Nonlinear Error-Correction Models in the Greek Money Market

Abstract

This paper explores the use of multilayer perceptron neural networks in modelling nonlinear error-correction mechanisms. Based on financial time series from the Greek Interbank Interest rate market, comparisons are drawn between neural network and linear error-correction models, regarding their out-of-sample forecasting ability. We establish that each of the series is I(1) and find a significant cointegrating relationship between them. The errors of the cointegrating regression are used to forecast one-day ahead logarithmic changes in a weekly interest rate, using a linear and a neural network error-correction model. We find that the nonlinear error-correction model has a superior out-of sample performance and is able to capture some of the nonlinearities in the series. We also show the specific nonlinear form of the error-correction relationship that was estimated via the neural network model.