In the digital age, data are collected in unprecedented volumes on a plethora of networks. These data provide opportunities to develop our understanding of network processes by allowing data to drive method, revealing new and often unexpected insights. To date, there has been extensive research into the structure and function of complex networks, but there is scope for improvement in modelling the spatio-temporal evolution of network processes in order to forecast future conditions. This thesis focusses on forecasting using data collected on road networks. Road traffic congestion is a serious and persistent problem in most major cities around the world, and it is the task of researchers and traffic engineers to make use of voluminous traffic data to help alleviate congestion. Recently, spatio-temporal models have been applied to traffic data, showing improvements over time series methods. Although progress has been made, challenges remain. Firstly, most existing methods perform well under typical conditions, but less well under atypical conditions. Secondly, existing spatio-temporal models have been applied to traffic data with high spatial resolution, and there has been little research into how to incorporate spatial information on spatially sparse sensor networks, where the dependency relationships between locations are uncertain. Thirdly, traffic data is characterised by high missing rates, and existing methods are generally poorly equipped to deal with this in a real time setting. In this thesis, a local online kernel ridge regression model is developed that addresses these three issues, with application to forecasting of travel times collected by automatic number plate recognition on London’s road network. The model parameters can vary spatially and temporally, allowing it to better model the time varying characteristics of traffic data, and to deal with abnormal traffic situations. Methods are defined for linking the spatially sparse sensor network to the physical road network, providing an improved representation of the spatial relationship between sensor locations. The incorporation of the spatio-temporal neighbourhood enables the model to forecast effectively under missing data. The proposed model outperforms a range of benchmark models at forecasting under normal conditions, and under various missing data scenarios.