Improving seasonal streamflow forecast performance

River and tree-lined banks

River Murray, near Corowa, NSW. (photographer: Merryn Coutts)

While our seasonal streamflow forecasts are better now than they were five years ago, or even two years ago—whenever we ask users how we could improve, a common response is ‘provide forecasts with greater accuracy and reliability’.
Increased forecast performance will allow water managers to more confidently make planning and operational decisions.
To continually improve the forecasts, we have developed a long-term collaboration with the University of Adelaide and the University of Newcastle.
Recent research from this collaboration has provided a method to reduce uncertainty in hydrological predictions, which is likely to increase the performance of the seasonal streamflow forecasts.

This research collaboration is providing guidance on the best error model to use to reduce the uncertainty in the forecasts.  
We are currently testing the recommended error models, with initial results showing an increase in seasonal streamflow forecast performance for locations across Australia. Using the new method, the number of catchments where forecasts are now more reliable and precise for at least seven months has increased by 20 per cent (taking it to 75 per cent of seasonal streamflow forecast catchments).
Map of Australia showing seasonal streamflow forecast sites, indicating greater than 75% of catchments have forecast precision and reliability between 10 and 12 months
 The map shows the number of months for which forecasts are both reliable and more precise than climatology. The number of catchments where forecasts are more reliable and precise for at least 7 months has increased by 20 per cent using the new method.
The Bureau provides monthly updates of 3-month streamflow forecasts at more than 300 locations to registered users, and more than 160 locations for the public. The forecasts are presented as probabilities showing the likelihood, or chance, of a given volume of water flowing into a stream over the specified period.
This provides the user with the range of possible outcomes. Ideally, they want probabilistic predictions that are reliable, precise and unbiased. However, achieving this is challenging, as no single error model performs best in all these aspects of predictive performance.
The choice of error model is based on whether flows in the catchment are permanent or intermittent, and whether reliability, precision or low bias is most important.
Before any new models are released publicly, they are tested by registered users. To become a registered user please email