PSVM are selected using different feature selection techniques.•A new performance measure called joint prediction error (JPE) is proposed.•On the basis of JPE, it is observed that the hybrid approaches perform better than basic classifier.

In the recent years, various financial forecasting systems have been developed using machine learning techniques. Deciding the relevant input variables for these systems is a crucial factor and their performances depend a lot on the choice of input variables. In this work, a set of fifty-five technical indicators has been considered based on their application in technical analysis as input feature to predict the future (one-day-ahead) direction of stock indices. This study proposes four hybrid prediction models that are combinations of four different feature selection techniques (Linear Correlation (LC), Rank Correlation (RC), Regression Relief (RR) and Random Forest (RF)), with proximal support vector machine (PSVM) classifier. The performance of these models has been evaluated for twelve different stock indices, on the basis of several performance metrics used in literature. A new performance measuring criteria, called joint prediction error (JPE) is also proposed for comparing the results. The empirical results obtained over a set of stock market indices from different international markets show that all hybrid models perform better than the individual PSVM prediction model. The comparison between the proposed models demonstrates superiority of RF-PSVM over all other prediction models. Empirical findings also suggest the superiority of a certain set of indicators over other indicators in achieving better results.