Accurate prediction of extreme temperatures is vital for human health, energy management and agricultural practices and planning. We have developed and implemented a short-term and sub-seasonal statistical–dynamical forecasting scheme that performs significantly better than climatology or persistence forecasts through a lead-time of three weeks. The climatology of extreme temperatures since 1989 was completed by analyzing the historical European Centre for Medium-Range Weather Forecasts (ECMWF) ERA-Interim reanalyses, ECMWF Monthly Forecast System hindcast runs, and METAR observations from 105 stations across the U.S. Then the predictability of extreme temperatures for various U.S. regions based on the ensemble forecast spread and the phase/amplitude of the Madden Julian Oscillation was investigated using the ECMWF Atmospheric Variable Ensemble Prediction System (1-15 days) and Monthly Forecasting System (1-32 days). The information available from these products was statistically-downscaled and corrected for model biases and distributional errors using a quantile-to-quantile correction. To account for station-specific temperature tendencies, an error minimization procedure using past forecasts and observations was developed to determine the optimum forecast window that would remove station optimally biases based on the current and forecast weather regime. Finally, using the hindcast runs and historical observations, theoretical temperature distributions were generated from extreme value theory, which led to the production of real-time temperature forecasts and anomalies out to a lead-time of 30 days.