GISS climate modeling is focused on decadal time scales. But seasonal climate predictions provide a valuable test of the model with both research and educational benefits. Our experimental forecasts are carried out by a student-educator-researcher team and used to study and contrast the roles of forcings and chaos in climate variations. This activity also contributes to our model assessment and model improvement activities.

These "forecasts" are not intended for operational purposes. Two major limitations that we would like to emphasize are:

1. These forecasts are driven by observed sea surface temperature (SST) anomalies for the week May 24-30, with these anomalies taken as constant for the period June-August. In reality observed SST anomalies are in a period of rapid transition in the tropics. Thus presumably we could obtain a better forecast by using predicted SST anomalies. But given our objective, summarized above, our approach is to compare two extreme simulation results: (a) those presented here, with fixed SST anomalies, and (b) those that will be obtained after the fact using observed SSTs, the latter representing a perfect SST prediction. This approach is well suited for research and education objectives, but it is not expected to yield the best forecast.

2. We do not initialize either the atmosphere state or the ground moisture. Although the atmospheric state may have little influence after the first week, we have found in previous simulations that the forecasts are sensitive to the initial soil moisture (Hansen et al., 1997; ELLFB, June 1996) and thus these initializations would be desirable in forecasts intended for operational purposes.

The global climate model (SI97) employed for the present seasonal forecast experiments is basically the same as the SI95 version of the model, which was recently described by Hansen et al. (1997). The three changes between the SI95 and SI97 models, involving changes in the large scale clouds, the sea ice puddling parameterization, and the vertical resolution (increase from 9 to 12 layers), are briefly described by Wilder et al. in the December 1997 issue of the Experimental Long-Lead Forecast Bulletin.

We expect the changes between SI95 and SI97 to have little effect on the model's forecast skill. This expectation tends to be confirmed by global maps of the correlation of the ensemble-mean of model simulations with observations for the period 1979-1997. This measure of the model's skill is shown in Fig. 1 for the SI97 model for surface air temperature and precipitation for Northern Hemisphere (NH) summer (Jun-Jul-Aug). The ensemble of model simulations in this case was five runs with the 12-layer version of the SI97 model driven by observed sea surface temperatures and sea ice cover.

The median of the histogram of correlation coefficients (not shown) is positive with respect to surface temperature, suggesting that SST has some influence upon the summertime circulation. However, the correlations rarely exceed 0.5 within the interior of the extratropical continents, away from the coastal regions where SST has a comparatively greater influence. If the model were to have no predictive skill and the 18 summers are statistically independent, then the absolute value of the correlation would be within 0.47 as a result of random chance 95% of the time. This suggests that summertime predictability within the interior of the extratropical continents is small, as might be expected from the chaotic nature of the atmosphere. However, the model's skill may be revealed to be statistically distinct from zero in these regions if a longer observational record is used for comparison. Furthermore, predictability may increase during a period of unusually strong forcing, such as an intense El Nino or La Nina.

The forecasts were obtained by extending 5-run ensembles of runs that had been made with the SI97 model using observed SSTs for the period January 1979 through April 1997. On May 1 each of these 5 runs was split into 5 runs by perturbing the atmospheric temperatures randomly by up to 1C. The resulting 25 runs were extended through May 30 using observed SSTs. For the 3-month extension from May 31 through August 30 1998, the SST anomalies were kept fixed at the observed values for the week May 24-30. Thus the only mechanism providing any potential predictability was the observed (late May) SST anomaly. No attempt was made to initialize the atmospheric or surface state (soil moisture, for example) with observations.

The calculated surface air temperature and precipitation anomalies for Jun-Jul-Aug are shown in Fig. 2. The results are based on 25-run means.

The consistency of the prediction is illustrated in Fig. 3, which shows the frequency with which the simulated anomaly has the same sign as in the ensemble mean. For example, if at a given gridbox 24 of the 25 runs yield an anomaly of the same sign, we define the consistency as 96%.

In those regions where the consistency is high the model is making a strong prediction, but that does not necessarily mean that it has a high reliability. On the other hand, a low correlation coefficient (i.e., a poor track record of the model in its 17 year simulations) does not necessarily mean that predictability in a specific year, which may have an unusual forcing, is small. But if the consistency of the prediction is not much more than 50%, we clearly do not have much confidence in the forecast.

Several features in the simulated surface air temperature and precipitation anomalies stand out. In the following list we indicate in parenthesis the model skill (correlation coefficient) in that particular region as well as the "model's confidence", i.e., the percentage of time that the model yields the same sign for the simulated anomaly in that region (Fig. 3).

1) North Africa: prediction is for unusually warm (cc ~ 0.5, ssa > 90%) and dry (cc ~ 0.25, ssa ~ 90%) conditions in the western Sahara.

2) Brazil: prediction is for unusually warm (cc ~ 0.7, ssa > 90%) and dry (cc ~ 0.6, ssa ~ 90%) conditions in Nordeste region.

3) China and Indochina: prediction is for warm (cc ~ 0.5, ssa > 90%) and dry (cc ~ 0, ssa ~ 60%) in Southeast China and the Vietnam-Thailand region. But note that confidence in the precipitation forecast is so low that it is not meaningful.

4) India: prediction is for greater than normal precipitation (cc ~ 0, ssa ~ 0.8). The model's "self confidence" (consistency of prediction) is high for precipitation, but its track record is poor.

5) Greenland: prediction is for unusually warm (cc ~ 0.5, ssa > 90%) and wet (cc ~ 0, ssa ~80%).

6) Antarctica: prediction is for unusually warm (cc ~ 0, ssa > 90%).

7) United States: prediction is for cool (cc ~ 0, ssa ~ 70%) and wet (cc ~ 0, ssa ~ 70%) conditions in the upper midwest.

8) Alaska: prediction is for warm (cc ~ 0.7, 0.85) conditions.

9) Asia: prediction is for the bulk of the continent to have a temperature (cc ~ 0.2, ssa ~ 80%) above normal.

10) Mexico: prediction is for most of Mexico to be warm (cc ~ 0.6, ssa ~ 80%) and dry (cc ~ 0.3, ssa ~ 0.8).

References:

Del Genio, A.D., M.S. Yao, W. Kovari and K.K.W. Lo, 1996: A prognostic cloud water parameterization for global climate models, J. Climate, 9, 270-304.  

Hansen, J., K. Beckford, S. Borenstein, E. Brown, B. Cairns, S. DeCastro, L. Druyan, M. Kelly, A. Luckett, R. Miller, R. Ruedy and J. Wilder, 1996: Forecasts of surface air temperature and precipitation using GISS SI95 model driven by SST and soil moisture anomalies, Exp. Lon-Lead For. Bull., 5, 1-6.

Hansen, J., M. Sato, R. Ruedy, A. Lacis, K. Asamoah, K. Beckford, S. Borenstein, E. Brown, B. Cairns, B. Carlson, B. Curran, S. de Castro, L. Druyan, P. Etwarrow, T. Ferede, M. Fox, D. Gaffen, J. Glascoe, H. Gordon, S. Hollandsworth, X. Jiang, C. Johnson, N. Lawrence, J. Lean, J. Lerner, K. Lo, J. Logan, A. Luckett, M.P. McCormick, R. McPeters, R. Miller, P. Minnis, I. Ramberran, G. Russell, P. Russell, P. Stone, I. Tegen, S. Thomas, L. Thomason, A. Thompson, J. Wilder, R. Willson and J. Zawodny, 1997: Forcings and chaos in interannual to decadal climate change, J. Geophys. Res., 102, 25,679-25,720.

Wilder, J., K. Beckford, S. Borenstein, L. Druyan, A. Estrella, J. Hansen, J. Knox, R. Miller and R. Ruedy, 1997: Winter forecast by the GISS SI97 model based on fixed SST anomalies, Exp. Long-Lead For. Bull., 6, 2-8 (see also "Erratum" on ELLFB web page).

Figure Captions: 

Fig.1. Correlation of observed 1979-1997 Jun-Jul-Aug surface air temperature (A) and precipitation (B) with 5-run mean of SI97 simulations, the model being forced by observed SST.

Fig. 2. Surface air temperature and precipitation anomalies for 25-run mean of simulations of Jun-Jul-Aug 1998 with the GISS SI97 model.

Fig. 3. Frequency with which the simulated anomalies have the same sign as the mean simulated anomaly for surface air temperature and precipitation.