The weather forecasts and seasonal forecasts have been verified – many "forecasts" for the past have been compared to what really happened. One of the key properties of a good forecast is reliability. A forecast system is reliable if it rains on 60% of all the days forecasted with a "60% chance of rain".

In our study we checked the reliability of regional climate-model trends. Because we only have a single trend per grid point, we collected the trends of all grid points into a verification graph, the rank histogram.

We discovered that the recent CMIP5 climate-model ensemble is not reliable but overconfident – the trends are more often in the top 5% and bottom 5% of the ensemble than expected by chance. The ensemble seems reliable for temperature but this is caused by variations in the rate at which the global mean temperature rises, not by the correct spatial pattern. The maps show the locations of the trends in the tails of the ensemble. Some of this is due to the natural fluctuations of weather and climate, but a large part must also be due to other causes.

The reasons the trends were not simulated correctly up to now may be that the natural variability is underestimated, that the input of the climate models was not correct (for instance as a result of factors such as aerosols or land use), or that the models do not correctly represent some processes. Further research is needed to improve the forcings and climate models and so reduce these trend biases. In the meantime we have to accept that we cannot make reliable climate forecasts, and to present the useful information in climate-model outputs in other ways, such as climate-change scenarios.

The team reported the study in Environmental Research Letters (ERL).

Related links