Data revisions in macroeconomic time series are typically studied in isolation ignoring the joint behaviour of revisions across different series. This ignores (i) the possibility that early releases of some series may help forecast revisions in other series and (ii) the problems statitical agencies may face in producing estimates consistent with accounting identities. This paper extends the Jacobs and van Norden (2011) modeling framework to multivariate data revisions. We consider systems of variables, where true values and news and noise can be correlated, and which may be linked by one or more identities. We show how to model such systems with standard linear state space models. We motivate and illustrate the multivariate modeling framework with Swiss current account data using Bayesian econometric methods for estimation and inference.