A Comparison between Neural Networks and GARCH Models in Exchange Rate Forecasting

Abstract

Modeling and forecasting of dynamics nominal exchange rate has long been a focus of financial and economic research. Artificial Intelligence (IA) modeling has recently attracted much attention as a new technique in economic and financial forecasting. This paper proposes an alternative approach based on artificial neural network (ANN) to predict the daily exchange rates. Our empirical study is based on a series of daily data in Tunisia. In order to evaluate this approach, we compare it with a generalized autoregressive conditional heteroskedasticity (GARCH) model in terms of their performance. Results indicate that the proposed nonlinear autoregressive (NAR) model is an accurate and a quick prediction method. This finding helps businesses and policymakers to plan more appropriately. 1. Introduction During the latest years economists have shown a great interest in exchange rate forecasting. Therefore, a new notion appeared lately in the scientific literature that is the artificial neural network. It aims to improve prediction that helps achieve accurate results and go beyond traditional linear approaches. Neural networks started to be a forecasting tool that appeals to time series thanks to its modelling of noisy and incomplete time series. Dhamija and Bhalla (2010) have compared the predicting performance of the neural network model to other heteroscedastic models namely ARCH, GARCH, GARCH-M, EGARCH et IGACH for the exchange rate series such as BP/USD, DEM/USD, JPY/USD, et EUR/USD. Results show a great forecasting performance of the neural network model at the expense of heteroscedastic models. Indeed, it is commonly stated that the neural network model remains the most performant compared to heteroscedastic models. Deniz and Akkoc (2013) have compared GARCH model to neural network model in terms of forecasting stock index volatility ISE30 of Turkey. They finally agreed that neural network is highly superior compared to other traditional models such as Arch and GARCH and this was clearly seen in different fields of finance namely investment decision, stock prices and risk management. The nonlinear model, specifically ARIMA model, was compared to neuronal technique. The latter was claimed to be the best technique in time series forecasting (Zhang, 2001). Bildirici and Ersin (2012) studied nonlinear models, support vector regression and GARCH model in which they combined GARCH model and neural network to finally get MLP-GARCH and SVR-GARCH model in an attempt to improve the forecasting performance of GARCH model. In their study which was applied on the daily yields of the stock index of Istanbul ISE100. They compared the different techniques using the error values MSE and RMSE to prove that artificial models are more robust than classical econometric models (GARCH). This study results are based on a comparison between linear models MA and ARIMA and neural network model (Mitrea et al., 2009) in an attempt to achieve a modeling capital structure (Hsiao, 2008).