Parameter Sensitivity of Support Vector Regression and Neural Networks for Forecasting

I. INTRODUCTION ccurate corporate decisions in an uncertain future environment requires accurate forecasting [1, 2]. As a consequence, significant effort has been invested in developing forecasting methods with enhanced forecasting accuracy, extending established statistical approaches of Exponential Smoothing and ARIMA-methods towards nonlinear methods of ARCH, GARCH, STAR etc. and methods of computational intelligence [3]. While statistical methods are embedded in a methodology following iterative model building and parameterization based upon a prior theory, methods from computational intelligence such as support vector regression (SVR) and neural networks (NN) are semi-parametric, data-driven methods which capture the underlying linear and/or non-linear model form and suitable parameters directly from the data [4]. Thus they offer promising features to applications to various business forecasting domains where limited tangible knowledge on an underlying model form exists, such as accounting, finance, marketing, economics, production, tourism, transportation etc. [2, 5]. However, in contrast to statistical methods, NN as well as SVR have not been established in business practice despite their attractive theoretical properties. In particular, the limited empirical studies on time series forecasting with NN