An Econometric Approach to Neural Network Model Selection for Financial Time Series Analysis


Speaker


Abstract

The purpose of this talk is to present a family of neural network (NN)-GARCH models that jointly parametrise the mean and the variance of the conditional distribution. We show how this flexible modelling framework can accommodate most of the stylised facts reported about financial prices or rates of return (nonlinear trends, asymmetric GARCH effects and non-gaussian errors). Following standard econometric practice, we analytically discuss several strategies for the specification of the mean and variance components of the model by means of sequential statistical hypotheses tests. We propose variations of the standard testing framework that are robust to model misspecification, i.e. they preserve their asymptotic validity even when the model is not correctly specified for the entire conditional distribution. Based on simulation evidence, we shed light on the finite-sample performance of the aforementioned testing procedures under various data-generating processes compatible with the outcomes of financial studies on asset returns. Finally, we discuss a range of possible - and often divergent- applications for the family of NN-GARCH models, such as forecasting the conditional distribution of equity returns, establishing an empirical pricing model for S&P 500 index options and detecting statistical arbitrage opportunities among fundamentally related assets.