Recent media headlines have again discussed the issue of whether climate models are overly sensitive to greenhouse gases. These headlines have misinterpreted a study by Millar et al. which was discussing carbon budgets to limit global temperature rise to 1.5°C above pre-industrial levels.
A recent study by Medhaug et al. analysed the issue of how the models have performed against recent observations at length and largely reconciled the issue. An overly simplistic comparison of simulated global temperatures and observations might suggest that the models were warming too much, but this would be wrong for a number of reasons.
In the Medhaug et al. paper they show the range of models (blue shading in figure with median in light blue), compared with the HadCRUT4 observations and their estimated uncertainty (orange shading with light orange line). There are a number of well understood reasons why the light orange line might not follow the light blue line, namely: radiative forcings, variability, observational biases and choice of reference period.
Radiative forcings: The simulations were produced using observed information on sources of radiative forcing up to 2005, and made various assumptions for subsequent forcings. For example, the simulations assumed no volcanic eruptions after 2005, whereas the real world did have some eruptions. In addition, the sun dimmed slightly and this was not included. Retrospectively we can estimate the effects of these assumptions on the simulations, and this moves the light blue line to the mid-blue line. In other words, if the models had known about future forcings they would have been closer to the observations.
Variability: It is also understood that natural fluctuations in the climate (e.g. ENSO, PDO) can temporarily offset or enhance the warming during certain periods. These effects can also be accounted for, producing the dark blue line. In other words, if the models had produced the same phase of variability as the real world then, again, they would have been closer to the observations.
Observational biases: We also understand that our observations are not perfect. The HadCRUT4 dataset has relatively few observations over the Arctic and also uses sea-surface temperatures over the ocean, whereas the model analysis uses simulated air temperatures everywhere. Accounting for these issues moves the observations warmer, to the dark orange line. In other words, in an ‘apples-with-apples’ comparison, the observations and models are closer together.
When accounting for these three factors together, the dark blue and dark orange lines now show a very similar warming trend – the models and observations have been reconciled and there is no clear evidence from the recent period that the models are therefore ‘running too hot’. About 1/3 of the apparent discrepancies are due to each of these three factors.
Choice of baseline: One further subtlety is the choice of ‘baseline’. Medhaug et al used a 1961-90 reference period, whereas the IPCC AR5 chose 1986-2005. This difference can also slightly move the relative position of the observations within the model spread higher or lower. There is no perfect choice.