The probabilistic forecasting method described in this work is designed to leverage spatial and temporal dependency of urban traffic networks in order to provide accurate predictions for an horizon up to a few hours. By design it can deal with missing data both for training and running the model. It is able to forecast the state of the whole network in one pass with an execution time scaling linearly with the size of the network. The method consists in two learn a sparse Gaussian Copula of traffic variables, compatible with the Gaussian belief propagation algorithm. The model is learned automatically from an historical dataset through an iterative proportional scaling procedure well suited to impose this compatibility constraint. It is tested on three different datasets of increasing sizes ranging from 250 to 2000 detectors corresponding to flow or/and speed and occupancy measurements. The results show very good ability to predict flow variables and a reasonably good performance on speed or occupancy variables. Some element of understanding of the observed performance are given by a careful analysis of the model allowing
to some extend to disentangle modelling bias from intrinsic noise of the traffic phenomena and its measurement process.