Seasonal Forecasting for Australia using a Dynamical Model: Improvements in Forecast Skill over the Operational Statistical Model

In 2013 the Bureau of Meteorology transitioned from issuing seasonal climate outlooks based on statistical relationships between sea surface temperature (SST) patterns and Australian rainfall, to a dynamical model-based system, the Predictive Ocean Atmosphere Model for Australia (POAMA). The case for the move to POAMA is threefold: a) when assessed over a common period, POAMA shows higher forecast skill than the previous operational model; b) dynamical models are less susceptible to changes in statistical relationships, whether these occur naturally or through climate change; and c) dynamical models generate physically consistent forecasts of a range of climate phenomena over a number of timescales. The seasonal climate outlook for rainfall is the most popular product within the Bureau's climate prediction service. It is expressed as a probability of the seasonal (three month mean) rainfall exceeding the long-term median. In this study, independent retrospective forecasts from the statistical model and the dynamical model were assessed over a common period from 1981-2010. Previous assessments of dynamical model-based forecasts have identified that the forecast probabilities, as generated by a simple count of the frequency of ensemble members exceeding the median, tend to be overconfident. This overconfidence is overcome in practice through a time lagged ensemble strategy. By combining successive burst ensembles initialised on multiple start dates, the spread of forecast outcomes increases and emphatic probabilities are reduced, resulting in outlooks that are more reliable. A comparative assessment of accuracy and reliability of the new rainfall outlooks with the previous statistical seasonal forecasting system is made using percent consistent, attributes diagrams and the Brier score decomposition. This assessment shows that over the assessment period, the dynamical model-based system is sharper, more reliable and consistently more accurate over a larger spatial domain than the statistical model.