Kleeman et al. [1995] showed that for the period 1982-9 1, the ingestion of subsurface data improved both the anomaly correlation and rms error of the NINO3 hindeasts, particularly at lead times in the range 4-14 months. At shorter lead times the impact is neutral for the anomaly correlation, perhaps suggesting a conflict between wind information and thermocline information (the use of 400m depth-averaged temperature as a proxy is probably less appropriate local to the NINO3 region). Such results were sufficiently encouraging for the model to be accepted as part of the Bureau of Meteorology operational climate outlook systems.
These experiments have been extended here to the 1970-1994 period to test possible interdecadal variations in model skill and to ascertaln whether the limited subsuiface data for the first half of this period could still provide some useful information. Balmaseda et al. [1995] and others have found that there are significant variations in predictability and model skill with time, some perhaps due to extemal interactions and changes in "climate" [Kleeman et al. 1996]. It might be expected that such influences could confound the Kleeman et al. [1995] model to some extent since it is founded on simple dynamics and coupled air-sea interactions. The major events of the 1980's (the 1982-83 and 1986-87 warm events and the 1988 cool event) are apparently well suited to the mechanisms captured in this and other similar models.
Figure 4 shows the anomaly correlation and rms errors for model hindcasts of NINO3 over the early period 1971-1980 and for the entire period 1971-1994 for experiments using wind initialisation and no subsuiface data (CNTL), wind forcing and all available subsuiface data (FCST), and persistence (PERS). For the 1970's and lead times out to 12 months the anomaly correlation of the model with ocean data assimilation is consistently higher than the model using wind initialisation only, though the difference is rather small. There is a corresponding decrease in rms error of around 0.20C. In contrast to the 1980's, the model skill drops off rapidly beyond around 6 months - the inclusion of subsuiface data is perhaps giving an extra month in lead time. The fact that at longer lead times the CNTL experiment appears better (though with inconsequential skill) is probably due to the fact that it can take advantage of the full history of wind forcing and not just the most recent 12 months. The results accumulated over 197 1-94 indicate a positive impact on anomaly correlation for subsuiface data at lead times greater than 4 months, and positive impact on rms error at all lead times out to 16 months. The marginal positive impact in the 1970's, and perhaps the confounding and unusual circumstances of the 1990's, counterbalance to some extent the very good results from the 1980's. Because of the more extensive period used here (96 realisations c.f. 40) and the unambiguous nature of the positive impact of subsuiface data, the conclusions of Kleeman et al. [1995] might now be regarded as having even more solid foundation.
