We used the analogue method to generate ensemble projections of local daily mean, maximum and minimum air temperatures in the NW Iberian Peninsula until the middle of this century. A 3-step method was followed. (1) The error of the analogue method under optimal conditions was estimated, using air temperatures at 850 hPa and mean sea level pressure from reanalysis data as predictor variables. (2) The method's error under suboptimal conditions was assessed by taking these predictors from control runs of a multi-model, multi-initial-conditions ensemble of global climate models. Neither the predictor data nor the downscaled series were corrected. Under these suboptimal conditions, none of the individual downscaled series could robustly reproduce the cumulative distribution function (CDF) of the observations in any season of the year. However, when the single downscaled series were combined into a multi-model series, CDFs were reliably reconstructed for summer and autumn. (3) Temperature series were downscaled from the ensemble’s scenario runs and compared to observations in the reference period to detect local climate change. In addition to the mean relative warming, it can be shown that the less frequent the event in the reference period, the higher its frequency increase and the broader its uncertainty interval in the scenario period. This tendency is more pronounced for daytime than for night-time heat/warm events, leading to a tripling to quadrupling of the former in summer. The local projections’ uncertainty intervals are dominated by model errors rather than by forcing or initial-conditions uncertainties.