Efficient Statistical Analysis of Financial Time-Series Using Neural Networks and GARCH Models

Abstract

We show how neural networks can be effectively combined with NN-GARCH models into a flexible modelling framework that can accommodate most of the statistical features observed in financial time-series (nonlinearities in mean, asymmetric GARCH effects and nongaussian errors). We analytically discuss several strategies for the specification of the NN component of the model and propose variations of the standard testing framework that are robust to heteroskedasticity. We demonstrate various aspects of the model-building strategy by employing NN-GARCH to forecast the conditional distribution of daily returns on the German DAX Stock index.

Keywords: Neural Networks, GARCH models, Maximum Likelihood, Conditional Densities