If, therefore, we can predict the boundary conditions of the atmosphere at a specific time, particularly SST and sea ice, we will have some information about the statistics of the atmosphere at that time. We will not be able to predict the precise state of the atmosphere, because it can vary in equilibrium with the predicted boundary conditions, but we will know something about its average conditions. We may be able to predict the average monthly or seasonal precipitation over a region even if we cannot say on what specific day the precipitation will fall. Knowing only a mean value at a given time can still be helpful if the associated variability is small. Predictions of tropical boundary conditions at a certain time are likely to be useful because tropical climate variability is low, while predictions of mid-latitude boundary conditions would be less useful because mid-latitude variability, especially during winter, is high.
Two more important points need to be made. First is the distinction between initialized and uninitialized prediction. To make a prediction about a specific time in the future, say the summer of 2009, there must be some connection to the actual conditions now. We call this estimation of the actual beginning state of the system "initialization" (while recognizing that this term is sometimes used elsewhere to mean the act of bringing a model system to a state of equilibrium without estimating its current conditions). If we do not make this initial estimation, we will not be able to forecast the time at which the climate will assume a given state, though we may still draw conclusions about its statistics (that is, changes of its mean and the nature of its variability). The difference between initialized and uninitialized prediction becomes important in discussing greenhouse-warming predictions versus ENSO predictions. The second important point is the potential for making empirically or statistically based (analog) predictions. Sufficient information is available from past climate records to allow predictions to be made (with specified uncertainty) whenever specific climate states exist that have in the past been accompanied or followed by particular regional or local climate conditions. "Climate state" is defined here, as in NRC (1975), as the average of the complete set of atmospheric, hydrospheric, and cryospheric variables over a specified period of time in a specified domain of the earth-atmosphere system.
For climate prediction on all time scales, whether initialized or not, the tool for predicting the boundary conditions of SST and sea ice is the coupled climate modela model that consistently links the atmosphere, ocean, and ice together in responding to a specified external forcing.
There is no accepted terminology describing the various time scales for prediction. This report will use "short-range climate prediction" to denote prediction on time scales up to interannual, "medium-range climate prediction" to mean prediction at decadal time scales, and "long-range climate prediction'' (sometimes called "greenhouse prediction'') for prediction on centennial time scalesthe scale of a human lifetime.
Short-range climate prediction is an established enterprise: Skill has been demonstrated for predicting the SST changes in the tropical Pacific that are characteristic of the ENSO phenomenon on lead times of 6 to 12 months. Atmospheric properties elsewhere may then be inferred from these forecasts. These predictive skills, which vary as a function of several factors (including season, model type, and decade), have been well documented (Battisti and Sarachik, 1995; Glantz, 1996; Latif et al., 1998). ENSO prediction is initialized prediction (in the sense defined above), so a real-time observing system in the tropical Pacific was put in place by the TOGA research program. It has been kept in place even though TOGA has ended, which should permit us to develop our skill further.
Long-range climate prediction has so far been limited to predicting forced climate change in response to the anthropogenic addition of radiatively active gases and aerosols to the atmosphere. Because this type of prediction is essentially uninitialized, it cannot predict the actual state of the boundary conditions at some specific future time. It can, however, be used to derive the statistics of the boundary conditions (and therefore the statistics of the atmosphere in equilibrium with the statistics of the boundary conditions) at some future time. Thus, initialized short-range climate prediction can predict the SST in the tropical Pacific for January of 1999, say, while greenhouse predictions can only say that annually averaged SST will be warmer in the year 2050 by some specified amount, or within a certain range. Such greenhouse predictions are still valuable if the forcing changes the mean boundary conditions enough for a difference beyond natural variability to be apparent; again, small shifts of the mean may be noticeable in the tropics where the variability is low, while larger shifts may be masked in mid-latitudes where variability is high. Long-range forecasts permit the assessment of shifts in average precipitation, or length of the growing season, or changes in patterns of runoff; as indicated by Karl et al. (1996), even subtle shifts in the mean state can have considerable implications for the frequency and magnitude of extreme climate events.
Medium-range climate prediction, prediction on time scales of a decade or so, is the most problematic type of prediction. Its value as uninitialized prediction is limited: The year-to-year variability of climate, together with the relatively slow approach of the climate system to equilibrium with anthropogenically added radiatively active atmospheric constituents, limit the value of prediction of the statistics of boundary conditions a decade in advance. Even this type of prediction may be useful under certain circumstances, however. When regional changes are fast and crossing the threshold of a new climate state can be predicted,