Assessment of merged seasonal streamflow forecasts for 40 catchments across Australia

The Bureau of Meteorology operates a probabilistic seasonal streamflow forecasting service based on a statistical Bayesian joint probability modelling approach developed by CSIRO. While operating the statistical service, the Bureau has advanced the development of a dynamic forecasting approach. Most of the development has occurred in-house with assistance from CSIRO and universities. The Bureau is planning to modify its service to issue forecasts derived by merging the statistical and dynamical forecasts. Methods for merging probabilistic seasonal streamflow forecasts have been developed. Two methods that work well are quantile model averaging (QMA) and Bayesian model averaging (BMA). BMA averages forecasts densities whereas QMA averages forecast quantiles. A recent evaluation of BMA and QMA for 12 catchments in eastern Australia found that forecast merging is able to take advantage of the better forecasting approach in a particular catchment or season, boosting overall forecast skill. BMA and QMA perform similarly in terms of overall skill scores and reliability in ensemble spread. However, QMA merged forecasts can be narrower, uni-modal and generally more smoothly shaped, which has advantages for forecast communication and user interpretation. Before operationally deploying forecast merging, further evaluation of the merging methods at more locations under more varied climatic conditions is required. In this study, forecast merging is assessed for 40 forecast locations in varied geographic and climatic regions of Australia, including very dry catchments. Findings are consistent with the previous study. Forecast merging improves overall skill across forecast locations and seasons, BMA and QMA perform overall similarly in terms of skill scores, and merged BMA and QMA forecasts are similarly reliable in ensemble spread.