PURPOSE to evaluate the reduction of correlation errors for both the clinical user-dependent and an automatic correlation model (CM) retraining procedure based on patient data acquired during real-time tumour tracking (RTTT). METHODS During clinical RTTT, tumour positions were detected using 0.5 Hz orthogonal kV fluoroscopy. Correlation models were retrained by three scenarios: (1) the current clinical procedure with a treatment interruption and rebuild of the correlation model using 20s fluoroscopy, (2) a single CM for the entire treatment without retraining and (3) a continuous retrained CM using an input of 0.5 Hz orthogonal verification images acquired during RTTT. The correlation errors (CE) were calculated by the difference between predicted and detected target positions. For all three retraining procedures, PTV margins were calculated using margin recipes. With the dose reconstruction by convolution of a static planned dose (5 mm PTV-margin) with a point spread function based on CE, the coverage of the CTV periphery was evaluated for RTTT with the different CM update approaches Results: The clinical CM update procedure showed a 0.9 mm margin reduction compared with omission of intra-fraction retraining. An additional 0.9 mm margin reduction was achieved by on-the-fly CM updating. Without interruptions for correlation remodelling, the treatment time could be reduced by 20% (1-5 minutes). On average, RTTT treatments showed a reduction of 2.4±3.1% for CTV D99 when using a clinical update and was equal to reconstructed CTV D99 values for a single CM treatment (1.6±3.3%) and for continuous CM update (1.6±2.0%). For 2 patients with a larger dose-loss (>5%) on the CTV periphery using the clinical RTTT routine, the continuous CM approach reduced the dose-loss to 2%. CONCLUSION A fast automatic update of the correlation model reduced PTV-margins and maintained an adequate CTV dose coverage for a large variety of breathing motion during RTTT. Corporate funding by BrainLab AG and Hercules Funding by the Flemish Government.