When Bias Correction Goes Wrong
Climate model output is often bias corrected to better match observations. These techniques transform the output so that the model and observed distributions agree. However, the authors demonstrate a range of cases where the distributions agree, but other characteristics of output are wildly different. These situations arise from applying bias correction to climate models with fundamental deficiencies in simulating the variable of interest.
This study raises awareness of the limitations of bias correction. Users of climate model data will better understand the reasons why bias correction techniques are inadequate for correcting climate models with fundamental deficiencies simulating physical processes.
Global climate models (GCMs) are typically bias corrected to better match observations. However, if a GCM does not accurately simulate the underlying physical processes, then common bias correction techniques may not be able to fix the problem. The authors found many examples where bias correction is unable to adequately correct process-related issues with GCMs. For instance, in Peru, tremendous amounts of precipitation fall every few years, typically during El Niño events. However, some GCMs simulate very regular precipitation with similar amounts year-to-year. Common bias correction techniques can get the precipitation distribution to match observations, yet the connection to El Niño—a crucial aspect of these events—is missed completely. This study raises awareness of the limitations of bias correcting GCMs. Bias correction alone may be unable to repair models that inadequately simulate the underlying physical processes. While some GCM cases are irredeemable and should be discarded, others could be remedied with the adoption of advanced bias correction techniques, which use stochastic generators or process-oriented approaches.