Climate Models Fail to Match Observed Historical Data

A key consideration for placing confidence in climate models is the extent to which they mimic actual climate. Surprisingly few analyses of this critical factor have been performed. In this study, 55 weather stations around the globe are analyzed for various monthly and longer-term aspects of temperature and precipitation, and the contiguous United States (70 stations using a gridded Thiessen polygon mean) is analyzed for long-term large-scale correspondence. The stations used were chosen only if they had at least a 100-year record and few missing values. Six climate models were used for comparison. For evaluating single stations to GCM outputs, the 4 nearest GCM grid cell model outputs to that station were combined using a linear weighting regression that gave maximum benefit of the doubt to nearby similarity of model outputs to real data. So what did the authors find?

For individual stations, at the monthly time scale the models reproduced the seasonal fluctuations in temperature (correlation 0.91) and precipitation (correlation 0.26) as well as latitudinal gradients pretty well. However, for maximum and minimum monthly temperature and precipitation, and annual values as well, the correlations fall below 0.1. So what does this all mean?

It has been argued that GCMs make effective large scale and long-term predictions, although the authors note that no proof of this has ever been offered. To address this issue, they compared the model output with the 30 year moving average temperature and precipitation values for the continental United States. With respect to precipitation, they found that the temporal pattern of the 30 year mean does not match very well. Precipitation is much higher in the models at the continental scale at seasonal and annual resolution. The annual total is near 700mm but the models go over 950mm (250mm or 36% higher). When a model is this far over-stating the precipitation, any slight change in the model forecast will produce an inordinate change in extreme values. Likewise, no prediction of flood frequency distributions can be made with such outputs.

For temperature, the mid-century peak was missing and the models showed an acceleration in warming in recent decades for mean annual temperatures, as well as for monthly maximum and minimum temperatures, that is not found in the data. Additionally, the models are all warmer than the observations (by up to 4°C) for mean annual and minimum monthly temperatures. According to the authors, this large offset indicates a model problem with radiative physics because longwave radiation is proportional to the fourth-power of the temperature of the surface, so this large an offset is not realistic.

Based on the above findings, the authors conclude that in "examining the local performance of the models at 55 points, we found that local projections do not correlate well with observed measurements. Furthermore, we found that the correlation at a large spatial scale, i.e. the contiguous USA, is worse than at the local scale." These results do not support the contention that the model simulations become more accurate as one goes to larger spatial scales and longer temporal averages, a key assertion of climate modeling proponents. Furthermore, if used for evaluating biotic response (e.g., extinction risk), the unrealistic temperature and precipitation behaviors and absolute levels will produce unrealistic forecasts of impact.