The correlations coefficients (CC) between the observed and the forecasted rainfall anomaly at different pentad leads and for the five selected regions are shown in Figure. It is observed from the figure that except NEI, there is no significant improvement in the prediction skill of CFST382 run and both T382 and T126 skills are comparable in all the four pentad leads. On the other hand if we compare the same with that of GFSbc, it is observed that GFSbc shows better prediction skill in the pentad lead 2, 3 and 4 over CEI, SPI and MZI and in pentad lead 2 over NEI. Over NWI none of the models shows any improvement in the prediction skill.

The verification skill scores for the ensemble mean deterministic forecast are presented in

Figure (a),
Figure (b),
Figure (c),
Figure (d),
Figure (e).

The different skill score measures used in this study are: The Kupper skill score (KSS) and bias score proposed, the Heidke skill score (HSS) and Gerity score, the correlation coefficient and the root mean square error (RMSE). All these scores are shown over MZI and four homogeneous regions. It is observed that over all the regions the large correlation at pentad lead 1 significantly drops at pentad lead 4. CC values of 0.5 are chosen as the threshold value for useful prediction on pentad scale. After pentad 2 lead, the deterministic ensemble mean prediction skill drops below 0.5 over all the selected regions.

This confirms the uselessness of deterministic forecast alone after pentad 2 lead. However, considering large sample size of 288, all CC values above 0.14 are significant at 99% confidence level. The HSS and GSS are found to be significant at 99% confidence level even at pentad lead 4 over CEI and NWI. It is also evident from the figure that skill in predicting break is higher, followed by active and then normal over MZI and CEI at all lead pentads and at pentad lead 1 and 2 over SPI. Over NWI, skill in predicting break is higher at all lead pentads but active follows normal at pentad lead 4.

ROCs for the discussed three observed categories are also evaluated for the four homogenous regions and also for MZI using CFST382 and compared with CFST126 and GFSbc in

Figure (a),
Figure (b),
Figure (c),
Figure (d),
Figure (e).

The area under the ROC curve is used for the calculation of the probabilistic skill score and is plotted in Figure. It is observed that break is more predicable in all the pentad leads over MZI and CEI and in pentad lead 1 and 2 over NWI and SPI. It is also observed that GFSbc outperforms CFS in predicting the active phases over MZI, CEI (pentad lead 2, 3, 4) and SPI (pentad lead 3, 4).