Hybrid Approaches in Financial Time Series Forecasting: A Stock Market Application

Abstract

The hybrid approach in time series forecasting is one of the key methodologies in selecting the most accurate model when compared to the single models. Applications of machine learning algorithms in hybrid modeling for stock market forecasting have been developing rapidly. In this paper, we propose hybrid modeling through machine learning approach for four stock market data; two from the developed stock markets (NASDAQ and DAX) and the other two from the emerging stock markets (NSE and BIST). A stock market is known with its volatile structure and has an unstable nature, so we propose several combinations for the hybrid models considering volatility to reach the most accurate time series forecasting model. In hybrid modeling, first ARIMA (Autoregressive Integrated Moving Average) models combined with GARCH models (Generalized Autoregressive Conditional Heteroscedasticity) are used for modeling of time series, then intelligent models such as SVM (support vector machine) and LSTM (Long-Short term memory) are used for nonlinear modeling of error series. We also compare their performances with single models. The proposed hybrid methodology markedly improves the prediction performances of time series models by combining several models which reflect the time series data characteristics best.