In the context of Smart Cities, there is a growing need to inform drivers, anticipate congestion and take action to manage the state of the traffic flow on the road network. This need has driven the development of a large number of traffic forecasting methods. The last decades have seen the rise in computing power, in storage capacity and in our ability to process information in real-time. More and more road segments are equipped with traffic sensors. These evolutions are new elements to take into consideration in order to design accurate traffic forecasting algorithms. Despite the large amount of research efforts on this topic, there is still no clear understanding of which criteria are required in order to achieve a high forecasting performance at the network scale. In this thesis, we study two real datasets collected in two main French cities: Lyon and Marseille. The Lyon dataset describes the traffic flow on an urban network. The Marseille dataset descrobes the traffic flow on urban freeways. We evaluate the performance of methods from different fields: time series analysis (autoregressive models), and different subfields of machine learning (support vector machines, neural networks, nearest-neighbors regression). We also study different neighborhood selection strategies in order to improve the forecasting accuracy, while decreasing the complexity of the models. We evaluate a well-known approach (Lasso) and apply for the first time on traffic data a method based on information theory and graphical models (TiGraMITe), which has shown very effective on similar physics applications. Our experimental results confirm the usefulness of neighborhood selection mechanisms in some contexts and illustrate the complementarity of forecasting methods with respect to the type of network (urban, freeway) and the forecasting horizon (from 6 to 30 minutes).