Forecasting stock market volatility using artificial neural networks

Abstract

Stock markets play a very significant role in the economy of capitalist countries. Millions of people trade everyday in order to have more and more profit. At this point, being able to know the future of the market gets more importance. Future prediction in a stock market is the prediction of the probability of the future losses. This probability forms the future risk profile for the interested market participant. The important thing is the estimation, measurement and the definition of the risk in mathematical norms. In this context, the variance in the time series is measured and the predictions are based over this variance. Financial time series exhibit time dependent heteroskedastic variance known as the conditional variance (volatility) that is not a directly observable feature. In this thesis study, we focus on the volatility modeling using artificial neural networks (ANNs) and future predictions with these volatility models. Specifically, we are investigating the use of ANNs in risk estimation of asset returns. On the contrary to traditional methods, we have used ANNs to model the relationship, the dependence in time in volatility. We have divided the space into clusters with the help of the Mixture of Experts (MoE)'s divide and conquer technique and we have assigned local experts to each cluster. Having localized experts learn their region of interest and having combined the outputs of these local experts via a gating expert we have been able to model the relationship in time and we have used this technique for the future prediction of Istanbul Stock Exchange (ISE) National-100 index. Also we modeled the relation between the input and output space by the help of the hybrid recurrent neural networks (RNN)'s multiple feedback mechanism. As a result, we have determined that MoE and hybrid RNN are very promising in modeling the volatility of ISE National-100 index.