Testing for Granger causality in large mixed-frequency VARs

Abstract

Reproduction permitted only if source is stated.ISBN 978-3-95729-217-9 (Printversion) Non-technical summary Research QuestionWhen dealing with time series sampled at various frequencies it has become common practice to directly incorporate high-frequency information into the econom(etr)ic model at hand. These specifications were first restricted to the single regression case; with the development of the (stacked) mixed-frequency vector autoregressive (MF-VAR) system (Ghysels, 2015) it is now possible to treat all series similarly and investigate causal effects between them. However, if the difference in frequencies between the series involved is large (as, e.g., in a month/working day scenario), estimation accuracy of the system coefficients is exacerbated, implying the detection of causal effects to be potentially inaccurate. To overcome this issue various parameter reduction techniques are introduced and analyzed. These methods are then evaluated in terms of their ability to detect causality patterns between the series under consideration in the resulting restricted model. ContributionTwo parameter reduction techniques are discussed in detail: three reduced rank regression (RRR) model variants and a Bayesian MF-VAR. Using a Monte Carlo experiment both approaches are compared in terms of their Granger causality testing behavior with an unrestricted VAR, a (time-aggregated) low-frequency VAR and the max-test (Ghysels, Motegi, and Hill, 2015a). To further enhance their finite sample properties we develop and evaluate (whenever possible) two bootstrap variants of these tests. Finally, the methods are applied to U.S. data by investigating channels of causality between the monthly growth rate of the industrial production index (IPI) and daily bipower variation (BV) of the S&P500 index. ResultsWe find that, depending on the direction of causality under consideration, a different set of tests results in the best Granger non-causality testing behavior. For the direction from the high-to the low-frequency series, standard testing within the Bayesian MF-VAR, the max-test and the restricted bootstrap version of the Wald test in two RRR model versions performs best. For the reverse direction, the unrestricted bootstrap variants of the Bonferroni-corrected Wald tests within the unrestricted VAR and the RRR models dominate. As far as our application is concerned, Granger causality from BV to IPIgrowth is clearly supported by the data; evidence for causality in the reverse direction, however, only comes from a subset of tests. Maastricht UniversityAbstract We analyze Granger causality testing in a mixed-frequency VAR, where the difference in sampling frequencies of the variables is large. Given a realistic sample size, the number of high-frequency observations per low-frequency period leads to parameter proliferation problems in case we attempt to estimate the model unrestrictedly. We propose several tests based on reduced rank restrictions, and implement bootstrap versions to account for the uncertainty when estimating fac...

Testing for Granger causality with mixed frequency data

“…16 A recent work by Götz and Hecq (2014) attempts to resolve the parameter proliferation issue by using reduced-rank regressions and Bayesian estimation.…”

Features of the assessment of factors influencing productivity

“…Granger causality test is closely related to the vector autoregression (VAR) model (Götz, Hecq, & Smeekes, 2016). C. Granger (1969) presented the concept of causality, which states that the existence of X dependence on Y and the knowledge of the values of X and Y can predict the trend of Y. the Granger causality test for time series is based on the assumption that if X affects Y, then X changes should occur before Y changes, but not the other way around.…”

Google data in bridge equation models for German GDP

“…(vi) Adaptive LASSO: To address the potential inconsistency of the usual LASSO estimator, adapted Lasso (AdaLASSO hereafter) versions have been introduced (see, e.g., Zou, 2006 8 PLS can be interpreted as a middle ground between PCA and canonical correlations analysis (CCA hereafter), where the target variable is usually a vector rather than a single time series (Götz et al, 2016). In CCA, linear combinations on both sides of the equation are determined in such a way as to maximize the covariance between them (again conditional on them being orthogonal to the previous factors).…”