In this study, a method is presented to maintain real-time positioning at the decimetre-level accuracy during breaks in reception of the measurement corrections from multiple reference stations. The method is implemented at the rover by estimating prediction coefficients of the corrections during normal RTK positioning, and uses these coefficients to predict the corrections when reception of the corrections is temporarily lost. The paper focuses on one segment of this method, the on-the-fly prediction of orbital corrections. Frequently, only a few minutes of data representing short orbit ‘arcs’ are available to the user before losing radio transmission. Thus, it would be hard for the rover to predict the satellite positions using equations of motion. An alternative method is proposed. In this method, GPS orbital corrections are predicted as a time series and are added to the initial positions computed from the broadcast ephemeris to compute relatively accurate satellite positions. Different prediction approaches were investigated. Results show that the double exponential smoothing method and Winters' method can be successfully applied. The latter, however, has a better performance. The impact of the data length used for estimation of the prediction coefficients and the selection of seasonal lengths in Winters' method were investigated and some values were recommended. In general, the method can give orbital correction estimation accuracy of less than 5 cm after 15 minutes of prediction. This will result in a positioning accuracy better than 5 cm.