PREDICTING STOCK MARKET DIRECTION USING MACHINE LEARNING MODELS

Abstract

Forecasting of stock prices has been a challenging area due to its complex and dynamic nature. There are several evidences that traditional econometrics based predictive models encountered significant challenges due to parameter instability. The aim of this study is to apply three classifiers namely, Random Forest (RF), Support Vector Machines (SVM) and Neural Networks (NN) to predict the Pakistani stock market’s direction and to compare the prediction accuracy. Daily closing prices are collected from yahoo server from 2013 to 2018. Famous 30 market indicators are applied to predict the market direction by using Random Forest, Support Vector Machines and Neural Networks. Model accuracy is evaluated using the confusion matrix. The empirical findings reveal that Neural Network performs best with the highest accuracy of 91%. Model specific, top five input indicators are used by applying feature selection in all classifiers. Interestingly, optimization improves the prediction accuracy in case of neural networks (NN) and support vector machine (SVM)) models while Random Forest’s (RF) accuracy did not improve. These findings have great importance for institutional investors and management companies having flexibility to accelerate or postpone their investment decisions.