In the context of Connected and Smart Cities, the need to predict short term traffic conditions has led to the development of a large variety of forecasting algorithms. In spite of various research efforts, there is however still no clear view of the requirements involved in network-wide traffic forecasting. In this paper, we study the ability of several state-of-the-art methods to forecast the traffic flow at each road segment. Some of the multivariate methods use the information of all sensors to predict traffic at a specific location, whereas some others rely on the selection of a suitable subset. In addition to classical methods, we also study the advantage of learning this subset by using a new variable selection algorithm based on time series graphical models and information theory. This method has already been successfully used in natural science applications with similar goals, but not in the traffic community. A contribution is to evaluate all these methods on two real-world datasets with different characteristics and to compare the forecasting ability of each method in both contexts. The first dataset describes the traffic flow in the city center of Lyon (France), which exhibits complex patterns due to the network structure and urban traffic dynamics. The second dataset describes inter-urban freeway traffic on the outskirts of the french city of Marseille. Experimental results validate the need for variable selection mechanisms and illustrate the complementarity of forecasting algorithms depending on the type of road and the forecasting horizon.