Empirical statistical downscaling methods are becoming increasingly popular in climate change impact assessments that require downscaling multi-global climate model (GCM) projections. Here, empirical statistical downscaling methods are classified based on calibration strategies [bias correction (BC) and change factor (CF)] and statistical transformations (mean based, variance based, quantile mapping, quantile correcting and transfer function methods). Ten combinations of calibration strategies and transformation methods are used to represent a range of empirical statistical downscaling methods. To test the performance of these methods in downscaling daily precipitation and temperature, an inter-model cross-validation is carried out using an ensemble of 16 GCMs from the Coupled Model Intercomparison Project Phase 5 (CMIP5) dataset over the Huai River Basin in China. These downscaling methods are further applied to downscale the climate for the future period to assess the associated uncertainties. The results show that the CF-based methods outperform the BC-based methods in projecting the probability distribution of downscaled daily temperature, while both calibration strategies give comparable results in the case of precipitation. With the CF calibration strategy, simply adding (for temperature) or multiplying (for precipitation) the mean CF is sufficient to represent most of the relative changes projected by GCMs. The use of quantile-based methods appears to be advantageous only at the tails of the distribution. More sophisticated BC methods are needed to remove the biases in the higher-order statistics of the GCM outputs. The two calibration strategies lead to fundamentally different temporal structures and spatial variability of the downscaled climatic variables. The BC-based methods produce larger uncertainty bands of inter-annual variability than the CF-based methods. For downscaled future precipitation, the uncertainty arising from the downscaling methods is comparable to the uncertainty arising from GCMs, while more uncertainty is introduced by calibration strategies than by statistical transformation methods.