Speed up of the majority voting ensemble method for the prediction of stock price directions

The prediction of stock price directions is important in finance. The Majority Voting Ensemble method is superior in prediction accuracy to single classifier models including Logistic Regression, Decision Tree, K-Nearest Neighbors and Support Vector Machine, but the computational cost is very expensive since it considers all the hyperparameters of single classifier models. The current study proposes a revision of the majority voting method to improve the computational efficiency. The proposed method lets each single classifier model find its own hyperparameter values and this modification speeds up the computation by 500 times compared to the standard majority voting method while maintaining the accuracy. The numerical experiments show the ranking of the classifier models in the order of the proposed majority voting, the standard majority voting, and then other single classifier models including the support vector machine. This improvement will allow the majority voting ensemble method to be applied in the financial market in practice. The algorithms are tested on 7 national indices from 3 continents for the past 3 years, and the performance is measured in two criteria, the area under the receiver operating characteristic curve and the percent correctly classified.