Robust skill of decadal climate predictions


D. M. Smith, R. Eade, A. A. Scaife, L.-P. Caron, G. Danabasoglu, T. M. DelSole, T. Delworth, F. J. Doblas-Reyes, N. J. Dunstone, L. Hermanson, V. Kharin, M. Kimoto, W. J. Merryfield, T. Mochizuki, W. A. Müller, H. Pohlmann, S. Yeager, X. Yang

npj Climate and Atmospheric Science
Work Package 1



Predictions of the climate up to a decade ahead are in great demand but have been held back by problems around models producing small climate signals in relation to noise, the so-called signal-to-noise paradox. This study uses a much larger ensemble of model simulations than previous work to assess this problem, revealing that a large ensemble can overcome the signal-to-noise paradox and produce confident predictions of temperature, rainfall and atmospheric pressure on a decadal scale. The benefit of initialising models with observations is also demonstrated more clearly than in previous studies, though most of the prediction skill arises from external factors rather than internal variability. The results of this study help us make better predictions of how the climate might change in the coming years, allowing people and policymakers to adopt more effective measures and policies to adapt to and mitigate the impact of future climate change.

There is a growing need for climate predictions up to a decade into the future as the climate continues to change. Although these decadal predictions can determine surface temperature well, there is much less confidence in their predictions of rainfall and atmospheric circulation. Recent studies have shown this problem is at least partly due to the climate signal in the model results being too small in relation to the random noise inherent in the climate system; this is known as the signal-to-noise paradox. This problem can be addressed using a large ensemble of model simulations from which the climate signal can be picked out, even if the individual simulations do not contain large signals. This study looks at this issue, using a larger model ensemble and assessing other ways to improve our confidence in the model results.

The team found that the signal-to-noise paradox is a widespread problem in decadal predictions, highlighting problems in the way the models simulate future climate. The team also found, though, that using a large enough ensemble of models goes some way to overcoming the individual models’ deficiencies. They were able to produce confident predictions of rainfall and atmospheric circulation, as well as the surface temperature predictions the models were already capable of, by using a much larger ensemble than had been used in previous studies. This means that, although our climate models individually have serious problems predicting the climate at the decadal scale, a large enough group of them can still produce useful predictions. The team also proposed a new method to assess the impact of initialising the models, that is running them over a period in the recent past that can be compared with observations, then allowing them to run into the future. Initialisation was found to provide significant benefits in certain predictions, such as temperature over Europe and atmospheric pressure over the Atlantic. However, the overall patterns of prediction skill were not affected by initialisation suggesting that decadal climate is mainly driven by external factors rather than internal variability. Understanding the reasons behind this could allow us to further improve our models.

This study used climate models from seven leading forecast centres around the world. Each of these models was run multiple times with slightly different conditions, producing a final ensemble of 71 climate simulations. Retrospective forecasts starting each year from 1960 to 2005 were checked against observed data to test the models. This is the same protocol used for the CMIP5 project. To assess the impact of initialisation these predictions were compared with an uninitialized ensemble including 56 simulations from the same seven models.

Policy relevance
Credible climate predictions on the decadal scale are crucial for adapting to climate variability and change in the coming years. This study’s demonstration of our ability to produce unexpectedly useful predictions from existing models will help people and authorities take more effective measures and develop more effective policies to adapt to and mitigate the impacts of climate change. These results also contribute to the WCRP Grand Challenge on Near Term Climate Prediction and ongoing World Meteorological Organisation annual to decadal prediction activities.


There is a growing need for skilful predictions of climate up to a decade ahead. Decadal climate predictions show high skill for surface temperature, but confidence in forecasts of precipitation and atmospheric circulation is much lower. Recent advances in seasonal and annual prediction show that the signal-to-noise ratio can be too small in climate models, requiring a very large ensemble to extract the predictable signal. Here, we reassess decadal prediction skill using a much larger ensemble than previously available, and reveal significant skill for precipitation over land and atmospheric circulation, in addition to surface temperature. We further propose a more powerful approach than used previously to evaluate the benefit of initialisation with observations, improving our understanding of the sources of skill. Our results show that decadal climate is more predictable than previously thought and will aid society to prepare for, and adapt to, ongoing climate variability and change.