In comments on a previous post, JohnV asked how trends based on the AR4 model realizations would compare to observations if we selected a start date of 1990. This is a fair question, but I told him I would defer answering for a little while.
Here’s a portion of JohnV’s comment (click to read the entire thing.)
So, if going back to 1990 is a good thing and using trends is a good thing, then using trends from 1990 must be a good thing. The observed trends from 1990 to 2007 are:
GISTEMP: 2.16 C/century
HadCRUT3: 2.04 C/centuryAnd from 1990 to 2008 (preliminary numbers) are:
GISTEMP: 1.87 C/century
HadCRUT3: 1.74 C/century
Though JohnV didn’t say this directly, I suspect one of his points is to suggest these are close to the “about 2C/century” projected to apply to the early portions of this century.
They are close. But… unfortunately, there’s a problem. The 2C/century is not the trend models predict if one happens to compute the trend using the start date of 1990. During the period from 1980 to 2000 GMST is expected to exhibit a large amount of non-linearlity due to something other than “internal variability”: Volcanoes erupted.
I’ll demonstrate the issue qualitatively. (That is to say, the quantitative aspects will be crude.) To discuss the issue I will:
- Use all realizations continuing with the A1B SRES provided at The Climate Explorer. Most come pre-assembled, as permitted by Geert Jan script. One has an issue, and I assembled it myself. Also, two of these are not included in the AR4.
- Show the graph “pinning” all realizations using a reference of Dec. 1989 to Dec. 1990 average. The only purpose is to show how things happen to look using that baseline. (For discussing the AR4, I prefer to use the average temperature from Jan. 1980-Dec. 1999.)
- Fit trends to the 12 month averages GMST — computed monthly. (This has weird effects because the some monthly values contribute more than others. But, since I’m showing a graph, it’s convenient. My values for trends will differ from JohnV’s primarily because of the 12 month averaging. I don’t know if John’s are based on monthly averages, annual averages etc.)
Graphical Comparison “pinned” at 1990
Trends
If you examine the trends in the graph, you will see that using the method described above the trend for GISSTemp (outlined in red) was about 2.1 C/century. HadCrut’s was about 1.9 C/century.
These are very close to the nominal value of about 2 C/century the IPCC AR4 projection for underlying trend for the current period.
However, in the AR4, the authors did not happen to provide specific numerical values for projection referenced to a start date of 1990. The projections are stated relative to the average from Jan 1980-Dec. 1999. If we dig up the models and concoct a hindcast by averaging over all model runs, the OLS trend from 1990 to now using that average is about 2.5 C/century.
This exceeds the actual trend by 0.5C/century or 20%. If we focus on the average trend based on runs that include volcanic aerosols, the projected trend is 2.8 C/century, which differs even more noticeably from the observed trend.
Are these differences statistically significant? I’m not going to discuss that today! ( The answer appears to depend on the statistical model used, the confidence intervals, the metric chosen, whether we focus on volcano only or not. I may eventually go into more detail, but I’m checking a few things about the model data I downloaded.)
Whether significant or not, if we selected 1990 as a start date, the trends predicted by models do over-shoot the trends we actually observed.
The whole premise of this post seems flawed. Why would we be surprised to see model runs accurately predicting observed temperatures that occurred before the model runs were published?
AR4 predictions were published in 2007 and probably not finalized more than a year or two before that.
Of course the model runs are in rough agreement with observed temperatures from 1990 until the numbers were finalized.
Does anybody believe that the algorithms and data inputs that produced the AR4 model runs were uninfluenced by observations made post-1990?
Any climate modeler who cares about his craft is going to make sure that his model is at least somewhat consistent with observed temperatures of the preceding 17 years.
Even if we imagine a world in which all model runs were immutably based on principles discovered prior to 1990 (in reality, this is a provably false assertion) what would happen to those algorithms and inputs that produced results which were obviously NOT consistent with the past 17 years of observations? Would the researchers who created those models continue using them? Would the results be submitted to journals and the IPCC? Would the IPCC have chosen to publish them? Is not the most likely outcome that any model runs which are inconsistent with observed temperatures never get published by the IPCC?
Why then would we be surprised to see numbers published in 2007 agree with temperatures over the preceding 17 years? Of course they do.
The interesting question is whether or not these model runs will be consistent with observations made AFTER they were published.
Of course, by the time we can make such an assessment, AR5 will be out.
Who thinks that the model runs in AR5 will be inconsistent with observed temperatures between 1990 and one year before the report is released?
OT but I did not want to load the troposphere thread. Have you seen this:
http://pangea.stanford.edu/research/Oceans/GES205/Barnett_Science_Penetration%20of%20human%20warming%20into%20ocean.pdf
Fingerprints all over the place.
on topic: trends in the form of lines assume that the functional dependence is linear. Nothing in the shape of the temperature graphs, except the gradual rise from last century points to linearity. A best curve fit would certainly not show a line.
Re Jason: “Why would we be surprised to see model runs accurately predicting observed temperatures that occurred before the model runs were published?”
That’s exactly what I would expect, unless the models were bogus.
Anna–
The average of the model runs that include volcanic eruptions is definitely not linear.
Bob and Jason: Whether or not the models agree with the observations depends on your definition of agree. I haven’t written a detailed post, but those claiming “agree” should note that the difference between 2.8 C/century (predicted by “volcano” runs) and 2.0 c/century is not exactly tiny. If the IPCC best estimate of the projection changed by ±0.8 C/century, this would be considered a rather large change in the projected value. Even ±0.5C/century is not a small amount.
anna v (Comment#7736) December 27th, 2008 at 5:36 am
The only finger print that paper shows is, the amount of solar energy reaching the earths surface (ocean or land) has a greater influence on the climate than does CO2. How they can conclude that CO2 has a larger forcing magnitude over the oceans than it does land. It’s contrary to even the most basic laboratory experiments. Using the same wattage over the same surface area. Try heating a black bowl of water with a hair dryer and then with a solar lamp. Let me know which one of the two has the greater heat content per volume of liquid when calculated against the original liquid volume.
You ask a very naive question. To answer whether the observed trend is linear, you either have to have a lot more data, or data with a lot less annual variation, or a model that predicts the form of the observed trend.
Can you, for example, show that a quadratic, or exponential, or whatever trend is statistically preferable for this set of data. Frankly I doubt it, at least if you do the statistics correctly.
Eli–
I am of the opinion what we might call the deterministic component of the trend is non-linear during this period. That is: if we consider the volcanic forcing to be exogenous or known, and and things like ENSO or the PDO to be randome, then the deterministic component will have a dip after the eruption of Pinatubo.
As it happens, those models that are forced with volcanic forcings predict a non-linear trend during this period. The green line represents the sum over quite a few model realizations and show a pronounced dip after Pinatubo. There is even a phenomenological basis for suggesting that volcanic eruptions cause the dip.
I’m not sure what question you think is being asked, or why that question would be naive. JohnV asked a question. He wanted to know how the OLS trends would fits if we happened to fit them. No one has suggested the true underlying trend is linear– only that one can compute a trend. The trend is, in some sense, an observable– this is true whether or not the average over an infinite number of model runs converges to a linear trend during this period.
If we treat the OLS trend as an observable and compute it based on the data, and the models, the trend for the models is larger than for the data. You can make what you wish of that, but it’s the way things happen to shake out.
Actually, isn’t the trend line off far more than even that?
The plot starting at 1990 contains two major events which tend to tip the trend line in a far more positive direction: (1) Mt. Pinatubo, and (2) the 1998 El Nino.
In this case, the order of the events matters. Mt. Pinatubo fires first, dramatically lowering the left side of the graph for 3-4 years (I thought I remember seeing that Gavin admits that the effects of Pinatubo lasted for years). Next, the biggest El Nino event fires next. This dramatically raises the right end of the graph. The combined effect is to greatly rotate the trend line.
This is not trivial. The timing of Mt. Pinatubo matters in this case. As proof, one only need reverse the timing of the events – put the values from 1991-1994 at the end for 2005-2008 – and then see what your trend line looks like. Here’s a hint, it’s not positive.
What’s even worse, 1991 was considered a “strong” El Nino year, and 1992 was just an ordinary El Nino. Simply “remove” Mt. Pinatubo as if it never happened, and replace the anomaly values for 1991-1994 with values representing a dissipating El Nino event (Gavin says that El Nino gives values ranging from 0.1-0.2C). Now, instead of a large dip from 1991-1994, you have a peak showing the El Nino. Guess what, the trend line is barely discernible.
In short, Mt. Pinatubo, almost by itself, greatly affects the whole trend line just by the timing of the event. Take it out, and there is no real warming trend. Or, it is very small.
Furthermore, consider that the 1998 El Nino was so big, and adjust that downward for a “normal” El Nino, and again, the trend line is diminshed.
So, what are we left with … the timing of two events since 1990 greatly affects the look of the trend line. Factor the volcano out, and reduce the effect of the 1998 El Nino … and what trend line is there? Answer: hardly anything.
I eyeballed the values, made some adjustments, and came up with something approximating a trend of 0.03C per decade since 1990.
My $0.02.
Brian–These are just the specific values I get if I download data, create 12 month averages and fit a line. I’m not sure how exactly you remove Pinatubo from the observed data. But, I think you’d still get a warming trend if you just drop out a few years of low temperatures after the eruption.
I don’t think you can just mentally drop out Pinatubo, swap years, erase the El Nino, find a new trend and then decree away warming. The temperature did rise since 1990, since 1970, and also since 1900. There has been warming. You may be able to argue about the cause, but it has been warming.
I’m not saying that there’s not been some amount of warming. But the point is that you can remove the known signals from the noise to reveal an underlying trend line that is not nearly as great as Gavin has claimed, or that appears to be somewhat slowing. Like I said, make some adjustments for Mt. Pinatubo and the 1998 El Nino, and all of a sudden, the trend line since 1990 is not near what is claimed (0.2C/decade).
What I’m pointing out is that the timing of these events does matter if you reset at 1990 and greatly affects the trend line in a positive fashion, and some compensation is needed.
The *timing* of two major climatic events does matter in this case as it tends to skew the trend. No effort is made to factor out these overriding signals, and Gavin wants to claim that the 0.2C/decade trend is still viable, but that doesn’t appear to be the case at all. It would seem to be much less than that over the last two decades.
That’s really all I’m saying – that the trend appears to have slowed substantially (even dramatically) since 1990 when you factor out the two really major biases since that time.
Brian–
If your point is that the least squares trend since Pinatubo is as large as it is in part because of Pinatubo: Yes. It is for the observations. It also is so for the models! So, the theory as coded into the models says the same thing. If you pick a start date of 1990, part of the high trend is due to Pinatubo. Had I picked an end date of 1998, the trend would have looked whopping huge for the observations. But I didn’t use 1998, I used 2008.
Oddly, because 1998 is near the center of 1990-2008, it barely affects the trend at all. (That’s a weird thing about least squares. Outliers at the ends do end up mattering more.)
You remove the El Ninos and the volcanoes by figuring out what their forcings were are running a GCM several times. Even better, you leave the El Nino’s and volcanoes in and run the GCMs. This is called hindcasting and it works. What you will get is an ensemble of forecasts, each of which contains red noise, but what you want to compare them to is not the simple global temperature anomaly, but maps of precipitation and local temperatures and more. From this what you want to figure out is if the differences you get are noise or the result of some physics that you have omitted. And yes, all the GCM groups do this.
In a real sense looking only at the global temperature anomaly is simplistic and trying to tease meaning out of it without modeling is intellectual onanism
Eli-
Refusing to compare the model predictions claimed most robust (i.e. GMST) to the most robust observations (GMST) and insisting we limit comparisons to less robust data and less robust predictions is mental masturbation.
BTW: There are those who insist the weather noise is not red. If the weather noise is red, the models are predicting GMST incorrectly. All the comparisons to precip or accurate hindcasts won’t change this.
First of all none of the data is truly “robust”. There is significant instrumental and system noise. So much in fact that you are left with the issue of how to characterize the noise which even for better data is not a trivial problem. With respect to noise being red, pink or orange, the need for longer records is even more pressing. Ability to measure the frequency spectrum of the noise is, of course, limited by the length of the record. There is also a definitional question as to exactly what is meant by red noise, but the limit would be white (equal noise in each frequency band) or reddish ( decreasing noise with frequency to some power). I don’t think that anyone is proposing blue noise over any period.
Further, it is not clear to me what you mean by saying that instrumental (note that word) global climate surface temperature ANOMALY is considered to be the most robust observation? Is it robust in the sense of telling us something about climate, well that depends on the period you look at the observations over. Over 20-30 years since maybe 1900, yes. Over 1-5 years, no. The internal variability is too high. Since the models also have internal variability, one has the same problem there and of course, it makes it big time silly to compare over short periods. In either case, IGMSTA is only one parameter, and even where you learn something from it, you also have to look at other things, such as the distribution of temperature changes over regions comparable to those modeled, precipitation, circulation patterns and more. The shorter the period you are looking at, the more necessary a multivariate comparison and a pattern matching approach to the data and models. Of course then you have the problem of defining what a useful method of matching is.
None of this is to say that one dimensional models are useless (see 1970 Earth models) but that you learn quantitative things from them. The interesting thing with respect to greenhouse gases is that there has really been little difference in the 2x CO2 global result as we went from pen and paper (Arhennius) to one dimensional computer models, to 3 D models with silly oceans to better ocean coupling and so on. While the better, more complicated models have taught us a lot, there has not really been a huge difference in the modeled IGMSTA, which should tell you not to obsess to much about it.
Eli says:
“The interesting thing with respect to greenhouse gases is that there has really been little difference in the 2x CO2 global result as we went from pen and paper (Arhennius) to one dimensional computer models, to 3 D models with silly oceans to better ocean coupling and so on.”
Which suggests the complex climate models have been tuned until they produced the expected result. After all, any climate modelling group that came up with model that did not confirm the previous results would likely be presumed to be wrong until they could prove otherwise – an impossible task in a field where laboratory experiments are impossible. This would force modelling group back to the lab and come up with a model that did confirm previous findings if they wanted others to accept their model.
Eli– Most robust measurements in the sense of a) being less vulnerable to instrumental errors b) having more complete spacial coverage and c) having more replications. Most robust prediction in the sense that the model agree on GMST better than other features.
I have nothing against 1-D models. Often, at this blog, I’ve said that I’m not sure detailed 3-D models are any more accurate or precise with regard to predicting things like sensitivity.
I’ve pulled apart the components of GISS’ ModelE from the simulations page which I think might help understand what’s really included in some of the climate models. Based on this, I think we can also extend it to see if it is still working up to today or 5 years from now.
First comment is we can forget about using “forcings in W/m^2” for anything since ModelE calculates the total net forcings increase to be 1.92 W/m^2 to 2003 while the temperature increase is only about 0.6C to 2003.
That means the trend to date calculations are based on only 0.32C / W/m^2 versus the equilibrium temperature response they keep using of 0.75C / W/m^2. gavin says this is due to the ocean thermal lag response.
Second, the GHG forcing component of the model has an increase of +1.0C to 2003 built in while the “Other forcings” included in the model provide a net -0.4C to date (mostly Aerosols and Volcanic negatives).
I’ve charted this here.
http://img183.imageshack.us/img183/6131/modeleghgvsotherbc9.png
The GHG forcing component follows “4.053*ln(CO2) – 23.0” extremely closely (R^2 is 0.97) so we can extend that out to the future based on the CO2 forecasts. This increase works out to 0.231C per decade (and will stay very close to that rate for a long time based on the characteristics of the Log warming formula and CO2 forecasts). It reaches +3.7C by 2100.
The Other forcings component seems to be flatlining now with the Volcanic forcing going back to Zero after Pinatuba. The Aerosols component is flatlining in the final years up to 2003 as well but who knows what they will do with this in the future. There could be small Solar 0.1C decline given the state of the Sun recently. I’ve charted all the significant Other forcings individually if someone wants to see them.
Based on these trends, I’ve extended ModelE out to 2013. It is already out by 0.15C (or more) up to today.
http://img135.imageshack.us/img135/8594/modeleextend2013gi9.png
Bob, thanks for the interesting graphs. It might be enlightening to plot image 1 and image 2 together starting at year 1980. I think it would show a couple of features. Such as, why was 1998 claimed to be such a pentultimate year when 1998 could (did?) show how much the models missed? And, I believe it would show that with possibly increased aerosols from India and China, and the opposite of 1998, that there is good reason the models have not been invalidated as yet.
To John, here is image1 and 2 plotted together since 1980 – starting to get a little too busy now.
GISS’ record did not show the big bump for the 1997-98 El Nino like all the other temp record agencies did so it is even hard to see it on this graph. I think the Asian brown cloud didn’t really get going until the early-2000s and it certainly doesn’t show up in ModelE’s Aerosol forcing up to 2003. There is nothing included in this extension for a big increase in Aerosols from Asia (but Hansen will be including it from now on I imagine).
http://img175.imageshack.us/img175/2107/modeleextraev0.png
I changed the Non-GHG extension slightly as the formula was so close to flatline, I just left it at that before. But to chart it on this page I had to use the formal formula so it is just slightly different (not more than 0.01C).
Here is the Aerosols Forcing which looks a little “forced” to me.
http://img175.imageshack.us/img175/6919/modeleaerosolshb4.png
I agree with “”Here is the Aerosols Forcing which looks a little “forced†to me.”” I note that the June 1991, Pinatubo Volcano eruption is modelled quite well in comparison to the 1998 El Nino. That was my point. To me, it indicates, as Willis has pointed out ,that doing one pertubation such as Pinatubo has little, if anything, to do with others. In fact, looking at your graphs, to me, it is obvious that with the modelled effects of aerosols in general and the modelled effects of aerosols in particular to Pinatubo, that the general aerosols have been “overtuned” with respect to Pinatubo, a known, well measured input. i.e. Pinatubo is a large sudden aerosol injection, the general aerosols have lots of caveats/assumptions.
The other thing about this type of backcasting. The Pinatubo has a good correlation with the model. 1998 does not. It is just repeated evidence that Lucia’s comments of using noise her way rather than Gavin’s is correct, IMO.
John, the Aerosols impact is from smokestacks, vehicle emissions, burning, that kind of thing – Black Carbon and Sulfates.
Volcanic aerosols impact is a separate forcing and, other than Pinatubo, ModelE misses the other volcanoes by quite a bit in my mind. I should have put this chart up too (I’ve moved GISS temps down so that one can see if the impact closer.)
http://img101.imageshack.us/img101/7802/modelevolcanoesmr4.png
Thanks Bill. I had noted how the GCM peole had pointed out Pinatubo as validating the predicted cooling. Interesting that other forcings are not inline as well. Though from what I remember from RC, their Chichon looked to track better. Perhaps they used smoothed data, or a lag of about 6 months.
Though of interest, is that several of the major components of volcanic activity are the same as antropogenic emissions. I would assume that the relative amounts of similar constituents would be modeled similarly.