TAR-cointegration neural network model: An empirical analysis of exchange rates and stock returns

Abstract

The study aims to propose a family of Neural Networks (NN) model to achieve improvement in modeling nonlinear cointegration compared to Hansen and Seo (2002) Threshold Autoregressive Vector Error Correction (TAR-VEC) model. Our proposed TAR-VEC-NN family consist of TAR-VEC Multi Layer Perceptron (TAR-VEC-MLP), TAR-VEC Radial Basis Function (TAR-VEC-RBF) and TAR-VEC Recurrent Hybrid Elman (TAR-VEC-RHE) models. TAR-VEC-NN models are also discussed under two modeling strategies, first based on TAR-VEC modeling and the second based on a NN modeling approaches. The TAR-VEC-NN models proposed are analyzed for modeling monthly returns of TL/$ real exchange rate and ISE100 Istanbul Stock Exchange Index. For the data analyzed in the study, the TAR-VEC-NN models and their nonlinear cointegration structure improve forecast accuracy compared to TAR-VEC models; for both modeling strategies, we obtained similar results. Even though TAR-VEC-MLP model provides comparatively significant forecast improvement, TAR-VEC-RHE and TAR-VEC-RBF models achieve better forecast accuracy as expected given the dynamic memory structure of RHE and given the basis functions of RBF models which capture nonlinear error correction more efficiently. Further, our results show that, though with in sample accuracy, TAR-VEC-MLP and TAR-VEC-RHE produced the low RMSE values, in terms of long run predictions, the RBF model produced best results which is expected given the basis functions' capability in capturing deviations with the gaussian functions in a nonlinear error correction system. Thus, in the literature the forecasting ability of VEC type models are commonly criticized. With the use of our approach, there is an important improvement in VEC based models with NN specifications in terms of forecasts which cannot be disregarded. (C) 2009 Elsevier Ltd. All rights reserved.