FORECASTING BIST100 AND NASDAQ INDICES WITH SINGLE AND HYBRID MACHINE LEARNING ALGORITHMS

Abstract

The aim of this paper is to investigate stock market return forecasting performance of single and the developed novel hybrid machine learning (ML) algorithms. Daily returns of BIST100 and NASDAQ indices are predicted by series specific GARCH and ARMA-GARCH as well as three different ML algorithms that are Random Forest, XGBoost and Artificial Neural Networks (ANN). New hybrid ML models incorporating forecasts of the traditional (ARMA-)GARCH and the three ML algorithms are developed. Accuracy of the out-of-sample predictions of the methods are reported both for the single and hybrid models including pre-COVID-19, post-COVID-19 and the full sample test periods. Moreover, a simple trading strategy is applied in order to assess the economic impact of employing a specific forecasting model. According to the obtained accuracy metrics and the results of the trading strategy, developed novel hybrid models suggest quite promising results compared to the forecasts of the other models, especially (ARMA-)GARCH.