A General Class of Neural Network-GARCH Models for Financial Time Series Analysis

This paper introduces a general class of combined neural network-GARCH models suitable to financial time series analysis. We put special emphasis on designing a full model-building cycle for this class of models that includes all stages of econometric modelling (specification, estimation and evaluation). Based on the maximum likelihood theory, we device procedures for statistical inference in the framework of NN-GARCH models and thus offer the modeler the opportunity to test hypotheses of interest concerning both the mean and variance structure of the data-generating process. We examine various issues in the model-building cycle by means of a simulation study and an empirical application to the DAX Stock Index