The Blackboard

Lucia:

Congratulations!

A very clean presentation

Lucia:

Variation in Hadcrut3 data appears in your graph greater than remainder of indices, especially after 1994. I also read (below) the exchange between the commentators over at the latest WUWT posting by Anthony.

Given Zeke Hausfather’s comment here @ 8725 that “where the IPCC displays just one temperature record, they use HadCRUT3, not GISS”, and you have from time to time focused on the IPCC, I am wondering whether manipulation of data is being applied over at the MET. Any comment?

“Steve M. (06:28:05) :
OT, maybe: From the MET office website regarding HARCRUT3:
“We have recently changed the way that the smoothed time series of data were calculated. Data for 2008 were being used in the smoothing process as if they represented an accurate esimate of the year as a whole. This is not the case and owing to the unusually cool global average temperature in January 2008, it looked as though smoothed global average temperatures had dropped markedly in recent years, which is misleading.”
Am I missing something? They have to change their process because January 2008 was “unusually cool.” And I thought HARCRUT might have better data than GISS.

Sven (06:54:05) :
Re: Steve M. 06:28:05
Where did you find that statement? I can’t see this.
I was wondering for quite some time why Metoffice has not renewed their graphs, some from November, some even from April 2008. I sent hem an e-mail yesterday asking about that and got a reply that it was forwarded to some specialists who would get back to me. Nothing so far apart from the fact that they have (interestingly soon after my mail?!) replaced the year 2007 with 2008 without redrawing the graphs (that still show only 2007 with 0.4C anomaly as the end) themselves at
http://www.metoffice.gov.uk/cl.....crut3.html
and these plots still end with Nov. 2008
http://www.metoffice.gov.uk/cl.....tures.html
Metoffice used to redraw the graphs simultaneously with new data coming in and it seems to me that they are uncomfortable showing the picture of temps going visibly dow. Reading what you just posted, it really seems to be the case – they are sitting and thing about what to do? Sounds like a conspiracy theory but…?

Sven (07:19:04) :
Re: myself 06:54:05
OK I found the Metoffice statement:
http://hadobs.metoffice.com/ha.....bal/nh+sh/
It seems that, in order not to show cooling, they have just not used the 2008 data for the smoothing at all?
And the grapphs I was referring to earlier still are not up to date

Sven (07:23:15) :
Oh, one more thing. Actually I remember now that this statement “We have recently changed the way that the smoothed time series of data were calculated. Data for 2008 were being used in the smoothing process as if they represented an accurate esimate of the year as a whole. This is not the case and owing to the unusually cool global average temperature in January 2008, it looked as though smoothed global average temperatures had dropped markedly in recent years, which is misleading. ” is actually old. It was put there earlier last year. Before that they did smoothing with data from an incomplete year and after this announcement stopped doing this. But now it seems they are not doing it even when the year is over?!

Brian,

I am wondering whether manipulation of data is being applied over at the MET. Any comment?

I don’t have any reason to believe anything untoward is happening at HadCrut, if that’s what you mean. Of course, decisions do have to be made on how to deal with historic data which may have various measurement issues.

My opinion about their former smoothing issue is this:
1) The used to use a poor smoothing method that distorted the end point. The problem tended to be worst during years that started with either sharp jumps or drops in temperature.
2) In years starting with sharp jumps, the images didn’t bother them so much, and they didn’t think about the distortion.
3) Last year, the temperatures dropped early in the year. This also followed a period of non-increasing. They found the image puzzling. After scratching their heads, the finally realized their method of smoothing was poor.
4) The fixed it.

It’s not a conspiracy. However, the process is an example of confirmation bias: As long as the method resulted in smoothed curves that communicated an overall message they considered “right”, they didn’t notice the method itself was poor.

Now, should the revert to the crummy method when temperatures start to rise, then you can accuse people of conspiracy. But, I bet they won’t do that because it’s not a conspiracy.

Am I reading it right? It looks to me like the divergence of GISS and HadCRU from the multi-model mean only really starts in 2004; prior to that both measurements were pretty much spot on. I thought the models had been reading high for longer than that.

Lucia,

What happens to the lower 2SD limit after 2008? Is the up tick in 2007 a spike or a step in the long term trend? (i.e. if temps stay the same in 2009 will they fall belong the 2SD range?)

Doug,
If you do a hindcast test using data from 1980-2000, the models look fine. I haven’t done 1880-2000, but I suspect that’s not too bad either.

It’s only when you add data collected after the SRES used to drive the models were published that the projections look high. But the disagreement is sufficient that if we compare trends using the Santer method using data from 1980-now, the models look fairly poor. That result is not restricted to starting the comparison in 2001. (The results are worse if we test the 2001-2008 trends. But they are still poor if we test the 1980-2008 trend.)

Raven
The standard deviations are computed based on the distribution of annual averages for that month. Just as “model weather noise” affects the anomalie, it affect the SDs. In fact, the SD’s are noisier than the trend. So, if we add the temperature anomaly + 2 SD’s we get a very squiggly line.

There is no particular information in the squiggle. It’s just that th annual average anomaly for the models had a little less scatter in Dec 2008 than a few months earlier. Rest assured it will squiggle back out shortly!

Lucia: I think its time climate modelers threw in the towel. Now its J. Scott Armstrong, founder of the International Journal of Forecasting saying ALL the models are absolute tripe excuse the expression…..
http://wattsupwiththat.com/200.....g-climate/

What this indicates to me is that we will just have to wait and see where the temp goes in the next 2 or 3 years. If it continues to depart from the GCM downward beyond the 2 SD then I think that would indicate we are quite probably in the down part of the, what seems to be, another 60 year cycle. The GCM model guys and gals will have to go back and rethink the models some more.
We were in the up part from the late 1970s and that just happened to roughly match the GCM.

Time will tell. – just have to be patient. Don’t think we should be spending any money shoving CO2 into the ground though.

VG–
I read the the claim

“We conducted an audit of the procedures described in the IPCC report and found that they clearly violated 72 scientific principles of forecasting (Green and Armstrong 2008).”

I wonder what the principles are?

Steve–
Because both models and observations are rebaselined to 1980-1999, there is a degree of match imposed during that period. The average anomaly for all cases is simply set to zero during the baseline.

It’s worth knowing that when examining these types of graphs. If I felt like playing I could probably find baselines to make the models look bad on this sort of graph. But… the IPCC states is projections relative to this baseline, so others would be be very fair in this context.

Lucia:

So does this mean that the departure of observed temperatures and GCM is is less meaningful and may be somewhat expected? In future, when you have more data and a longer baseline (say 30 years) you could expect better agreement but continued deviation would be more meaningful?

You are light years ahead of me on this stuff and statistics beyond the elementary puts my brain into seizure so you may go huh? on this question.

Love your blog though. Visit it often. Great stuff!!

Lucia:

“I wonder what the principles are?”

Here’s the paper:

http://www.forecastingprincipl.....udit31.pdf

Lucia:

The 139 Principles:

http://www.forecastingprincipl.....dshort.pdf

just to keep you posted Anthony has now put up both original letters from Dr J Theon so its 100% believable I’d place my scientific trust in this guy any day over his previous underling sorry…
http://wattsupwiththat.com/200.....r-muzzled/

Lucia,

I think the caption to your graph really ought to be leading 12 “month” averages, not leading 12 “year” averages!

I’m sorry that my contribution here seems to be limited to pointing out minor errors.

Jorge– Editing points are important. Thanks!

Steve

So does this mean that the departure of observed temperatures and GCM is is less meaningful and may be somewhat expected?

No. It just that this method of trying to detect deviations isn’t very efficient. It both lacks power and is ambiguous.

It’s useful to look at this graph partly because everyone does look at this sort of graph. Still, the trend method has more power, and lets us use an estimate of real earth weather noise.

On Hadcrut adjustments at year end see http://hadobs.metoffice.com/ha.....thing.html. Note particularly these two paragraphs:

In March 2008, some diagrams were placed on this web site which showed smoothed annual series that included data for 2008. The annual value for 2008 was based on the only two months of data – January and February – that were available at the time. January and February 2008 were cooler than recent months, leading to a marked downturn towards the end of the smoothed series (Figure 2, orange line) that caused much discussion.

followed by:
A similar, albeit less extreme, situation occurred in March 2007 (Figure 3, orange line). January 2007 was nominally the warmest January in the HadCRUT3 record and the anomaly for February was only a little lower. This led to the smoothed curve (including the annual estimate for 2007 based on only two months of data) showing stronger warming at the end of the series (Figure 3, orange line) than when data for only whole years were used (Figure 3, blue line).

Please read the whole page

If the Hadley Centre had updated their global graph it would look like this:
(You have to click on the link)
http://www.freewebs.com/fourdj.....annual.png

I used the same method to smooth it as they do as described on the page mentioned above.

Rich,

I followed this debate about endpoint smoothing last year. I had no real problem when Hadley decided not to use an incomplete year in the calculation but I was wondering what they would do if, in fact, 2008 finished up cold.

As far as I can see from your posting and the Hadley website, they have decided to postpone updating the graph with the 2008 figure. This may simply be tardiness on their part but it could be they are looking at alternative smoothing methods.

If they do change the method to reduce the apparent fall in the smoothed temperatures, I will certainly cry foul. To be fair, it always was a stupid method which gave such a strong weight to the end value and was virtually guaranteed to exaggerate the existing trend, up or down.

I will also contact Mr. Kennedy so they know that people are taking an interest.

Hi Lucia – please could you do two graphs? I think it would help presentation if the measured data and the modelled data were on separate sheets with exactly the same scales. Thanks if poss.

tryingtokeepup.
I could… but why do you ask? I can make all sorts of graphs!

Jorge–
I suspect it’s just tardiness. I seem to recall last year there was some delay in finalizing the year end data. I suspect they like to be more than usually careful at year end.

Hadley’s January data was out later than usual too. For all we know, they are still waiting for Russian data!

Hi Lucia – I think given the no. of series and the limited range of colours it would make the visual identification of similarities and differences between the measured and the modelled data more obviously identifiable, esp. to those of us less familiar with the data on a day by day basis. Thanks

Thanks for the graph Rich.

I think this one needs to get trotted out everytime one of the warmers says there has not been a cooling trend recently.

It is obvious why the Met and the Hadley Centre do not want to show the continuing trend (even though they were pretty fast doing so when the line was going up).

I also do not like how GISS squeezes the width of their X-axis and uses the smallest scale possible on the Y-axis to make the slight warming over the past 130 years look like a catastrophe. The average person has not looked into the temperature data as closely as some of us and does not recognize how chart scaling can exagerate a trend.

If GISS put their model’s predictions on the same graph they would have to increase the Y-axis scale by a factor of 2 and it wouldn’t look like quite the catastrophe in that case. An average person would look at this graph and wonder how this global warming case became so exagerated.

tryingtokeepup–

Are you trying to identify point by point similarities? We shouldn’t expect point by point similarities. By super-imposing, we can see:
1) The general trends match pre-2000.
2) Both observations and models have dips after the eruptions of Pinatubo and El Chichon.
3) The observations are “wigglier”. We’d expect this because of internal variability. The individual mode runs are also “wiggly” and we should not expect any “wiggles” to match.
4) The two series appear to diverge after 2000. Either this is “weather noise”, or it’s a deterministic effect. However, they diverge.

I guess I’m reluctant to make two graphs if the goal is to try to do a mental subtraction to focus on the difference due to the “wiggle” due to natural variability. Even though you might wish to try to look at that, you would not really learn anything about models from that.

Hi Lucia – thanks for taking the trouble to clarify.

It’s just a simple presentation issue for point 4 – I think the divergence post 2000 would be more evident by having the modelled and measured data on separate charts of the same scale and limits side by side (or one above the other). I’m not seeking to prove/disprove anything by this and take on board your points 1-3.

On a wider issue I agree Bills comment above re: the resolution of some chosen scales – not sure what the solution is but I wonder if possibilities such as all data from public funded work having to be presented with along with its error limits.

tryingtokeepup–
The other issue with graphs is that I don’t want to end up in a situation where people request loads of custom graphs. Each takes some time to make, proof read, upload etc. It’s not a lot for one or two graphs. But the difficulty is that, looking forward, everysingle person will want some specific graph.

If you click the graph and see the larger version, does that solve your difficulties? Of is there something about superimposing the graphs that makes things difficult? Or the particular colors? Excell has loads of colors.

My preference is to make one graph that shows as much as possible, while remaining readable. I realize you may believe you can compare the data and observations more easily if they are placed on different graphs, but I honestly don’t see how this is so. Ordinarily, it’s easier to compare two sets of data by placing them on the same graph.

Jorge – There is no sensible way to extend a time series so that you smooth it to the end. However you cut it, it means you’re faking data. We ought to be stopping 10 years before the end and saying, “Look, the trend, as shown by the smoothed data, is still up”. That said, if I eyeball the graph and ask, “Does the Hadley-style smoothing seem to honestly represent a smoothed version of the given data?” I’d have to say, “Yes, right up to the end.”

You know, if they used a simple 21-point moving average filter, extending as before, it would still be rising in 2008. If I was making propaganda that’s what I’d do. Everybody knows what a moving average filter is so it’s easy to justify. Who’s heard of a 21-point binomial filter? Where’s the justification for it?

I suppose we could label them incompetent propagandists but I’ll prefer honest but tardy until I’m forced to reconsider.

Rich–
I’d heard of binomial smoothing. The hope is to smooth away some noise, while still keeping dips that are caused by deterministic effects. One might wish to show that temperatures really did drop after the eruption of Pinatubo. The binomial filter is a little better than a 21 year moving average if that’s your goal.

That said, they should always have done something to indicate that averaging of an sort does weird things at boundaries. They caused themselves a PR headache by overlooking the problems with the sliding definition of “averaged” at the endpoints. But, I don’t think anything nefarious was going on.

Hi Lucia – It was only a suggestion! No problem for me to see the larger graph and the request wasn’t meant to be a “me only” version, just seemed to be a clear way to illustrate the comparison you are making. No one else has raised so obviously not a problem. ttku

tryingtokeepup– I didn’t think you were asking for a “me only” version. I just like to know precisely what the issue is so that I can eventually come up with a solution that lets me communicate better.

I know for some people, the smallness of the image is a problem. I need to remember to add “click for larger”. For other people, it’s the color choices. Some of the guys here admitted to red/green color blindness. Lots of things can be changed. But when at all possible, I prefer solutions that don’t create multiple graphs if one graph will do.

So… I know I may sound like a butt head because I could just make two graphs– I’d really rather figure out some method of illustrating this with one graph rather than 3. (Three being a) the one I prefer plus b) the two you prefer.)

Rich, Bill,

I’d be highly reluctant to use the graph you posted because, as Lucia stated, the smoothing method is highly misleading. Because the smoothed line is extended to the end of the graph, the last point includes only 58.8 percent of the measurements of earlier points. This effectively exaggerates the trend at the end of the series by implicitly assuming that the next few years will have the average temperature of the prior few.

I much prefer a system that weighs all points equally, even if it does not necessarily extent to the end of the series. HadCRU3 with a simple 5-year moving average looks like this:

I recall Tamino having developed some fancy smoothing method that does away with the end point variability issue, but I prefer to keep things simple. Every point is the average of itself, the prior two years, and the subsequent two years.

If you want something more sensitive to the relatively “cool” past three years, you can do a three year rolling average smoothing:

Smoothing the data should show make it easier to identify trends in noisy data; it should not be effectively forecasting future data!

Reading through the HadCRU smoothing methodology, its actually worse than I assumed. They use the current year to fill in the next 11 years in their smoothing function at the end of the series:

“Extending the data series can be done in a number of ways, but the method used on these pages is simply to continue the series by repeating the final value.”

To see how much this skews things, lets pretend that 2009 had the same temperatures as a recent warm year, say 2005. In this case, the graph Rich posted would look like this:

I think we can all agree that any system weighted so heavily on the last datapoint in a long series is rather silly. I’m surprised no one called the folks at Hadley out on this before last year.

Rich,

It is interesting that statisticians talk about smoothing but electronics bods talk about low pass filtering. The 21 point binomial has a much faster roll off than any averaging type thus allowing for a greater suppression of high frequencies (noise?) for a given remaining low frequency.

At one time we used to talk about physically realizable filters which simply meant that a future value could never alter any previously smoothed value. This effectively means that only trailing averages, or more powerful equivalents, can be used.

Oddly enough this did not prevent using derivatives of past values to form part of the last smoothed value. In that sense, the last smoothed value was partly based on predicted (faked) future values. However with that kind of filter we did not go back and alter that value when the actual future data became available.

There are good arguments for taking data either side of a particular data point into account when smoothing past data but clearly this leads to endpoint problems.

In a strange way you could argue that the Hadley method of treating the next 10 years as having the same value as the current one is actually quite conservative! The snag is that we have to wait 10 years before we finally find out where the current point on the curve should be plotted.

As I am a lukewarmer, I am willing to be patient about these things.

Zeke,

I think one can be led astray by any smoothing method. I only work with the monthly data for the climate and I don’t smooth it at all.

Five-year rolling averages are completely inaccurate and are probably the worst smoothing timeline to use.

The temperature of 913 days ago, 30 months ago, should not be incorporated into the measurement of today’s temperature or the temperature of this month.

The temperature of 913 days in the future, 30 months in the future should obviously not be used in the measurement of today’s temperature.

A five year rolling average just averages the temperature over 1,826 days and pretends that number is today’s temperature. All kinds of inaccurate analysis and conclusions can result from that practise.

The climate is chaotic and noisy and hard to interpret without some kind of smoothing. But the drivers of that noise, the drivers of the climate operate on time-scales of daily, monthly and seasonal, never five years.

When one gets into oceanic influences, the time-scales then move to monthly, 3 months and 800 year time-scales, but there is no five years in the ocean influences either.

GHGs operate on daily, 3 month and 800 year time-scales as well. With GHGs, we are talking about photons of light here which operate at the speed of light. The speed of light influence is, of course, moderated by the time-scales of atmospheric molecules and ocean molecules ability to store them up – daily, 3 months and 800 years.

We need to drop all this smoothing and averaging unless there is logical reason to average the figures which matches the time-scale of the climate driver-influence you are trying to analyze.

“I think one can be led astray by any smoothing method. I only work with the monthly data for the climate and I don’t smooth it at all.”

Monthly data has been smoothed, to a one month resolution.

And the concept of a global temperature is the biggest smooth of all.

Pipped on UAH January 2009 and they are strange, very strange compared to RSS who show their number higher for January compared to same month last year and 0.567C higher for NH, whereas UAH show it lower form 0.299 for January 2008 down to 0.229 for Jan-2009. A whole 0.22C difference? Who’s flunking this time… John & Roy? A Doors song just sprung to mind.

To add, more puzzling is how the sum of the UAH NH+SH temps is 0.158C lower than the global figure of 0.307C, where RSS’s NH+SH add but for 0.002C up to Global. Let’s see who’s going to get the royal treatment as was bestowed several times on the other temp compilers who made data publication errors?

Correction on last post, looking in wrong columns. The UAH adds NH+SH as 0.308C

Sekerob– I haven’t looked at this month’s UAH yet. I’ll have to post today. I was thinking of doing something else interesting. (Answering a question either Zeke or Arthur asked me weeks ago!)

Yes, RSS+UAH are so close on hindcast, hardly worth looking at, but ~0.56C Higher than same month last year is interesting foremost because it contradicts the IBMY “it’s so cold” anecdotals.

The HadCRU’s had me wondering for longer also because they do not consider the Arctic where most deviation in temps is collecting.

Sekerob– Over the long term, HadCrut has a larger trend than GISS. I seem to recall the difference is about 10%, but I’m not sure.

Lucia, up or down? Eyeballing my 30 year chart that has the 6 major collectors in plus their 36 monthly moving average trend, The HadCRUT3v looks flatter in the first half of the decade then going down, whilst GISTEMP / NCDC was upwardish, then followed HadCRU. That’s suggesting there’s something to consider where in the latter part of this decade the strongest Arctic sea ice decay, annual, extent, but mostly area and volume reduction is being observed. There’s heat needed to do that.

Sekerob–
Since 1990 the HadCrut surface trend exceeds the GISS trend by about 10%. I don’t know whether this is due to UHI corrections, different choices for SST, the method GISS uses to include data for the poles or whatever. It’s just an observation of what you get if you compute the trends. I’ve posted the graphs here.

[...] For you see, their forecasts have consistently demonstrated themselves to be too high.  You can see above how Hansen’s forecast to Congress 20 years ago has played out (and the Hansen A case was actually based on a CO2 growth forecast that has turned out to be too low).  Lucia, who tends to be scrupulously fair about such things, shows the more recent IPCC models just dancing on the edge of being more than 2 standard deviations higher than actual measured result…. [...]