The use of High Frequency Radar (HFR) data is increasing worldwide for operational oceanography and data assimilation, as it provides real-time coastal surface currents at high temporal and spatial resolution. In this work, a Lagrangian based empirical real-time, Short-Term Prediction (L-STP) system is presented in order to provide short term forecasts of up to 48 hours of ocean currents from HFR data. The method is based on the finding of historical gridded analogues of Lagrangian trajectories obtained from HFR surface currents. Then, assuming that the present state will follow the same temporal evolution as did the historical analogue, we obtain a short-term prediction of the surface currents. The method is applied to two HFR systems covering two areas with different dynamical characteristics: the southeast Bay of Biscay and the central Red Sea. The L-STP improves on previous prediction systems implemented for the SE Bay of Biscay and provides good results for the Red Sea study area. A comparison of the L-STP methodology with predictions based on persistence and reference fields has been performed in order to quantify the error introduced by this Lagrangian approach. Furthermore, a temporal sensitivity analysis has been addressed to determine the limit of applicability of the methodology regarding the temporal horizon of Lagrangian prediction. A real-time skill-score has been developed using the results of this analysis which allows to identify periods when the short-term prediction performance is more likely to be low and persistence can be used as a better predictor for the future currents.