When the Models Compete, the Forecast Wins: Dynamic Weighting for NWP

Operational forecasting systems routinely blend outputs from multiple numerical weather prediction (NWP) models, yet these blends often rely on static weights that do not reflect how model skill fluctuates in time and space. This study explores a simple dynamic weighting approach that updates the contribution of two global NWP systems - ECMWF and ACCESS, using both their deterministic and ensemble variants - based on their recent performance. By assigning weights proportional to inverse RMSE over a rolling 30-day window, the method adapts automatically to evolving model behaviour across regions, lead times, and variables.

The experiment focuses on 2-m temperature, 2-m dewpoint, and 10-m wind speed during two contrasting seasonal periods (December-February and June-August). Model fields are regridded to a common domain, and no bias correction or calibration is applied. This design isolates the effect of the weighting strategy itself, providing a clean assessment of how dynamic weighting compares to static, fixed-weight blends. Verification is performed against the Bureau of Meteorology's gridded Mesoscale Surface Analysis System (MSAS) and an extensive network of Automatic Weather Stations (AWS).

The largest gains occur for temperature and dewpoint at longer lead times, where the dynamic blend outperforms both individual models and the static blends. A distinctive result from the MSAS-based evaluation is the reduced amplitude of the diurnal RMSE cycle for temperature and dewpoint. This behaviour reflects the method's ability to exploit complementary biases - such as opposing warm and cold tendencies across models at different valid times - resulting in partial cancellation of systematic errors without explicit bias correction.

Wind speed forecasts also benefit, though improvements are more modest, reflecting the inherently higher variability and sensitivity of wind fields. Small regions of degradation emerge in areas with sparse observations or complex terrain, highlighting the limitations of retrospective RMSE when reference data are uncertain or unrepresentative.

Overall, this work demonstrates that a transparent, performance-based dynamic weighting strategy can deliver meaningful forecast improvements using only raw model output. Its low operational cost and ability to adapt to evolving model skill make it a promising candidate for next-generation blending approaches and a foundation for future regime-dependent or probabilistic adaptive systems.