Our aim is to investigate long range predictions (up to several wavelengths) of the small-scale fading of radio channels. The purpose is to enable advanced 5G downlink transmission schemes that require accurate channel state information at transmitters, such as massive MIMO and coherent joint transmission, for vehicular users. We here present a proof of concept for the recently introduced predictor antenna scheme which promises a significant increase in prediction horizon compared to conventional techniques. Predictor antennas utilize the exterior of moving vehicles by placing antenna arrays on top of their roofs. They are used to estimate the fading radio channels that are encountered later by the following antennas. The level of predictability is determined by the correlation between the channel measured at the predictor antenna and the channel that is later encountered by the following antennas when they move to that position. That correlation, and the resulting prediction errors, are assessed on a large set of measurement data sampled at vehicular velocities, at a carrier frequency of 2.53 GHz, from a multitude of urban fading environments. These represent a wide variety of propagation environments, including narrow and wide roads, intersections, dense urban environments and residential areas. Using low-pass filtered predictor antenna measurements, the obtained average prediction Normalized Mean Squared Error (NMSE) is -11 dB for prediction horizons of 0.25 wavelengths and -8.5 dB for horizons of 3 wavelengths. This represents an order of magnitude increase of the prediction horizons as compared to time-series prediction that typically, in practice, fails to work for prediction beyond 0.3 wavelengths in space. As a result, we have a tool that enables advanced 5G transmit schemes for vehicular users and vehicle-to-infrastructure links.