Estimating causal relations is vital in understanding the complex\ninteractions in multivariate time series. Non-linear coupling of variables is\none of the major challenges inaccurate estimation of cause-effect relations. In\nthis paper, we propose to use deep autoregressive networks (DeepAR) in tandem\nwith counterfactual analysis to infer nonlinear causal relations in\nmultivariate time series. We extend the concept of Granger causality using\nprobabilistic forecasting with DeepAR. Since deep networks can neither handle\nmissing input nor out-of-distribution intervention, we propose to use the\nKnockoffs framework (Barberand Cand`es, 2015) for generating intervention\nvariables and consequently counterfactual probabilistic forecasting. Knockoff\nsamples are independent of their output given the observed variables and\nexchangeable with their counterpart variables without changing the underlying\ndistribution of the data. We test our method on synthetic as well as real-world\ntime series datasets. Overall our method outperforms the widely used vector\nautoregressive Granger causality and PCMCI in detecting nonlinear causal\ndependency in multivariate time series.\n