A1451
Title: Nonparametric time-varying Granger causality using exponentially smoothed density estimators
Authors: Sicco Kooiker - Vrije Universiteit (Netherlands) [presenting]
Abstract: Where parametric methods require assumptions about the unknown type of Granger causality, and static methods tend to over-reject and lack power in dynamically changing time series environments, the nonparametric time-varying Granger causality (NPTVGC) testing framework proves to be a useful method. A procedure is presented that combines the popular Diks-Panchenko (DP) test with exponentially weighted moving average local density estimators to assess Granger causality in environments that change gradually over time. Originally, DP suggested setting the bandwidth using their rule-of-thumb. This method becomes invalid under smoothing. A cross-validation hyperparameter optimization algorithm is introduced, providing an alternative method for selecting the bandwidth of the DP statistic. Two simulation studies demonstrate the importance of correct hyperparameters and the use of exponential weighting. In the empirical experiment, the NPTVGC framework identifies Granger causality from the Hang Seng stock index to the KOSPI stock index at times when traditional linear Granger causality methods do not. This demonstrates that the framework is beneficial when the functional form of Granger causality is misspecified.
