Climate Model Hindcasting
Compare the predictions of global temperature made by several different climate models to the actual obserations.
See the for more information about the model you've chosen.
The dashed line marks the year that the model was published. The model's prediction after this time is called a "forecast". The model's "prediction" for earlier times is called a "hindcast".
The "forecast" label is from the model's point of view. Looking back on these predictions today, they are really all "hindcasts" now, so we can see how good their predictions really were!
You may notice that this learning resource does not include major recent climate model projections. That's just because they're so new, there hasn't been enough observation time to compare them to. See the for more details.
All models make assumptions. The most important assumptions for these models are shown in the controls below the graph. You can adjust these assumptions to see how much difference they make to the model's prediction.
Hausfather, Z. (2017) Analysis: How well have climate models projected global warming? CarbonBrief website, accessed August 2021.
Hausfather, Z., Drake, H. F., Abbott, T., & Schmidt, G. A. (2020). Evaluating the performance of past climate model projections. Geophysical Research Letters, 47, e2019GL085378. https://doi.org/10.1029/2019GL085378
Resolution: ° by °.none (1D model)
What is "hindcasting"?
One important way to check how well a model did is to have it "predict" things that we already know. This is also an important part of the process of making a model. Climate model predictions are also called "forecasts", so this sort of check is sometimes called "hindcasting".
In order to evaluate something, we need to decide what it's for so we know what to focus on. Climate models like these are predicting how global patterns will change over several decades, depending on our behaviour and other factors. In the process, many climate models predict changes from year to year (and even from day to day!). These "short term" changes are an important part of how the model works, but modelers are really only interested in the "long term" results: the averages and trends over decades. This works because those small, short-term variations naturally "average out" over longer times.
So we need to squint a bit when we're evaluating these models. A model doesn't need to get all of the wiggles right (and many of them give smooth lines!). Instead, we look at things like what the slope would be if we drew a smooth line through all the wiggles, for example, or how much warmer the Earth is in the 2000s than it was in the 1970s.
This is also why more recent models are not included in this learning resource. In order to evaluate a model's predictions like this, we need a decade or more of observations to compare to, since that's the time scale the models are really designed for.
It's also important to remember that how the climate changes depends on our own behaviour, since we're the ones putting most of that CO2 and other greenhouse gases into the air. With some models, researchers were able to make predictions under several different scenarios, representing different assumptions about our future behaviour. At other times, researchers did not have the resources for several predictions, so they had to based their assumptions on what seemed to be our most likely path. Always think of climate model predictions as "if we do this, then that will happen."
In this learning resource, you can see roughly how much these assumptions matter to the predictions. Changing the slider for CO2 concentration will bend the forecast curve up or down appropriately. This usually won't be exactly the same as you'd see if you actually re-ran the original models, but it's quite a good approximation.
How do we measure the Earth's temperature?
When we talk about "the Earth's temperature" we're referring to the mean global temperature at the Earth's surface. Land surface air temperatures are measured by thermometers at weather stations around the world. Sea surface temperatures are measured by thermometers in buoys scattered around the oceans. All of these measurements are combined into a single mean surface temperature for the globe. Different sources (listed below) have different methods for combining the measurements, but the results are all quite similar.
On land, we measure the temperature of the air, but on the ocean we measure the temperature of the water. Until recently, climate models would use air temperatures over both land and sea to calculate their mean, which introduced a slight bias in the output since the ocean tends to be cooler on average. This only makes a difference of around 0.01°C, too small to matter for the model comparisons we're making here.
Annual global mean surface temperature data were from NASA GISTemp, NOAA GlobalTemp, Hadley/UEA HadCRUT4, Berkeley Earth, and Cowtan and Way, as provided in the supplemental material of Hausfather et al. (2020). The "Observed" line in the graph shows the mean of the five measurements. For any given year, the sources generally agreed to within 0.01°C.