Performance-based weighting improves forecast skill without calibration

Post-processing in Numerical Weather Prediction (NWP) is essential for transforming raw model outputs into forecasts that are both accurate and actionable. This study investigates a dynamic weighting approach for multi-model blending, in which the contribution of each model is adjusted in real time based on recent performance, using root mean square error (RMSE) as a skill metric. We benchmark this approach against three baseline approaches: equal weighting, optimised static weighting and spatially static weighting derived from a training period. Experiments are conducted using deterministic and ensemble model outputs from two global NWP systems (Australian Community Climate and Earth System Simulator, ACCESS, and European Centre for Medium-Range Weather Forecasts, ECMWF) over two seasonal periods and are verified against both gridded analyses and surface station observations. Results show that dynamic weighting consistently improves forecast skill relative to both equal and static weighting approaches, particularly at longer lead times, and closely matches the performance of the operational Australian Bureau of Meteorology’s calibrated and bias-corrected IMPROVER products, even though the experiments presented here omit the bias-correction and calibration stages applied operationally. This suggests that adaptive performance-based weighting alone can recover a substantial fraction of the skill typically attributed to more complex post-processing.