An Application of Neural Networks Trained with Kalman Filter Variants (EKF and UKF) to Heteroscedastic Time Series Forecasting
Mauri Aparecido de Oliveira, Escola Paulista de Política · 2012
In this work, two Kalman filters variants are applied to recurrent neural network training. The Unscented Kalman Filter (UKF) has been presented outperforming the Extended Kalman filter (EKF). Due to this a comparison between GARCH model and a neural network using EKF and UKF was implemented to heteroscedasticity time series prediction. Our experimental results and analysis confirm that a neural network using UKF perform better prediction than the other approach.