Advanced driver assistance systems like cooperative adaptive cruise control (CACC) are designed to exploit information provided by vehicle-to-vehicle (V2V) and/or infrastructure-to-vehicle (I2V) communication systems to achieve desired objectives such as safety, traffic fluidity or fuel economy. In a day to day traffic scenario, the presence of unknown disturbances complicates achieving these objectives. In particular, CACC benefits in terms of fuel economy require the prediction of the behavior of a preceding vehicle during a finite time horizon. This paper suggests an estimation method based on actual and past inter-vehicle distance data as well as on traffic and upcoming traffic lights. This information is used to train a set of nonlinear, autoregressive (NARX) models. Two scenarios are investigated, one of them assumes a V2V communication with the predecessor, the other uses only data acquired by on-board vehicle sensors. Depending on the applied approach and the moving space of the controlled vehicle, the thus obtained (imperfect) prediction allows fuel benefits in a range of 5% to 25% in the case of moderate, non-congested traffic. This is confirmed both by simulation and measurement. Compared to existing prediction methods, the proposed strategy delivers quite promising results.