I’m feeling stupid about figuring out which TAS runs match TOS runs at The Climate Explorer. I’ll start by saying why I want to match these: I would like to create ENSO corrected graphs to test whether observations are consistent with models. The reason I want to do this is so I have both “Knight et al.” and “Easterling and Wehner” type analyses (but with my method of estimating variances and so on.) The Knight et al methodology requires me to obtain the (TAS, TOS) time series pairs from each run, then create an ENOS corrected TAS using the correlation between the TOS and TAS. It is essential that the (TAS,TOS) time series come from the same run.
Next: I’ve been getting the TAS from the climate explorer. The Time series computed from CMIP5 data are here. TOS data are also available. However, if you look carefully, you will note that the page states:
The ensemble numbering 00, 01, … is an internal convention. The same number may refer to a different ensemble member tomorrow. The netcdf metadata contains the CMIP5 ensemble numbers rNiMpL that are definitive.
If the ensemble size is different for different variables they do not match (so tos00 may correspond to tas02). Again, consult the netcdf metadata. If the ensemble sizes are all the same, they probably match.
Note that for EC-EARTH, 7 members are available for TAS driven by RCP45 while 6 are available for TOS. This means I don’t know which of the 7 TAS members match the 6 TOS members.
Evidently, (I think) the meta data are obtained by selecting a field (e.g. TOS for EC-EARTH, rcp-5 scenario) and clicking. This gets me to a page whose name is “http://climexp.knmi.nl/select.cgi”. On that page, I see little [i] symbols. Clicking the [i] to the right of text that reads ” Found ensemble members 0 to 6 ” reveals a page whose url is “http://climexp.knmi.nl/showmetadata.cgi?EMAIL=someone@somewhere&field=cmip5_tos_Omon_EC-EARTH_rcp45”. There is tons of information there. Geert Jan (who is quite patient and helpful) tells me to
“You should look under “experiment_rip” or “parent_experiment_rip”. If you select individual ensemble members on the previous page you can also see the data for the other ensemble members.”
I’ve found “parent_experiment_rip” to the right it reads “r1i1p1”. I don’t know how to interpret that. (Every page seems to say “r1i1p1”.) If anyone knows how to figure out which TOS runs match which TAS runs, please tell me what I am missing here. Because I am mystified.
If, in fact, figuring out which go with which cannot be done I will either (a) need to download gridded data and compute both time series myself or (b) try to write up a paper without ENSO correcting. I’d really like to ENSO correct. I prefer avoiding recomputing TAS or TOS from gridded data. (Yes. Lazy.) And in anycase, it may turn out the resource making gridded data is equally opaque to me. (Or not. I just don’t know.)
So if someone can explain to me how to figure out which EC-EARTH TAS series match which EC-EARTH TOS, that will help me with this model. The information will also help me figure out how to sort this out with other models.
Thanks!
Update: Ok… I think I’ve figured it out. I need to scroll down to “analyze individuall”, click and then the numbers are revealed. This will be tedious… but not ridiculously so. (Easier than writing a script to download gridded data, process and so on!)
Perhaps a little off topic; but is it not possible to examine the distributions of (model-real) for hindcast and for forecast. If the properties of the two are different, does this not indicate a failure of the models?
DocMartyn–
Distributions of what? And what do you mean by a “distribution” (do you mean probability density function)? And what would constitute a ‘difference’?
The distribution of the residuals (real-mean) should have a normal distribution in the hindcast, if the model has been trained to match the hindcast. The same Gaussian distribution of residuals should occur in forecast if the model is as good at forecast as hindcast.
DocMarty
Residuals from a line? Comparing the standard deviations of the residuals would be statistically more powerful than comparing the whether both distributions are normally distributed. For that reason, testing the ‘distribution’ is possible, but likely a poor allocation of resources (here: time to do an analysis.) That said: if someone else wants to do that, I think it’s possible.
Lucia, I was going to give it a go if I could work out how to download their model runs, and be sure I was working with what I thought I was working with.
“if the model has been trained to match the hindcast” “if the model is as good at forecast as hindcast.” These two sentences don’t really go together. If the model was _not_ trained to match the hindcast, but was “just physics”, then we would have a right to expect the model to be as good at forecast as hindcast. If it is not, then it was trained to match the hindcast. We would never expect it to be as good on out-of-sample data as on in-sample data. However, it shouldn’t be a tremendous lot worse unless it was overfitted:
https://en.wikipedia.org/wiki/VC_dimension
MikeR, they typically claim that the models are not strictly trained, but updated on the basis of things other than temperature output.
I do not believe them.
Lucia,
In order to explain ALL of the modern global warming we observed between the 70s and the 00s, we ONLY need to explain what happened in 1978/79, 1988 and 1998. At all other times, global temperatures simply follow NINO3.4 SSTa in its general progression.
These three instances were the only occasions where global temperatures parted permanently from NINO3.4. And each time it happened abruptly, within a year, and the new higher level established stayed in place, no further divergence until the next jolt ten years later. The three of them together hold the ENTIRETY of global warming since the 70s.
It’s right there in the data:
http://i1172.photobucket.com/albums/r565/Keyell/Globalshift_zps0e78db58.gif
http://i1172.photobucket.com/albums/r565/Keyell/NINOvsglOIv2_zps5be10b66.png
http://i1172.photobucket.com/albums/r565/Keyell/Step1b_zpsf9c04bc3.png
http://i1172.photobucket.com/albums/r565/Keyell/Step2b_zpsb5a3e862.png
http://i1172.photobucket.com/albums/r565/Keyell/Step25b_zps430fd462.png
It is only to those who are convinced that CO2 drives global temperatures that the stagnation seems ‘strange’, a problem that has to be resolved by ad hoc explanations.
Kristian,
I don’t find series of animated gifs enlightening. If you are advancing the “jumps” theory as somehow suggesting CO2 doesn’t cause warming: I don’t find that remotely enlightening.
Lucia, so look at the png’s instead. It seems you only opened the first link.
The “jumps” are not a ‘theory’. They’re in the data.
Kristian,
You seem not to have read to the end of my sentence,
For that matter: I think you are deluding yourself a bit about “jumps”. Autocorrelated noise + trend can look like jumps . It doesn’t mean that we need to explain “jumps” nor that “jumps” from time to time have any particular meaning. Looking like ‘jumps’ doesn’t mean the system isn’t “trend + noise”.
Lucia, you said: “For that matter: I think you are deluding yourself a bit about “jumpsâ€. Autocorrelated noise + trend can look like jumps . It doesn’t mean that we need to explain “jumps†nor that “jumps†from time to time have any particular meaning. Looking like ‘jumps’ doesn’t mean the system isn’t “trend + noiseâ€.”
In other words, you didn’t look at the png’s. It is quite obvious that those shifts are not ‘noise’. We know exactly when they happen, where they happen and how they happen. They are evidently process-related. We can track them. And all we need to do is look at and follow the available data. No preconceived assumptions necessary.
But interesting attitude towards data. And the ‘background trend’ of yours is of course already assumed to be there before you even start looking at them … Perfectly circular reasoning.
Well, I won’t bother you anymore then. Sorry for the inconvenience …
Kristian
I looked at at least 2 of them. They are similar to lots of pngs I’ve seen lots of places with people claiming to see “jumps”. You also did not surround them with any narrative.
No. It’s not obvious.
What process?
What do you even mean by “track”?
What in the heck does this even mean? I’ve looked at lots of data. Merely “looking” at data rarely tells us what the data ‘mean’.
No. What I said is that if there was a trend and noise, the data would look just like that. But you seem to be claiming the fact that data look like that mean it’s not trend + noise. Or the rise can’t be CO2. Or something.
Other than the word “jumps” and thinking some sentences that seem to suggest merely looking at these graphs “explains” the stagnation, I don’t even know what point you are trying to make. I’m just guessing based on what other people sometimes try to claim based on “jumps”.
Obviously, I’m not going to present some thorough fully fleshed rebuttal to whatever you claimed, because you don’t seem to be saying much beyond:
1) Look at these pictures. There are jumps!
2) That explains the stagnation.
3) Something about CO2.
You might not even be saying that. But obviously, to the extent I have no idea what you are trying to claim, I really don’t think I need to pour over all 4 pictures. Two were enough.
Kristian,
Seeing apparent patterns in random data is so common that there is even a word for it: apophenia..
DeWitt,
‘Random data’? ‘Apparent patterns?!’
Worse than I thought …
Kristian: I looked at all of the graphs you posted. Like Lucia, I don’t find them enlightening. If I understand your hypothesis and put it into mathematical terms correctly, you propose:
T = kN + b_i
where T is the observed temperature, N is the NINO3.4 Index, k is a proportionality constant relating T and N (possibly = 1) and b_i (b_1 b_s, b_3, and b4) are constants for four periods of time (before 1978/79, 1978/79-1988, 1988-1998 and after 1998). How about showing us a graph of the difference between the observed global temperature and the temperature predicted by this equation with your preferred values for k and bi? (If you can get this far, you can think about how you might attempt to prove that your equation has better predictive capability than one Lucia might propose that uses NINO3.4 and radiative forcing. Your equation has more adjustable parameters.
Re: Frank (Aug 14 13:28),
Even then, it’s still just curve fitting to an apparent pattern absent a mechanism. Meaningless. Not even wrong, i.e. it can’t be falsified, to quote Pauli.
DeWitt,
And even with the curve fitting, one is left with questions: Why are the jumps always up? And how does there being “jumps” explain “the pause”. Why has it been so long since a “jump”? Is one due? Is there a known time between “jumps”?
Even if one accepted “jumps” as enligtening, it would still not automatically “explain” the pause (i.e. long time without another jump up) and it would not contradict CO2 causing warming — as all the jumps seem to be “up”.
As for the observations that surface temperature are highly correlated with El Nino. Yep.
“DeWitt Payne
Seeing apparent patterns in random data is so common that there is even a word for it: apophenia..”
Brush up your Shakespeare
Hamlet: Do you see yonder cloud that’s almost in shape of a camel?
Polonius: By the mass, and ‘tis like a camel, indeed.
Hamlet: Methinks it is like a weasel.
Polonius: It is backed like a weasel.
Hamlet: Or like a whale?
Polonius: Very like a whale.
lucia:
That’s easy enough: Increased CO2 causes warming, which are exhibited as “upwards jumps. ”
No, wait… not sure that’s what Kristian wanted us to “see.”
But I agree that trend + autocorrelated noise (1/f^2 noise in particular) looks suspiciously like jumps. I’d guess you’d need to look at other variables besides temperature to distinguish shift in equilibrium state from “jump like” random walks. I have no clue ATM how you’d do that.
DeWitt and Lucia: Yes, this would be a curve-fitting exercise. However, if curve-fitting to a step function works much better, there is some reason to consider an alternative theory. (A devil’s advocate could suggest very low climate sensitivity due to negative feedbacks plus sudden changes in ocean circulation/upwelling to produce step-functions.) However, I assume that the result of curve fitting will be that the step function on the average over-estimates temperatures early in “flat” periods and under-estimates temperatures late in the “flat” periods – ie that the flat period is mostly an illusion created by noisy auto-correlated data.
Tisdale is constantly publishing graphs with step-functions and I’d like to see them discredited – or not – because they actually fit the (noisy) data more poorly than functions based on radiative forcing.