Koşullu Volatilitenin Modellenmesinde Destek Vektör Makinesi GARCH Modeli Ve Türk Finans Piyasaları Üzerine Bir Uygulama (Support Vector Machine GARCH Model in Modelling Conditional Volatility and an Application to Turkish Financial Markets)

Abstract

The English version of this paper can be found at: http://ssrn.com/abstract=2227747Calisma, temel GARCH modelinin Destek Vektor Makinesi ve Yapay Sinir Aglari ile iyilestirilmis modellerin incelenerek GARCH modelinin tahmin performansinin iyilestirilmesi bakimindan basarisinin test edilmesini amaclamaktadir. YSA yapisinin temel alindigi NN-GARCH modelinde Donaldson ve Kamstra (1997) NN-GARCH mimarisinin temel alindigi ve farkli ogrenme algoritmalari ile modellendigi Bildirici ve Ersin (2009) temel alinirken, SVR-GARCH modelinin olusturulmasinda Ou ve Wang (2010) ve Perez-Cruz (2003) calismalarindan hareket edilmistir. Modeller, tahmin basarisi bakimindan degerlendirilmistir. Bu kapsamda, IMKB-100 endeksinde gunluk getiriler modellenmesinde temel alinan GARCH, SVR-GARCH ve MLP-GARCH modelleri, orneklem ici ve orneklem disi tahmin basarisi acisindan degerlendirilmistir. Ampirik bulgular cercevesinde, seride one cikan leptokurtik dagilima ek olarak asimetri ve dogrusal olmama karakteristiklerinin SVR-GARCH ve MLP-GARCH modelleri ile modellenebilecegi; farkli hata kriterleri ve esit tahmin tutarliligi testleri kapsaminda temel GARCH modellerinden basarili bulundugu; orneklem disi tahmin basarisi bakimindan ise, Diebold Mariano testlerinden hareketle SVR-GARCH ve MLP-GARCH modelleri icin esit tahmin tutarliliginin 8 farkli gelecek tahmininden 6'si icin reddedilememekte iken; 2'si icin ise SVR-GARCH modelinin daha basarili bulundugu sonuclarina varilmaktadir. The study aims to augment commonly applied volatility models with support vector machines and neural networks. The proposed modeling strategy benefits from neural network based GARCH models of Donaldson and Kamstra (1997) and SVR-GARCH models discussed by Ou and Wang (2010) and Perez-Cruz (2003). Models are evaluated for in-sample and out-of-sample forecasting of daily returns in Istanbul ISE100 stock index. Results suggest that volatility clustering, asymmetry and nonlinearity characteristics are modeled more effectively with SVR-GARCH and MLP-GARCH models compared to the basic GARCH models. In terms of out-of-sample forecasting capabilities, Diebold-Mariano tests show that equal forecast accuracy could not be rejected for most of the horizons for the MLP-GARCH and SVR-GARCH models. On the other hand, SVR-GARCH is found to possess better forecast accuracy in 2 out of 8 forecast competitions.