Title: Forecasting of high-dimensional realized covariances with reservoir computing
Authors: Lyudmila Grigoryeva - University of Konstanz (Germany) [presenting]
Oleksandra Kukharenko - University of Konstanz (Germany)
Juan-Pablo Ortega - University St. Gallen (Switzerland)
Abstract: The problem of forecasting high-dimensional realized covariance (RV) matrices computed out of intraday returns of the components of the S\&P 500 market index is considered. The study focuses on a novel machine learning paradigm known as reservoir computing (RC) for producing multistep ahead forecasts for time series of realized covariances. Various families of reservoir computers have been recently proved to have universal approximation properties when processing stochastic discrete-time semi-infinite inputs. The goal is to implement with reservoir computers the forecasting of realized covariances. We examine the empirical performance of RC in comparison with many conventional state-of-the-art econometric models for various RV estimators, periods, and dimensions. We show that universal RC families consistently demonstrate superior predictive ability for various designs of empirical exercises.