Feature Selection Based Support Vector Machine For Stock Market Prediction

Stock market forecasting is a fascinating research topic for economists,mathematicians and computer scientists.A number of different approaches have been used for years,namely fundamental analysis,technical analysis and statistical analysis.However,with the advancement of technology,stock market prediction has been studied using soft computing methods.In which Support Vector Machines(SVMs)and Artificial Neural Networks(ANNs)have proven their strength.ANN works on empirical risk minimization while SVM works on structural risk minimization principle.Therefore,SVM is found to be better than ANN.However,SVM is prone to overfitting,when a large number of financial indicators are used.To reduce the number of features,feature selection techniques are used for selecting features,which are highly correlated with the target value.But most feature selection techniques do not consider the subsets of highly correlated features among a set of all features.This study deals with the overftting problem created due to a large number of features by considering the subsets of highly correlated features with the output among a set of all features.We introduced a new feature selection technique which uses 31 different technical indicators comprising 3 basic indicators,6 moving averages,20 momentum indicators and 2 volatility indicators.The proposed model named Feature selection based Support Vector Machine(FSVM)is divided into two parts,first is the feature selection technique and second is forecasting model based on SVM.Feature Selection technique uses Jenks optimization method to classify all 31 indicators into different classes based on the Pearson correlation coefficient.Then support vector regression(SVR)model is used to forecast the future price using the subsets of financial indicators generated by feature selection.NASDAQ stock index data was selected to train and test the model.We trained the model on 8 years data i.e.from January 2007 to December 2015 and next two years data from January 2016 to December 2017 was used to test the model.Model was compared with statistical model ARIMA and a simple support vector machine based model(SSVM)which is built without using feature selection technique.Results on the basis of performance measurements i.e.root mean squared error(RMSE)and coefficient of determination(R-squared)showed that our model outperforms SSVM and ARIMA.The obtained results provide a strong basis for implementing FSVM for real-time stock trading.Furthermore,results also show that our method can find better subsets of indicators.