Title: Inference of instantaneous causal relations in multivariate linear time series by stabilizing conditional distributions
Authors: Niklas Pfister - ETH Zurich (Switzerland) [presenting]
Jonas Peters - University of Copenhagen (Denmark)
Peter Buehlmann - ETH Zurich (Switzerland)
Abstract: The problem of inferring the causal variables of a response Y from a set of $d$ predictors $(X^1,\dots,X^d)$ is investigated. Given that $(Y_t,X^1_t,\dots,X^d_t)$ is a multivariate linear time series, we want to identify causal effects. This includes time instantaneous and lagged effects and therefore extends Granger causality. In contrast to only considering lagged effects, inferring also the instantaneous effects is a much harder task as the causal direction is unknown a priori. We present a method that makes use of heterogeneity patterns (or non-stationarity) in the data to detect the instantaneous causal relations and show that it satisfies some desirable statistical properties. To illustrate practical applicability we apply our method to a data set related to the monetary policy of the Swiss National Bank.