The validation of statistical downscaling methods traditionally involves using a subset of observations to evaluate the model's performance; however, this procedure does not allow the user to infer the model's future performance. To address this limitation, Vrac et al. (2007) used regional climate model (RCM) outputs as present and future "pseudo-observations" and compared them with the statistically downscaled series (SDS). Our proposed validation methodology extends the Vrac et al. approach by validating the SDS in terms of both the daily variability and the annual climate of extreme weather.

In particular, multiple linear regression (LR) and Bayesian neural networks (BNN) were used to statistically downscale the coarse resolution Canadian Coupled Global Climate Model (CGCM 3.1) inputs using the Canadian RCM (CRCM 4.2) maximum and minimum daily temperature outputs (TMAX and TMIN respectively) as "pseudo-observations" for the southern Quebec and Ontario region. The historical (1968-2000) and future (2038-2070) "pseudo-observations" from 10 CRCM 4.2 grid cells were compared against the SDS for both periods in terms of (i) mean absolute errors (MAE) to determine the models' performance in simulating the daily variability, and (ii) indices of agreement (IOA) calculated from the annual STARDEX climate indices used to determine the performance in simulating the annual climate of extreme weather.

Relative to the historical period, the future period shows that although the MAEs between the SDS and the pseudo-observations were larger for both the linear and nonlinear models, the nonlinear ones outperformed their linear counterparts in the MAE and in reproducing the 90th percentile of TMAX, the 10th percentile of TMIN, intra-annual temperature range, number of frost days, and to a lesser extent the growing season length and the heat wave duration (these being the 6 STARDEX climate indices studied).