It is critical in the study of crustal deformation to reduce TRF-related system errors and enhance the accuracy of GPS coordinate time series. To reduce system errors associated with the coordinate time series, which are related to the terrestrial reference frame (TRF) realization, we develop a recursive TRF realization strategy in regional GPS data processing. We have processed the whole set of CMONOC (Crustal Movement Observation Network of China) GPS data by using the Bernese 5.2 software and employed the controlled datum removal (CDR) filter to deal with the rank defect problem in the daily coordinate normal equations (NEQs). On the recursive TRF realization, we sequentially perform time series modeling with integral trajectory models, TRF realization with all continuous stations treated as “pseudo” fiducial stations, and frame alignment to the ITRF2014 with the 6-parameters Helmert transformation, in an iterative mode. We obtain the final coordinate time series through 3 times of iterations. Compared to the results derived from the conventional TRF realization strategy with the average root mean squares (RMS) being 1.98 mm, 2.62 mm and 5.39 mm for the east, north and up components, respectively, the average RMS scatters earn significant reduction up to 30%, 43% and 16% in the first loop, with their quantities being 1.41 mm, 1.51 mm and 4.57 mm for the east, north and up components, respectively, and negligible changes in the following 2 loops. In contrast to previous studies, our strategy is feasible in the processing of regional geodetic networks and concentrated on the TRF-related system errors without any pre-assumption and spatial limitation. In essence, the recursive strategy enhances the weights for the CMONOC stations and inevitably weakens the weights for stations exterior to the CMONOC to accommodate the misfits between the referential and estimated daily solutions in the least squares adjustment, thus resulting in the slightly increase of RMS scatters for the coordinate time series of globally distributed stations. For the CMONOC stations, the north component of coordinates time series has a maximum RMS reduction, revealing the identical precision for both horizontal components, thus indicating that our strategy remedies the frame defects stemming from the extremely uneven distribution of the fiducial network, and retrieves the “real” precision of GPS observations. The insignificant RMS reduction on the vertical component may be attributed to the insufficiency in the time series modeling. Our recursive TRF realization strategy can benefit the velocity estimation for campaign stations.