Stock Price Forecasting: A Comparative Study of Single and Hybrid Feature Selection Methods

In recent years, due to the nonstationary behavior of data samples, modeling and forecasting stock price have been difficult for the business community and researchers. Features for stock market forecasting has been selected using several state-of-art feature selection techniques. To improve the feature selection, in this work a novel hybrid feature selection technique that combines three feature selections technique, extreme gradient boosting, least absolute shrinkage and selection operator, and backward elimination to forecast the close prices of TVSMOTOR stock on India's National Stock Exchange. Specifically, these features are input to a support vector regression model to predict the closing price. Finally, the forecasting results based on the proposed model was compared with the existing feature selection methods. The study also evaluates the effectiveness of the proposed hybrid feature selection approach in its forecasting accuracy. The results show that the developed model performs better than the existing feature selection models as the mean absolute error ranges from 0.161388 to 1.35315, the mean absolute percentage error from 0.000251 to 0.002168, and the root mean squared error from 0.45483 to 1.9244.