This paper uses a smooth transition error-correction model (STECM) to model the one-year and five-year mortgage rate changes. The model allows for a non-linear adjustment process of mortgage rates towards their long-run equilibrium. We also introduce time-varying thresholds into the standard STECM specification, to capture the gradual structural changes in the error-correction term. We find that the STECM, whether with fixed or time-varying thresholds, yields better in-sample fit and lower forecast errors than the linear benchmark and univariate models. Our estimation results indicate non-linearities in the adjustment process of mortgage rates towards their long-run equilibria. In particular, we find that mortgage rates respond more significantly to a large than to a small disequilibrium. The improvement of the STECMs in forecasting is statistically significant over the univariate models, but insignificant over the linear model.