Title: Predicting financial risk with artificial neural networks: Whether there is information in high frequency returns
Authors: Christian Muecher - University of Konstanz (Germany) [presenting]
Abstract: High Frequency financial data is vastly used for modelling financial risk, often by utilizing Realized Variance estimators. This aim is to model financial risk directly using high frequency returns as inputs for an artificial neural network. Artificial neural networks are universal approximators and thus are able to learn the function that predicts the conditional variance. There exist approaches showing that the function learned by a neural network using past daily returns as an input and the next periods squared daily return as an output is a consistent estimator of the conditional variance function. We extend these approaches by directly using high frequency returns as input to learn the variance function. The results using this more extensive information are compared to the approach based on past daily returns, as well as existing, standard, approaches used for predicting the variance of financial assets. The comparison is done in terms of simulated price processes and an application to real data.