IUP Publications Online
Home About IUP Magazines Journals Books Archives
     
Recommend    |    Subscriber Services    |    Feedback    |     Subscribe Online
 
The IUP Journal of Computer Sciences :
Improving the Volatility Forecasts of GARCH Family Models with the Recurrent Neural Networks
:
:
:
:
:
:
:
:
:
 
 
 
 
 
 
 

The primary objective of this paper is to develop a family of GARCH models, combining them with popular Recurrent Neural Network (RNN) models, which can capture the high nonlinear relationships between past return innovations and conditional variance, which is overlooked by standard GARCH models. The next objective is to apply Markov Switching GARCH model and see the differences in the predictive accuracy (according to AIC and BIG criteria and in two viewpoints: MSB and MAD) between the standard GARCH models, GJRGARCH model, RNN-GARCH models and Markov Switching GARCH model by comparing their out-of-sample forecasts. The dataset consists of a series of daily returns obtained from the National Stock Exchange (NSE) for the Indian Equity Markets. The results indicated that the proposed RNN-GARCH model and RNN-Markov Switching models are accurate and quick prediction methods.

 
 
 

Volatility of financial returns is an important aspect of many financial decisions. For example, volatility of exchange rates is a determinant for pricing currency options used for risk management. Hence, there is a need for good volatility forecasts.
The three most significant characteristics of returns in financial assets could be stated as follows: volatility clustering property, as a result of the volatility changes over time in magnitude and in cases where prices hardly change but volatility increases in sizes of large clusters; asymmetric relation property of volatility to past return shocks (Nelson, 1991; Nelson and Cao, 1992; Robert and Victor, 1993; and Lawrence et al., 1993); and nonlinearity property—the path of volatility reacts differently in different regimes (Franc, 2002; and Porter and Kramer, 2006).
To capture this, many authors used Autoregressive Conditional Heteroskedasticity (ARCH) models, as introduced by Robert (1982) and extended to Generalized ARCH (GARCH) by Tim (1986). Such models usually improve the fit a lot compared with a constant variance model, and Torben and Tim (1998) proved that GARCH models provide good volatility forecasts.

 
 
 

Computer Sciences IUP Journal , Volatility, GARCH standard models, GJR-GARCH model, Markov Switching GARCH, Artificial Neural Network (ANN), RNN-GARCH, Conditional variance.