A0763
Title: Forecasting realized volatility with echo state networks
Authors: Michael Grebe - The University of Manchester (United Kingdom) [presenting]
Abstract: The ability of artificial neural Echo State Networks (ESN) to forecast daily realised volatilities is examined using high-frequency data. I apply the state-of-the-art Echo State Network to forecast realized volatility of the S&P500 and compare the results with the HAR model and HAR-Q modification. I use daily realized volatility constructed using intraday returns, sampled at 5-minute intervals from 12 November 2009 to 30 August 2013 and test different network configurations to achieve best forecast results. The trainable parameters of the network are estimated such that they minimize the quasi-likelihood (QLIKE) loss function. The results show that Echo State Networks can be a powerful forecasting tool for financial time series, if implemented with care. The performance strongly depends on the pre-specified hyperparameter and ESNs under right configurations can outperform alternative econometric models such as the HAR and HARQ. Overall, the analysis has proven ESNs to be a promising alternative forecasting tool to the HAR model with improvement potential through further meta-parameter optimization.
