Guest Post by Willis Eschenbach
The IPCC, that charming bunch of United Nations intergovernmental bureaucrats masquerading as a scientific organization, views the world of climate models as a democracy. It seems that as long as your model is big enough, they will include your model in their confabulations. This has always seemed strange to me, that they don’t even have the most simple of tests to weed out the losers.
Through the good offices of Nic Lewis and Piers Forster, who have my thanks, I’ve gotten a set of 20 matched model forcing inputs and corresponding surface temperature outputs, as used by the IPCC. These are the individual models whose average I discussed in my post called Model Climate Sensitivity Calculated Directly From Model Results. I thought I’d investigate the temperatures first, and compare the model results to the HadCRUT and other observational surface temperature datasets. I start by comparing the datasets themselves. One of my favorite tools for comparing datasets is the “violin plot”. Figure 1 show a violin plot of a random (Gaussian normal) dataset.
You can see that the “violin” shape, the orange area, is composed of two familiar “bell curves” placed vertically back-to-back. In the middle there is a “boxplot”, which is the box with the whiskers extending out top and bottom. In a boxplot, half of the data points have a value in the range between the top and the bottom of the box. The “whiskers” extending above and below the box are of the same height as the box, a distance known as the “interquartile range” because it runs from the first to the last quarter of the data. The heavy black line shows, not the mean (average) of the data, but the median of the data. The median is the value in the middle of the dataset if you sort the dataset by size. As a result, it is less affected by outliers than is the average (mean) of the same dataset.
So in short, a violin plot is a pair of mirror-image density plots showing how the data is distributed, overlaid with a boxplot. With that as prologue, let’s see what violin plots can show us about the global surface temperature outputs of the twenty climate models.
For me, one of the important metrics of any dataset is the “first difference”. This is the change in the measured value from one measurement to the next. In an annual dataset such as the model temperature outputs, the first difference of the dataset is a new dataset that shows the annual CHANGE in temperature. In other words, how much warmer or cooler is a given year’s temperature compared to that of the previous year? In the real world and in the models, do we see big changes, or small changes?
This change in some value is often abbreviated by the symbol delta,”∆”, which means the difference in some measurement compared to the previous value. For example, the change in temperature would be called “∆T”.
So let’s begin by looking at the first differences of the modeled temperatures, ∆T. Figure 2 shows a violin plot of the first difference ∆T of each of the 20 model datasets, as numbers 1:20, plus the HadCRUT and random normal datasets.
Figure 2. Violin plots of 20 climate models (tan), plus the HadCRUT observational dataset (red), and a normal gaussian dataset (orange) for comparison. Horizontal dotted lines in each case show the total range of the HadCRUT observational dataset. Click any graphic to embiggen.
Well … the first thing we can say is that we are looking at very, very different distributions here. I mean, look at GDFL  and GISS , as compared with the observations …
Now, what do these differences between say GDFL and GISS mean when we look at a timeline of their modeled temperatures? Figure 3 shows a look at the two datasets, GDFL and GISS, along with my emulation of each result.
Figure 3. Modeled temperatures (dotted gray lines) and emulations of two of the models, GDFL-ESM2M and GISS-E2-R. The emulation method is explained in the first link at the top of the post. Dates of major volcanoes are shown as vertical lines.
The difference between the two model outputs is quite visible. There is little year-to-year variation in the GISS results, half or less than what we see in the real world. On the other hand, there very large year-to-year variation in the the GDFL results, up to twice the size of the largest annual changes ever seen in the observational record …
Now, it’s obvious that the distribution of any given model’s result will not be identical to that of the observations. But how much difference can we expect? To answer that, Figure 4 shows a set of 24 violin plots of random distributions, with the same number of datapoints (140 years of ∆T) as the model outputs.
As you can see, with a small sample size of only 140 data points, we can get a variety of shapes. It’s one of the problems in interpreting results with small datasets, it’s hard to be sure what you’re looking at. However, some things don’t change much. The interquartile distance (the height of the box) does not vary a lot. Nor do the locations of the ends of the whiskers. Now, if you re-examine the GDFL (11) and GISS (12) modeled temperatures (as redisplayed in Figure 5 below for convenience), you can see that they are nothing like any of these examples of normal datasets.
Here’s a couple of final oddities. Figure 5 includes three other observational datasets—the GISS global temperature index (LOTI), and the BEST and CRU land-only datasets.
Here, we can see a curious consequence of the tuning of the models. I’d never seen how much the chosen target affects the results. You see, you get different results depending on what temperature dataset you choose to tune your climate model to … and the GISS model  has obviously been tuned to replicate the GISS temperature record . Looks like they’ve tuned it quite well to match that record, actually. And CSIRO  may have done the same. In any case, they are the only two that have anything like the distinctive shape of the GISS global temperature record.
Finally, the two land-only datasets [23, 24 at lower right of Fig. 5] are fairly similar. However, note the differences between the two global temperature datasets (HadCRUT  and GISS LOTI ), and the two land-only datasets (BEST  and CRUTEM ). Recall that the land both warms and cools much more rapidly than the ocean. So as we would expect, there are larger annual swings in both of those land-only datasets, as is reflected in the size of the boxplot box and the position of the ends of the whiskers.
However, a number of the models (e.g 6, 9, & 11) resemble the land-only datasets much more than they do the global temperature datasets. This would indicate problems with the representation of the ocean in those models.
Conclusions? Well, the maximum year-to-year change in the earth’s temperature over the last 140 years has been 0.3°C, for both rising and falling temperatures.
So should we trust a model whose maximum year-to-year change is twice that, like GFDL ? What is the value of a model whose results are half that of the observations, like GISS  or CSIRO ?
My main conclusion is that at some point we need to get over the idea of climate model democracy, and start heaving overboard those models that are not lifelike, that don’t even vaguely resemble the observations.
My final observation is an odd one. It concerns the curious fact that an ensemble (a fancy term for an average) of climate models generally performs better than any model selected at random. Here’s how I’m coming to understand it.
Suppose you have a bunch of young kids who can’t throw all that well. You paint a target on the side of a barn, and the kids start throwing mudballs at the target.
Now, which one is likely to be closer to the center of the target—the average of all of the kids’ throws, or a randomly picked individual throw?
It seems clear that the average of all of the bad throws will be your better bet. A corollary is that the more throws, the more accurate your average is likely to be. So perhaps this is the justification in the minds of the IPCC folks for the inclusion of models that are quite unlike reality … they are included in the hope that they’ll balance out an equally bad model on the other side.
HOWEVER … there are problems with this assumption. One is that if all or most of the errors are in the same direction, then the average won’t be any better than a random result. In my example, suppose the target is painted high on the barn, and most of the kids miss below the target … the average won’t do any better than a random individual result.
Another problem is that many models share large segments of code, and more importantly they share a range of theoretical (and often unexamined) assumptions that may or may not be true about how the climate operates.
A deeper problem in this case is that the increased accuracy only applies to the hindcasts of the models … and they are already carefully tuned to create those results. Not the “twist the knobs” kind of tuning, of course, but lots and lots of evolutionary tuning. As a result, they are all pretty good at hindcasting the past temperature variations, and the average is even better at hindcasting … it’s that dang forecasting that is always the problem.
Or as the US stock brokerage ads are required to say, “Past performance is no guarantee of future success”. No matter how well an individual model or group of models can hindcast the past, it means absolutely nothing about their ability to forecast the future.
Best to all,
DATA SOURCE: The model temperature data is from the study entitled Evaluating adjusted forcing and model spread for historical and future scenarios in the CMIP5 generation of climate models, by Forster, P. M., T. Andrews, P. Good, J. M. Gregory, L. S. Jackson, and M. Zelinka, 2013, Journal of Geophysical Research, 118, 1139–1150, provided courtesy of Piers Forster. Available as submitted here, and worth reading.