Monday
Jan282013
by 
Josh
Josh We told you so - Josh 197
Jan 28, 2013 Climate: MetOffice Josh ;)
Update on Jan 28, 2013 by
Josh
Josh
Richard Tol makes a very good point so I have changed the title from 'Predicting the past' to 'We told you so'.
Reader Comments (59)
Mystic Mug.
A welcome addition to Dr. Dan Streetmentioner's Time Traveler's Handbook of 1001 Tense Formations!
coffee on my screen :-) ... brilliant
Retrospectively brilliant!
How about a cast for Julia's mouth?
Another cracker!
castigat ridendo mores springs to mind.
Predicting the past is actually a valuable part of the scientific process. Newton, for one, predicted that apples fall to the ground. People knew that already, but they did not know why nor that apples falling to the ground and planets circling the Sun are manifestations of the same phenomenon.
If you type the name 'Senna the Soothsayer' into GOOGLE look who tops the list.
An interesting image too
That is very cruel.
Telling the truth can be very hurtful.
"If you type the name 'Senna the Soothsayer' into GOOGLE look who tops the list.
An interesting image too"
Awesome set of images!
Updated. I think Richard is right. I liked the alliteration of 'Predicting the past' but it isn't quite right in this context.
Anoneumouse, 'Senna the Soothsayer' into Google. That is very funny!
I'm not sure that's true is it Richard? Newton proposed an explanation for why apples fall to the ground which was experimentally verifiable and which worked for data other than apples and trees. He explained a large range of commonplace observations that were not understood.
Had he proposed that apples fall to the ground because there a force upon them exerted by trees, this would have been a perfectly valid explanation as well, but only until it was found that apples also fall to the ground if dropped out of a window. Or that pears fall to the ground as well. Neither are predicted in his model of the past, but both that and the real model make accurate predictions of apples in narrowly defined circumstances.
IOW, hindcasting is not the same as validating your model empirically using out of sample data, is it?
Global warming may be a result of the 90% of the universe which is dark matter. The concept of a theory which is wrong has been abolished in favour of continuous revenue.
Richard Tol is semantically confused.
@J4R, JS
The first test for any new theory/hypothesis/model is that it can explain things that have been observed already - predicting the past. Only after that test is passed, you move on to predicting things that have yet to be observed (either because they have not happened yet - predicting the future - or because the monitoring has yet to be put in place).
Not only do they provide perfect ‘hindecast ‘ by changing the models until they match the past, they claiming this means they can predict the future until that proves wrong and the circle starts again. They also indulge in rewriting past data so that it better supports current and future claims of climate doom. Now that is having you cake and eating it. And shows that for some, Orwells 1984 view on the manipulation of historical facts is taken has an instruction manual not a warning.
Richard Tol is quite right to say that the first test of a hypothesis is whether or not it can explain what is already known. In the case of climatology this would be tantamount to predicting the past. Can the models used by the Met Office and similar organisations retrospectively "predict" the major changes in climate over the past couple of millennia or so? Can they "predict" the climate in Roman times, the subsequent cooling, the Medieval Warm Period, the Little Ice Age and the rebound in temperatures from the Little Ice Age?
Surely if the models could do all that successfully there would be no need to "get rid of the Medieval Warm Period" as one researcher infamously put it. The fact that the current standstill in global warming was not predicted, except retrospectively, proves that climate researchers still lack an adequate understanding of natural variability and until they do have an adequate understanding of it they should not expect this country and other countries to take drastic and obscenely expensive action to combat threats predicted by their obviously flawed models.
To compare the insights of Newton with the hindsights of GCMs is quite a leap.
Imagine if the gravity model had had to wait for GCM-style analysis.
The first model predicted all apples would fly upwards, in close accord with a strongly-held belief that apples were lighter-than-air, and that seeing them falling was an illusion occasioned by the fallibility of our minds, and furthermore that all orchards had to be destroyed now, before the effect spread and crops of all kinds escaped into space. The threat of Catastrophic Apple Generated Famine, in other words.
But nets set up above orchards were not a success at harvest time. So the next model predicted they would fly sideways. But that turned out not to be of much practical benefit either. Improved information flows led at last to the modellers conceding that apples could really fall downwards under some circumstances, and behold the next models predicted that. A fruit-flux adjustment had been devised. In view of the seriousness of the problem, this was intitially called the gravitas forcing adjustment. The models had the flying apples returning some time in the future to avoid too much loss of face for all involved, but for all practical purposes a falling apple forecast was to apply for a sliding decade at a time until further notice. The destruction of orchards though, ought to proceed apace. Just in case.
Do they call their new computer model the MET-BTTF?
John Shade, ("Imagine if the gravity model had had to wait for GCM-style analysis")
But just imagine what Newton could have done if he had access to a sooper dooper computer like the Met Office. Funnily enough, someone has.....
http://kasmana.people.cofc.edu/MATHFICT/mfview.php?callnumber=mf202
Well, not a computer, but an electronic calculator. The result was - he went nuts! Could explain a lot?
A strong possibility for the 2014 Josh calendar?
Josh wrote: 'Anoneumouse, 'Senna the Soothsayer' into Google. That is very funny!'
What a pity it doesn't direct us to her pod cast. Senna, gettit? You know, pod...
[sigh]
JF
JF; that really would be predicting the passed...!
Unlike GCMs, with their uncanny ability to "hindcast", Newton's equations, underpinned by his theory of gravity, also allow "forecasts" in that they can predict the future position of the planets with uncanny ability.
An ability that GCMs are lack.
Please everybody, do be advised that you cannot, absolutely cannot, "predict" the past. To predict means eg "to say or estimate that (a specified thing) will happen in the future". A word means what it means, not what you'd like it to mean. (So, pre = before, in the past, it's already happened, like over and done with type of thing; dict = say). Got it?
Simon Abingdon...Well the blokes and bloke-esses at the MET office can't predict the future with any real accuracy so why not try to get something right for a change by looking backwards.
I'm pretty certain that some of them are looking over their shoulder's hoping not to see someone coming to give them a tap on the said shoulder. You never know what's ahead of you in this life, do you?
P45s anyone?
PW
Are computer models reliable?
Computer models are the only reliable way to predict changes in climate. Their reliability is tested by seeing if they are able to reproduce the past climate, which gives scientists confidence they can also predict the future.
Met Office Publication
I think what people find difficult to understand is what is this thing that we call a model? Well, it’s a huge computer code and it’s about solving the very fundamental equations of physics which describe the motion of the atmosphere, the motion of the oceans, how clouds form, how the land interacts with the sun’s rays, how it interacts with rainfall and so on and so on.
So what these models are is hundreds and thousands of lines of code which capture and represent our best understanding of how the climate system works. So they are not in a sense tuned to give the right answer, what they are representing is how weather, winds blow, rain forms and so forth, absolutely freely based on the fundamental laws of physics. How do we know that they’re good? Well we continually test them against observations of the current climate in lots and lots of ways. At the Met Office we use the same model to make weather forecasts as we do to make our climate predictions, so every day we are testing the model and saying, ‘how well did we do with the weather forecast?’ We know that on many occasions our weather forecasts are incredibly skilful and that’s increasingly giving us confidence that the science in our models is fit to do this ‘crystal ball gazing’ into the future to say what will happen to our climate as we go really into uncharted territory. Because we are taking this planet to somewhere where it has never been before, or at least for millions of years.
Thanks, Josh. The text and illustration are perfect!!!
"Please everybody, do be advised that you cannot, absolutely cannot, "predict" the past. "
Don't be silly, we do it all the time, especially in observational sciences. I will take an example from CRU. Have a look at A CRU temperature curve. That curve is a prediction, literally a prediction, of the average value of unobserved temperatures in the past. So, for example, in 1883, CRU has an estimate for the global average based on a sample of stations. That average represents the prediction for the average of all temperatures at unobserved locations. In short, we say that the sample can be used to predict what would have been observed at other locations had we been there to observe it.
Luckily, we can test this prediction of the past as more historical data is rescued from archives, so we can test the prediction.
In short, prediction normally has "time" asssociated with it but dont let this mislead you. Prediction at its heart means making a statement about values that you have not "observed" yet. Today I weigh 210 lbs. Tell me what I weighed yesterday. You are making a prediction about "the past" It is no less a prediction than if I asked you to predict what I weigh tommorrow.
@steven mosher (Jan 28 4:13PM)
Completely and I suspect mischievously wrong.
You admit that "Prediction at its heart means making a statement about values that you have not "observed" yet". The whole world can see the sequence of tenses implied by this definition.
Again, "Tell me what I weighed yesterday" invites me to predict what you'll soon (in the future, get it?) tell me you weighed yesterday, unless I already know whereupon it's no longer a prediction, is it?
@Rog Tallbloke Jan 28, 2013 at 10:20 AM says "Mystic Mug." Did you perhaps intend "Mystic Muggles"?
(with credit to JK Rowling)
Steven
I would agree, in that you're predicting what a data point will prove to be. This does strike me as subtly different from predicting the past in a general sense though.
The more common - or perhaps just "layman's" - term for what Richard Tol describes is surely "backtesting". Stock market traders do this all the time:if they think they've hit on a new way of identifying a leading indicator of price change, they will always backtest it to see if it works on past data where the prediction can be tested against the now-known outcome.
The problem they have is that because the data tends to lead the analysis, there's always an element of post hoc ergo propter hoc about it all. This is why technical traders haven't yet won everyone else's money off them.
It is as though you took the results of spins of a roulette wheel and tried to argue there was a predictable pattern. If you assumed enough complexity you could no doubt kludge one together, but it wouldn't work for predicting the results from any other roulette wheel or in the future.
steven, "retrodiction" is the appropriate word for what you're on about. I'm surprised you didn't know that.
Speaking of nomenclature, what are the accepted terms for:
(a) initializing a GCM with a past state, and running the GCM forward in time using no additional information [in other words recreating what a prediction at that time would have been]; and
(b) initializing a GCM with a past state, and running the GCM including the actual exogenous inputs (e.g. volcanic activity, solar)?
steven mosher
Can I use your method of predicting the past for tomorrow's Euromillions on a ticket I buy on Wednesday (30th Jan)? It would be great if you could tell me how to do that.
Thanks in advance
Sandy
Jan 28, 2013 at 3:57 PM | Unregistered CommenterAsk the expert — Prof Julia Slingo
Understanding physics does not mean one is able create a model (in this case computer program) that is able to mirror the complex chaotic behavior of the Earth's climate. Weather models possibly are able to forecast weather events over a few days but rapidly go off the rails when extended too far. So why should climate forecast models not go off the rails in a similar manner.
Prof Slingo also addresses us like we are a grade school class and then finishes her sermon with a statement that is most certainly wrong, providing that she is still referring about climate. A lesson in how not to combat skepticism.
Simon,
these people doesn't understand the basics of verbal communication, and they don't care to do so.
They could have just said "fitting the model to the past", but no, they have to be obscure.
Robert Austin,
I doubt very much that the real Slingo would comment here at the heart of the darkness. (joke)
The 'Ask the Expert' comment is surely a spoof. Only a political activist welcoming the ‘CO2 as a Crisis’ story would be persuaded by the facile argument that models used to help forecast the development of existing objects such as Atlantic depressions heading our way can be run and run and run without benefit of hourly or 3-hourly updates of surface observations, frequent cross-checks with satellite imagery, refinement with radar, and inputs from weather balloons. Without these, the model runs for weather forecasts would soon be producing nonsense. Here comes the act of faith. The nonsense will sometimes be one way, sometimes another, and it will all average out over the decades of model-time to give results which don't look crazy. Especially if someone stays on guard throughout to redirect runs away from destinations so haywire that they would harm the credibility of the average at the end. And of course, if the result is still crazy, or at least out of line with current pronouncements and positions, there is no need to publish it.
More comments on the Met Office, Models, and PR can be found here: http://www.bishop-hill.net/blog/2013/1/10/spot-the-difference.html, including one (Jan 12, 2013 at 3:37 PM) which describes Paul Nurse’s naïve response to a weather-tracking model shown on his own tv programme.
Jan 28, 2013 at 4:13 PM steven mosher
If you are estimating a past value (or values) from older and more recent values, I'd call it interpolation. I'd reserve prediction (or extrapolation) solely for estimating future values from past and present values.
(I think it's worth discussing what would normally be a minor point of terminology as we are in effect accusing the Met Orifice of a form of dishonesty.)
Sadly, it is not a spoof.
http://www.metoffice.gov.uk/media/pdf/q/e/julia-slingo.pdf
All I can say on reading mosher's post above is: wow.
Richard Tol, steven mosher
"forecast" and "predict" refer to future events.
"hindcast" and "retrodict" refer to past events.
Use the correct technical term appropriate to the context and you'll be less likely to mislead and create confusion.
@Simon A
Predict is Latin for foretell. The "fore" is not necessarily chronological. If you reinterpret the "fore" in foretelling an event as "before the event data are included", all is fine again.
@Josh (Update on Jan 28)
I think it is a great pity that you changed the title of your cartoon from "Predicting the past" to "We told you so".
Both Richard Tol (Jan 28 11:00AM and 12:33PM) and steven mosher (Jan 28 4:13PM) use that phrase (mosher "prediction of the past" to be exact) and to have kept the original cartoon title would have highlighted not only the absurd idea of predicting what is already history but also the neglectful insouciance of their use of language.
What they're doing is of course nothing more than comparing theory/computer-output to the historical record and making inferences about eg what a more detailed historical record might have said (Tol "the first test of any new theory is that it can explain things that have been observed already". mosher "the sample can be used to predict [he means infer] what would have been observed at other locations had we been there to observe it").
@Richard Tol (Jan 29 8:50AM)
"If you reinterpret the "fore" in foretelling an event as "before the event data are included", all is fine again."
All is not fine again Richard. What possible justification is there for altering the normal meaning of a word? To foretell is in the present tense. It refers to future events.
The lexicon already includes all the words you need for precise expression. (See my post above Jan 29 4:14AM). Please stop mangling the language.
Simon, thanks for repeating this! You can sort of see what Richard and Steve's point is. If a particular model suggests that a particular observation should be A and not B, then it is a reasonable way to test that model to see whether the observation was A or B. This can be done even if A and B took place in the past. Hindcasting is not a useless exercise. That being said, such tests are not as good if there is any way in which you were aware of the outcome - A or B - prior to constructing your model. Predictions of FUTURE events are always, when all other things are equal, more risky than retrodictions. I think that Richard and Steve know this very well. There is a grey area in that some retrodictions are more risky than others. For example, the logical link between the nature of the model and the parameters that are fed into it, and the outcome of the model, may be quite complex so that it is difficult to 'tune' the model to reproduce the known outcome. In such a case, the retrodiction is harder to get right than if you are just trying to retrodict whether the coin fell head or tails based on ... all available knowledge, including having seen it fall. An even greyer area is if the person making the retrodiction genuinely did not know what the outcome had been. But for climate models, we have good evidence that neither of these features apply: modellers are usually very aware of past global temperature anomalies, so they don't make their retrodictions blind. And, as e.g. discussed at Lucia's recently, in practice it is feasible to tune climate models to achieve the 'desired' output.
In my view, slippery use of language can lead to slippery understanding. Retrodiction is easier than prediction. No shame in that - I consider it quite remarkable that even retrodiction is possible for global temperature using global ocean-atmosphere climate models. But one shouldn't call successful retrodiction successful prediction. Ever.
There is a world of difference between a public prediction of some future event in the presence of considerable uncertainty, and the making public of the achievement of getting your prediction procedure to agree with an existing observation. The latter is fraught with risks of tampering, pampering, over-fitting, and perhaps other specific-goal dominated temptations of which the conscious mind is barely aware; we might view success of the former with tentative admiration and interest, and success of the latter with a bit of a yawn pending the real test.
That £60m comuter, 'Deep Black', which the Met office used our money to invest in, should be re-christened 'Deep Sh*t'....