The best model (or set of best predictors) based on AIC selects 2-3 predictors for almost all the lead times except for a few cases during withdrawal where it selects 4 predictors. The predictor sets and model skill scores can be found in Tables 3 (onset period) and 4 (retreat period). Furthermore, predictor sets were remarkably consistent between the south and middle clusters, notable exceptions being the 1 April south cluster model containing February North Atlantic SST, and the 1 May south cluster model containing April Guinea Coast 200mb meridional winds. Guinea Coast 600mb meridional winds were found in the north cluster replacing South Atlantic MSLP in the 1 March and 1 April models. Plots of observed, estimated and cross-validated estimates of south onset cluster relative humidity are shown in Figure 14. It can be seen that the model estimates and predicts the values very well at all the lead times. The corresponding scatter plots of the observations and estimates (Figure 14) shows that almost all of the predictions fall within the 5% of the observed values (dotted lines indicate ±5% lines). It is interesting to note that the models exhibit good skill at all lead times, especially at 74-day lead time – as can be seen by the R2cross validated R2 in Tables 3 and 4.
