Forecasting confidence decreases over time
If we are not certain about the weather next week, how can we be confident about climate change predictions 50 or 100 years into the future? The charts below show two predictions about the temperature in Europe.. one was forecast for a day in June 2015, some 198 hours ahead, the other is a prediction for European July temperature by 2100, some 876,000 hours ahead. If computer model forecasts beyond 120 hours are known to become significantly more unreliable pushed further into the near future, how confident can we be in similar computer models predicting atmospheric conditions some 876,000 hours into the distant future?
Two temperature anomaly predictions for 198 hours and 876,000 hours ahead.
Of course, weather and climate are different things: weather is short term atmospheric conditions over days and weeks and is usually forecast at high resolution over small areas, while climate describes long term weather over decades or more for larger regions or the whole planet. Nevertheless, much like the computer models that broadcast meteorologists use to issue short range weather forecasts, climate models use equations of fluid motion and thermodynamics to determine the behaviour of the atmosphere and ocean and to project predictions of Earth’s climate into the future. Both short range and long range predictions require powerful super-computers to run similar complex models that ingest millions of real time observations and perform trillions of calculations to produce predictions of atmospheric conditions on a huge three dimensional grid across different surfaces and altitudes, including the oceans.
three dimensional NWP grid
global warming prediction
The atmospheric system is essentially unpredictable. Even the most powerful super-computer short range weather predictions can become quite unreliable beyond about 120 hours. This is perhaps why the MetOffice still only widely publish short term forecasts out to 5 days. It’s also why charts beyond 300hours are sometimes called “Fantasy Island”… they are so unreliable and not to be taken seriously as a forecasting tool on their own.
MetOffice 5 days is your lot
beyond 5 days.. probably wrong
Fantasy Island..very unreliable forecasts beyond 300hours
In the medium range, accurate ten day forecasts are still something of a holy grail and longer range seasonal forecasts arguably remain largely experimental as they are based on the unreliable time-frame of most numerical model output and the application of complex teleconnections and /or the extrapolation of observations from historic analogue patterns. The MetOffice have even put their seasonal long range predictions into obscure parts of their website after notable public failures in past seasonal forecasts, where they still languish today despite huge investment in computer power. The charts below show examples of MetOffice contingency planner long range forecast information.
Mean 2 metre temperature anomaly for Jan/Feb/Mar
Mean 500mb height anomaly Nov,Dec and Jan
3 month outlook… experts only
Even with a “perfect model”, weather and climate prediction will always suffer from uncertainties due to the immense complexity of the climate system, the chaotic nature of the atmosphere and the simplifications and approximations necessarily built into models themselves. The ensemble charts below illustrate this growing uncertainty over just a short time scale of a few weeks. Note the increasingly wide range of possible outcomes from the individual members showing the growing uncertainty as time progresses, and this is in a relatively settled period of weather.
ensemble temperature forecast
ensemble wind speed forecast
So, if we really cannot be not certain about the weather next week, how can we be confident about related computer model predictions of climate for 50 or 100 years ahead? To answer this we need to outline the three different types of uncertainty over climate change prediction:
There are three main types of uncertainty over climate change predictions:
- Model uncertainty: climate models have to approximate and estimate feedbacks and processes, they do this slightly differently.
- Internal variability: weather is chaotic partly due to uncertain internal forcings, like volcanic eruptions.
- Scenario uncertainty: future estimates of human behaviour and emissions of greenhouse gases in particular are uncertain
So, do these uncertainties increase over longer time scales thus rendering computer model predictions of climate in 100 years completely unreliable? The answer is “No”! Or more precisely, most of these uncertainties actually decrease over time making model predictions of climate in 100 years possibly more reliable than a seasonal forecast for next summer! Time scale and geographic scale are two reasons that can explain why this happens:
- Time scale: climate models deal with longer timescales better than short. The chart below shows lines indicating different model output for temperature change between 1850 and 2010. Note the various models (shown as different lines) all reaching a similar overall temperature increase of +0.8C by 2010.
temperature change more certain over longer time scale
Over the long timescale shown above, the model runs all agree on an overall temperature rise of +0.8C due to initial changes in radiative forcing linked to combinations of, for example, increases in emissions of greenhouse gases, changes in solar radiation and frequency of volcanic eruptions etc. Models confidently handle these changes over long time scales because climate system and model uncertainty both decrease over time..
Now… if we zoom into one part of the chart above the predictive capability of the model is shown to be much more questionable as the lines wiggle about much more over shorter timescales and sometimes go in completely opposing directions. This shows uncertainty increasing over shorter timescales.
less accurate predictions over a short time scale
On the shorter timescale of decades or less, model uncertainty increases as climate variability over short time scales is naturally large from one year to the next i.e. one year can be colder or warmer than the previous due to internal variability of the climate system (i.e.chaos). Models don’t handle this small scale short term climate chaos very well! One model run predicts cooling in a particular decade, while another predicts warming for the same decade. It turns out that predicting climate change over smaller timescales is more unpredictable than predicting changes over broad sweeps of time! This is because of the internal chaotic nature of the climate system and model uncertainty being greater at this higher resolution. For example, models will handle these short term variables differently: How will ocean heat uptake respond? What will happen to ocean currents and regional climates? How will snow cover respond? Will volcanoes erupt? How will cloud cover and type change? … and many more besides.
climate model variables
Despite these transient climatic uncertainties over short time scales the overall direction of change towards a new climatic equilibrium in the long term is confidently predicted by all the models. So… longer timescales are better handled by models than short.
Scenario uncertainties increase over time in long range climate prediction models
The exception to this reduction in uncertainty over time is scenario uncertainty. Most long range climate model charts include a wide range of predictions. This wide range is not due to model uncertainty or uncertainties over internal climate variability. Scenario uncertainty is the uncertainty over predicting future emissions of greenhouse gases. In other words, uncertainty over our own human behaviour in the future, which is historically difficult to predict! This is the reason why the IPCC charts show several “RCP trends”. Representative concentration pathways (or emissions scenarios) show a range of human response to the climate crisis… ranging from drastic curbs to emissions and lower growth, through business as usual, to increased emission pathways presumably due to high growth with no curbs to greenhouse gas emissions. These all yield obviously different outcomes. The uncertainty over human response, known as the scenario uncertainty, increases over time. Despite scenario uncertainty, each RCP is an accurately modelled temperature change based on changes in radiative forcing, the wide range of results is due to our own unpredictable behaviour more than climate chaos or uncertainties in the models.
2. Geographic coverage
On a long term global scale, climate system uncertainties are reduced while uncertainties increase over how climate change will occur over small geographic regions. Predicted global mean annual temperature change is therefore more certain than, for example, regional European or UK temperature change in 100 years. Future climate change on regional scales is more uncertain than on a global scale due to small scale variabilities.
So… prediction of climate change over short time scales and regional scales is more uncertain than predicting longer range changes on a global scale. Models struggle to resolve climate change at small time scales and small geographic scales. Overall however, longer term climate outcomes are more certain than short term small scale weather action. Even with the immense complexity of the climate system, computer models can provide confident predictions of future climate within a range of scenarios. The IPCC quote levels of confidence and certainty for their climate predictions and, even for the end of the century, they claim high confidence in their predictions ranging from “likely” to “more likely than not” for various scenarios.
IPCC AR5 report on confidence in computer model temperature predictions
the range of climate predictions will hit the probable climatological outcome
range of outcomes can be predicted fairly confidently
but the detailed action is harder to forecast!
In conclusion, maybe this analogy might help: Perhaps climate prediction is a bit like a football match: complex uncertainties about precisely where the ball will go during the game are perhaps completely impossible to forecast beyond the first few seconds of the game. Any detailed minute-by-minute action on the field thereafter becomes increasingly uncertain over time as countless variables come into play, including some chaos. For example, individual player performance on the day, how the players and teams interact, the chaotic nature of the ball, the nature of the pitch… these are internal variables that make it almost impossible to model accurate step by step action of a game perhaps beyond the first few kicks after the whistle blows. This is why small scale high resolution weather forecasts are still limited to less than a week ahead. Nevertheless, despite such short term uncertainty the overall score and outcome of the game is still possible to predict with confidence, particularly within a certain range. The same can be said for long range climate models.
This post has only scratched the surface of climate and weather uncertainty. Please let me know of any errors you spot in this post. Further reading available here: