Are Multi-decadal Climate Forecasts Skillful?

In one of our July 11, 2005 posts, climate was defined so that climate forecasts are forecasts of the future state of the atmosphere, oceans, land, and continental glaciers, as defined using physical, chemical, and biological variables that we can measure. We can apply local, regional, or global averages over any time period we choose to characterize the future state of the climate. Weather forecasts are a subset of climate forecasts, in that we limit our forecasts to weather conditions, averaged over 12-hour periods, for example, out to a week or more, and generally assume a number of climate variables, such as vegetation and sea-surface temperatures, are invariant over this time period. It is important to note that the averaging time is not what distinguishes weather from climate (e.g., although called “seasonal climate predictions”, these forecasts are more accurately “seasonal-averaged weather predictions”).

As a necessary condition, climate forecasts must be able to skillfully reconstruct the observed temporal and spatial variability and change of local, regional, and global climate variables, when the forecast models are only given the external forcings (such as solar irradiance, volcanic eruptions, CO2 concentrations) as illustrated in Figure 1-2 in Radiative Forcing of Climate Change: Expanding the Concept and Addressing Uncertainties (2005). See also Tables 1 and 2 in Dynamical downscaling: Assessment of value retained and added using the Regional Atmospheric Modeling System (RAMS) where climate forecasts are called a Type 4 model simulation.

In 2000, we published a paper which demonstrated that the general circulation models were unable to skillfully reconstruct even the globally-averaged mid-tropospheric temperature trend during the 1979-2000 time periods. Thus, as of that date, the climate prediction models were shown to not be able to skillfully forecast the future climate even with respect to a single globally-averaged climate variable. (I am on a CCSP committee entitled “Temperature Trends in the Lower Atmosphere: Steps for Understanding and Reconciling Differences” and will update our assessment of the issue of climate prediction skill as soon as the report is public).

Mike MacCracken in his essay response to my Climatic Change essay seeks to distinguish a “prediction” from a “projection.” However, this only obscures the discussion, as GCM results are obviously packaged as forecasts in that specific time periods in the future are presented (see, as just one example, the 2070-2100 forecasts of the United Kingdom Hadley Centre). Even Mike recognizes there is no regional predictive skill in his paper entitled “Reliable regional climate model not yet on horizon.”

A conclusion of our evaluation is that papers which appear in the literature that present future values of a subset of (or all) climate variables are misrepresenting their results by implying that they are forecasts. They should be presented as sensitivity studies (as a process study; see my July 15 post on the types of model applications).

We can illustrate their misuse as forecasts by an analog. If we could run a numerical weather prediction model to provide a forecast of rainfall for tomorrow and publish a paper on it today, would this be considered sound science justifying a paper? Of course not. First we would want to wait to see if the forecast was skillful. This is possible with weather forecasts for tomorrow, but we cannot yet verify a climate forecast model’s skill, for decadal-averaged weather conditions decades into the future.

The climate modeling community runs ensembles of multi-decadal predictions (with different initial conditions, different models) and they average their results over decadal time periods, which they claim distinguishes their simulations from the numerical weather prediction community’s application. Of course the numerical weather prediction community also runs ensembles of simulations. The fundamental difference is that the weather community can validate their model results thousands of times. There is no such ability with multi-decadal climate prediction models.

Our conclusions are the following:

  1. Peer-reviewed papers, and national and international assessments, which present model results for decades into the future, or provide impact studies in response to these model simulations, should never be interpreted as skillful forecasts (or skillful projections). They should be interpreted as process (sensitivity) studies, even though the authors use definitive words (such as this “will” occur) and display model output with specific time periods in the future.
  2. The US National Assessment, which provided model simulations on regional scales for the coming decades, is inaccurately portrayed when their results are given to stakeholders with the interpretation that their results bracket what is expected in the future. This is misleading when transmitted to policymakers, as process studies are inappropriately interpreted to be forecasts.
  3. Climate forecasts (projections) decades into the future have not demonstrated skill in forecasting local, regional, and global climate variables. They have shown that human climate forcing has the capacity to alter the climate system, but we should not present these model simulations as forecasts. To present them as forecasts is misleading to policymakers and others who use this information.

Comments Off

Filed under Climate Models

Comments are closed.