Statistical downscaling techniques have been developed for the generation of maximum and minimum temperatures in Greece. This research focuses on the four conventional seasons, and three downscaling approaches, Multiple Linear Regression using a circulation type approach (MLRct), Canonical Correlation Analysis (CCA) and Artificial Neural Networks (ANNs), are employed and compared to assess their performance skills. Models were developed individually for each variable (Tmax, Tmin), station and season. The accuracy of downscaled values has been quantified in terms of a number of performance criteria, such as differences of the mean and standard deviation ratios between observed and modelled data, the correlation coefficients of the two sets, and also the RMSEs of the downscaled values relative to the observed. All methods revealed that during the cool season Tmax tends to be better reproduced, whereas Tmin is overestimated, particularly over western Greece, which is characterised by higher orography. With respect to the warm season, the simulation of Tmax reveals greater divergence, whereas Tmin is better generated. The distinction between the three techniques is somewhat blurred. None of the methods were found to be superior to the others and each has been shown to be a good estimator in some cases. This study concludes that all proposed methods comprise useful tools for simulating daily temperatures, as the high correlation coefficients, between observed and downscaled values, have demonstrated. However, the importance of local factors, which affect the natural variability of temperature, has been emphasised indicating that the geography of a region constitutes an important and rather complex factor, which should be included in models to improve their performance.