Selecting reference streamflow forecasts to demonstrate the performance of NWP-forced streamflow forecasts

We assess four different reference forecasts for the purpose of measuring the skill of streamflow forecasts generated from Numerical Weather Prediction (NWP) model rainfall forecasts. The reference forecasts we investigate are 1) streamflow climatology, 2) persistence, 3) a hydrological model forced by zero rainfall and 4) a hydrological model forced by an ensemble of resampled historical rainfall. We assess performance of reference forecasts to lead-times of 9 days. Reference forecasts should be simple to produce, but also must be reasonably accurate to establish a robust performance threshold. We show that because streamflows are strongly autocorrelated, streamflow climatology is a very low performance hurdle to clear for any NWP-forced streamflow forecasts, particularly at short lead-times (<2 days). Conversely, because the shape of hydrographs is broadly predictable, persistence forecasts generally perform very poorly at longer lead times (>1 day). Using a hydrological model substantially improves the accuracy of reference forecasts, with resampled-historical-rainfall forced forecasts outperforming zero-rainfall-forced forecasts, particularly at longer forecast lead times. We argue that streamflow climatology and simple persistence are not accurate enough to be used as reference forecasts. We recommend the use of reference forecasts generated by resampled historical rainfalls as a robust performance benchmark of NWP-forced streamflow forecasting systems. We demonstrate the use of resampled-historical-rainfall forced reference forecasts to assess the performance of a new Australian ensemble streamflow and flood forecasting system developed by CSIRO and the Bureau of Meteorology.