Post-processing of ensemble forecasts based on the standardized anomalies gradient boosting algorithm

In recent years, ensemble forecasting has become a vital tool for major global weather forecasting centers. However, ensemble forecasts often suffer from underdispersion and systematic biases, making the application of statistical post-processing methods essential. The Standardized Anomalies Model Output Statistics (SAMOS) is one commonly used post-processing technique that provides a complete description of the forecast distribution. Nevertheless, SAMOS typically relies only on predictors directly related to the forecast variable or selected based on subjective judgment, potentially overlooking other valuable predictors. Moreover, incorporating too many predictors directly into SAMOS may lead to overfitting. Therefore, effectively selecting key predictors from the numerous variables provided by forecast models remains a significant challenge. Boosting-based variable selection and optimization algorithms have proven effective in mitigating overfitting and identifying the most important predictors. This study proposes the Standardized Anomalies Gradient Boosting (SABST) method by integrating the strengths of SAMOS with a Boosting-based variable selection algorithm. SABST is applied to calibrate ensemble forecasts of 2 m temperature, 2 m relative humidity, and 10 m wind speed. The SABST model is developed using ECMWF high-resolution and ensemble forecast (ENS) data from 2019–2020 and is systematically compared with SAMOS, focusing on its calibration performance and bias correction ability. Results show that compared to ENS and SAMOS, SABST performs better in addressing underdispersion in probabilistic forecasts and improving the accuracy of deterministic forecasts. Based on the Continuous Ranked Probability Skill Score (CRPSS), SABST improves the average CRPSS by 9.5%, 15.3%, and 4.6% across all forecast lead times compared to SAMOS. These findings demonstrate the advantage of introducing additional potential predictors in the SABST framework and highlight the method’s potential for application in ensemble forecast post-processing.