We used empirical-statistical downscaling in a pseudo-reality context, where both large-scale predictors and small-scale predictands were based on climate model results. The large-scale conditions were taken from a global climate model and the small-scale conditions from the dynamical downscaling of the same global model with a convection permitting regional climate model covering South Norway. This hybrid downscaling approach, a “perfect model”-type experiment, provided 120-years of data under the CMIP5 high emission scenario. Ample calibration samples made rigorous testing possible, enabling us to evaluate the effect of empirical-statistical model configurations and predictor choices, and to assess the stationarity of the statistical models by investigating their sensitivity to different calibration intervals. The skill of the statistical models was evaluated in terms of their ability to reproduce the inter-annual correlation and long-term trends in seasonal 2-meter temperature (T2m), wet-day frequency (fw), and wet-day mean precipitation (μ). We found that different 30-year calibration intervals often resulted in differing statistical models, depending on the specific choice of years. The hybrid downscaling approach allowed us to emulate seasonal mean regional climate model output with a high spatial resolution (0.05° latitude and 0.1° longitude grid) for up to 100 GCM runs while circumventing the issue of short calibration time, and provides a robust set of empirically downscaled GCM runs.