1: \begin{abstract}

2: Causal inference from observational data is an ambitious but highly relevant task,

3: with diverse applications ranging from natural to social sciences.

4: Within the scope of nonparametric time series, 

5: causal inference defined through interventions (cf. Pearl~\citep{pearl}) is largely unexplored, although time order simplifies the problem

6: substantially. 

7: We consider a marginal integration scheme for

8: inferring causal effects from observational time series data, \mbox{MINT-T}

9: (\textbf{m}arginal \textbf{int}egration in \textbf{t}ime series), which is

10: an adaptation for time series of a method proposed by Ernest and B\"{u}hlmann (Electron. J. Statist, pp. 3155-3194, vol. 9, 2015)

11: for the case of independent data. Our approach for stationary stochastic

12: processes is fully nonparametric and, assuming no instantaneous

13:   effects consistently recovers the total

14: causal effect 

15: of a single intervention with optimal one-dimensional nonparametric

16: convergence rate $n^{-2/5}$ assuming regularity conditions and twice differentiability of a certain corresponding 

17: regression function. Therefore, \mbox{MINT-T}

18: remains largely unaffected by the curse of dimensionality as

19:   long as smoothness conditions hold in higher dimensions and it is

20: feasible for a large class of stationary time series, including nonlinear

21: and multivariate processes. For the case with instantaneous effects,

22:   we provide a procedure which guards against false positive causal

23:   statements. 

24: \end{abstract}