Here we assess the hindcast probabilities of early onset using ACCESS-S2, focusing on an ensemble of 3 hindcast members, that have been calibrated to observations using a quantile-quantile mapping technique to reduce biases. Each hindcast member has been initialised at short intervals and run out for over 7 months, meaning that an onset date is undetermined if 50 mm of rain is not accumulated by the end of the run. For the probability of early onset below (left panel below), we combine 11 different initialisation dates to create a 33-member sample. For example, for Lead month 0, we combine the dates of August 16, 21, 24-31 and September 1. For each year, probabilities are calculated by summing the total number of hindcast ensemble members that show an onset earlier than the observed median (with respect to a smoothed 1960-2018 or 1990-2018 median), and dividing through by the total number of members. For an onset date falling on the median onset date, a 50-50 split is given between an early and late onset. The right panel below shows the observed onset date anomaly referenced against the median onset to judge the performace of the hindcast probability.
