Application of Unsupervised Feature Selection, Machine Learning and Evolutionary Algorithm in Predicting Stock Returns: A Study of Indian Firms

Abstract

Prediction of stock prices has become an important area of research in the field of financial analytics and has garnered a lot of attention among academicians. Drawing on the literature on application of econometric tools and also machine learning techniques, this paper presents a framework for predicting stock returns using three unsupervised feature selection techniques, four predictive modeling techniques and finally an ensemble combining the four predictive modeling techniques. To design the ensemble, evolutionary algorithm is applied. In order to assess the results of our study, four different performance measures, namely, Mean Absolute Error (MAE), Mean Squared Error (MSE), Nash-Sutcliffe Efficiency (NSE) and Index of Agreement (IA) have been utilized. Our feature selection results indicate that all explanatory variables are not significant for different classes of companies and also for different time periods. This gives us insight into the fact that, for stock returns prediction, one has to be careful of the predictors to be chosen. Further, results indicate that for all the forecasting methods, namely, random forest, bagging, boosting and support vector regression, forecasting efficiency for large cap and mid-cap firms was better than that of small cap firms. Statistical analysis through Analysis of Variance (ANOVA) suggests that of all four predictive modeling techniques, boosting was the most efficient technique for forecasting the stock returns. We then proceeded to construct an ensemble of the above four methods. In terms of all four measurement metrics, performance of the proposed ensemble was better in both training and testing phase as compared to the efficiency of the individual predictive modeling techniques.