Predictions of Nino3.4 SST in CFSv1 and CFSv2: a diagnostic comparison

Diagnostic evaluations of the relative performances of CFSv1 and CFSv2 in prediction of monthly anomalies of the ENSO-related Nino3.4 SST index are conducted using the common hindcast period of 1982-2009 for lead times of up to 9 months. CFSv2 outperforms CFSv1 in temporal correlation skill for predictions at moderate to long lead times that traverse the northern spring ENSO predictability barrier (e.g., a forecast for July made in February). However, for predictions during less challenging times of the year (e.g., a forecast for January made in August), CFSv1 has higher correlations than CFSv2. This seeming retrogression is caused by a cold bias in CFSv2 predictions for Nino3.4 SST during 1982-1998, and a warm bias during 1999-2009. Work by others has related this time-conditional bias to changes in the observing system in late 1998 that affected the ocean reanalysis serving as initial conditions for CFSv2. A posteriori correction of these differing biases, and of a similar (but lesser) situation affecting CFSv1, allows for a more realistic evaluation of the relative performances of the two CFS versions. After the dual bias corrections, CFSv2 has slightly better correlation skill than CFSv1 for most months and lead times, with approximately equal skills for forecasts not traversing the ENSO predictability barrier and better skills for most (particularly long-lead) predictions traversing the barrier. The overall difference in correlation skill is not statistically field significant. However, CFSv2 has statistically significantly improved amplitude bias, and visibly better probabilistic reliability, and lacks target month slippage as compared with CFSv1. Together, all of the above improvements result in a highly significantly reduced overall RMSE-the metric most indicative of final accuracy.