Feature Selection for the Prediction Model of the Tehran Stock Exchange Index by Dimensionality Reduction Techniques

The main purpose of this study is to select an appropriate model for daily prediction of the total index of the Tehran Stock Exchange (TEDPIX). In this regard, dimension reduction techniques have been used to select effective and representative features to increase the accuracy of the selected model.

Since dimensionality reduction can be performed by two different methods (feature selection and extraction), in this study, two methods were used simultaneously to select the appropriate features of the prediction model. Hence, the MID algorithm was used to select the features, and the PCA algorithm was used to extract them. In this regard, after collecting 34 financial and economic features affecting the stock market, the features were prioritized by the MID algorithm. Then, the appropriate model was selected by comparing the performance of two different neural network models called RBF and DNN, which are respectively the most important and innovative of the extant models. Then, using two types of dimensionality reduction techniques, the prediction accuracy of the selected model was examined. The appropriate method for selecting the input features of the prediction model was identified, accordingly.

Analysis of the obtained results showed that the RBF model comes with more accuracy in the daily prediction of the Tehran Exchange Dividend and Price Index. Also, by comparing the performance of the two types of dimensionality reduction techniques, it was found that compared with the PCA algorithm, the MID algorithm brings better results in selecting the input variables of the RBF model. Therefore, according to the priority of features with the MID algorithm and the pattern of changing the level of error by increasing the number of features in the RBF model, the ISF-MID algorithm was proposed to select the appropriate features of the stock index prediction model. Using this algorithm, with the minimum number of features, can end in the highest accuracy in predicting the total index of the Tehran Stock Exchange.

The proposed method can identify, prioritize and select appropriate features for the prediction model, due to the simplicity and effectiveness of its use. It can also be useful in various areas of modeling, including the capital market, foreign exchange market, etc.