Two sets of experiments with the coupled system have been performed, differing only in the ocean initial conditions. In the first the ocean initial conditions are taken from the control (no-assimilation) run, while in the second they were taken from the assimilation run using scheme OI-1. The main integration period of the forecasts was 6 months. At the time of writing, 16 hindcasts have been carried out for the period 1992-6, consisting of four summer starts (July 1) for 1992-5, four autumn starts (Oct 1) for 1992-5, four winter starts (1 Jan) for 93-96 and four spring starts (1 April) for 1993-6 for ocean initial conditions obtained both with and without data assimilation. In addition, several additional forecasts have been made for 1996.
The coupled model is fully coupled: there is no anomally coupling in any field. The advantage of this approach is that it exposes any errors in the coupled model, be they in the atmosphere or ocean. The main disadvantage is that the drift in the coupled system can be large. Fig 8 shows the average drift for the EQ3 region for the first 6 months of a forecast, for 4 seasons of the year. It can be seen that the drift is as large as the interannual variability we want to predict. This is certainly of concern: if the variability is sensitive to the mean state, then the influence of drift will not be a linear modification to the signal. However, for the tropical response we assumed in the first instance that it was a linear modification, and subtracted the mean drift from the predicted SSTs.
In fig 9 the predictions for the Nino 3 region for 1994 are plotted, showing encouraging skill in reproducing the observed evolution of SST for the first 3 months and some modest skill for months 4 to 6. The observed anomally is shown by the continuous curve. Figure 10 shows the predictions for the Nino 4 region for 3 different years, showing good predictions out to 6 months. The observed SST anomalies are also shown by the dot-dash curve. The model was able to generate the warming of 94, the cooling in 95 and to maintain the weak anomalies of 1993. The mean drift was removed from the hindcasts before presenting the results in figs 9 and 10. For these events the skill of the model is very encouraging. Similar results occur at other start dates, not shown. All the forecasts shown are for start dates in the 1990s, a period when most models have done badly, as judged by the reports in the Climate Diagnostics Bulletin or the Experimental long-lead Forecast Bulletin. In that sense our results are encouraging. On the other hand the sample is small, and we do not yet know our ability to predict events in the 80s. The forecasts shown are for initial conditions obtained with data assimilation. There are cases when the analyses from the control, ie without sub-surface data assimilation, but with a strong feedback to SST lead to better forecasts than those initiated from assimilation analyses. However when averaged over a number of forecasts for the 1990s the rms errors are smaller in the case of the ocean inital conditions obtained using data assimilation as shown in fig 11.
For the present year the forecasts have differed in some respects from previous forecasts over the 92-95 period as seen in fig 1. First we had one of our worst forecast initiated from January initial conditions. The first three months of this forecast were good, the last three months were poor. This poor forecast occurred whether the ocean initial conditions were taken from the control run or the run with data assimilation. Forecasts started later, from April start dates also did poorly. The forecasts from other start times have been better and basically predicted no significant warming or cooling, as shown fig 12.
