# Arctic Sea Ice Minimum Forecast

It’s that time of year again, time to submit forecasts of this year’s upcoming September Arctic sea ice extent to ARCUS.

I don’t intend to submit a forecast (except here on my blog). But I thought this would be a good time to re-iterate my forecast, and add another, which turns out to be not substantially different.

The “original” forecast method is quite simple: fit a quadratic to September Arctic sea ice extent data from NSIDC and extrapolate that one year into the future. Use the standard deviation of the residuals to define a 2-sigma error range. That gives a forecast of 4.15 +/- 0.98 million km^2.

I also have a modified version of the lowess smooth, which I can use to estimate future values as well. That gives a forecast of 4.13 +/- 0.96 million km^2. As you can see, the two methods make essentially the same prediction.

Note that these forecasts are based on historical data from NSIDC, so should be compared to actual NSIDC figures after nature reveals the answer.

Note also that the uncertainty level is substantial. The range by method 1 is from 3.17 to 5.13 million km^2, by method 2 from 3.17 to 5.09 million km^2.

And for those who like graphs, here’s the September data from NSIDC together with the forecasts:

### 33 responses to “Arctic Sea Ice Minimum Forecast”

1. How does your method rank against others in accuracy of past predictions (assuming you were to travel back in time and make a “prediction” using your curve and the current data)? Granted, your method is based on knowing what the past outcomes were, to arrive at your curves, but still… it would be interesting to look at.

[Response: I haven’t done the comparison myself, but someone once did and concluded my forecasts were among the most accurate. Given that it’s such a simple statistical model, I suspect that’s just luck.]

2. Mark

Since all forecasts are problematic (albeit interesting) I have what I find a more interesting question for someone of your talents Tamino. I hope you don;t find it a “dumb” one

My “guess” btw would be 4 million.

But because (over time) samples tend to converge on their trend (absent an outside different direction perturbation) then the result will almost certainly (95% of course haha) fall within your predicted ranges.

What value outside that range would indicate a trend reversal and with what probability? For example if the value was 6.3 million (roughly the average for the three decades from 1980-2010) – what is the likelihood

What value would “convince” us (with reasonable certainty) that something different was indeed going on? (For example 8 million would be higher than 2sd above the 3 decade average)

My reason for posing this question is that I think the extent will be above the record 2012 minimum (like 2013 was – which was almost “half way” back toward the 3 decade average) – and we all recall what the WUWT and their fellow pseudo-skeptics made of that.

The moment that it is higher than 2012 (and especially if it is higher than 2013) I expect to see claims of a “recovery”. Can we head this off at the pass?

[Response: If it’s outside my prediction range, I would regard that as evidence that something is going on which we don’t understand. Not proof, but evidence, and it might be a perturbation which isn’t persistent rather than a reversal of trend. Only 1 or 2 data points out-of-pattern makes “trend reversal” a tough call. If it’s 8 million km^2 or more, I would say definitely “something’s up.” As for “head this off at the pass,” I expect Watts & Co. will claim “recovery” even if we set a new record low for Arctic sea ice extent.]

• Ernst K

For me, evidence of a trend reversal would be a sustained (multi-year) increase in ice volume and thickness.

FWIW, my statistical model:

Sept average extent in million km2 = 7.44/(((year-1951)/65.3)^6.03 +1)

gives 2014 = 4.14 +/- 0.95

September ice areas show very little correlation with the previous year or even the early summer. It’s all about what happens in August (give or take a few weeks).

The only aspect of it that appears to be predictable is the long term trend, the rest is basically weather.

• J Garland

Actually, Tamino, I’ve been waiting for the contrarian blogosphere to start posting the headline “Tamino Predicts Continued Recovery Of Arctic Ice” since you put up your post. It’s a very simple misinterpretation/miscommunication, well within the powers of Tony, et. al.

• Mark

LOL – well, yes, perhaps the only “certain” prediction here is that no matter what the actual result the denialist/pseudo skeptic blogosphere will find a way to perceive it as evidence that AGW is false.

I appreciate the scholarly caution in Tamino’s answer to my question. Was just curious because in my days of Statistical Process Control metrology in the 80’s we used to have a rule of thumb that said that any 3 points in a row outside 2sd limits indicated a process was going out of control. 3 in a row above 2sd on the same “side” (e.g. 3 above) indicated an external shift – though 2 in a row above 3sd (i.e. outside the 3sd processs control limits) was also a pretty good indicator..

Of course a manufacturing process, whilst subject to varying influences, is not the same as an experimental observation in nature.

Even so the physics tells us we have a “naturally varying” process (due to how heat gets moved around by the climate system both seasonally and on a longer timescale) subject to an additional external forcing due to increased planetary heat retention.

So I was wondering if they could be treated analogously

[Response: I’d want to “do the math” before giving a definitive answer, but I expect so. Off the top of my head, the “rule of thumb” you describe sounds to me like a reasonable choice.]

• Horatio Algeranon

“any 3 points in a row outside 2sd limits indicated a process was going out of control”

“Out of Tolerance”
— by Horatio Algeranon

Three points in a row:
Spiraling out of control.
First three posts by Watts:
Out of tolerance lots.

• Horatio Algeranon

“Ice Recovery For Dummies ”
— by Horatio Algeranon

Cover the eyes
Recover the ice

3. B Buckner

NOAA predicts a positive anomaly for artic sea ice extent this summer through October. Not sure about the units, but I think this puts the minimum in the neighborhood of 6 million km2.

http://www.reportingclimatescience.com/news-stories/article/noaa-predicts-above-average-arctic-summer-sea-ice-extent.html

[Response: Their graph indicates a September value of about 6.6 million km^2, closer to 7 than to 6.

Predicting sea ice extent is one of the things that computer models are known not to do very well. It’s also hardly the “centerpiece” of this model forecast. And the web page describing their results clearly states “CAUTION: Seasonal climate anomalies shown here are not the official NCEP seasonal forecast outlooks.]

• Rattus Norvegicus

Another important point is that this forecast seems to run consistently high for Sept. minimums. Best taken with a mine of salt.

4. The last few months of average ice extent have regressed to the mean linear decline, so the safest bet for September is 5million km^2 or thereabouts.

However my “gut” says it will beat 5m, but by how much I do not know.

5. “beat” up or “beat” down?

6. 4.44 +/- 0.5
Compare present levels and background conditions to 2011, which superficially looks comparable, then take a wild guess. :)

7. Ron C.

The measure appears to be Sept. monthly average from NOAA.
I am looking at the past 7 years, 2007-2013 inclusive, since it looks like a regime change occurred after 2006.

For the last 7 years, the Sept. Extents and Daily Min were:

Year Sept Daily Min
2013 5.35 5.08
2012 3.63 3.37
2011 4.63 4.33
2010 4.93 4.60
2009 5.39 5.05
2008 4.73 4.58
2007 4.3 4.16

So the Sept. Average is about 0.255 higher.

Base on my analysis of taking the melt % from the March max, I estimate a range of NOAA Sept. minimums for 2014:

High 5.14

Medium 4.54

Low 3.29

Personally, it looks to me around 5 MKm2.

8. Of possible interest, our statistical evaluation of ensemble skill among 309 individual predictions submitted to the SEARCH Sea Ice Outlook over 2008-2013:

http://onlinelibrary.wiley.com/doi/10.1002/2014GL059388/abstract

• By the way, I wanted to check the possibility of making a meta-prédiction based on all predictions with a significant performance history. In theory, this should help to remove the bias and reduce the noise. Do you have the data in an machine readable format so I could play with them? We could work on this together if you like?

• Yvan, I had hypotheses along these lines when I started working with the SEARCH Sea Ice Outlook (SIO) predictions too, but the data point a different direction. In “difficult” years, actual extent falls outside even the uncertainty limits (variously defined) given with most predictions. In “easy” years on the hand, many are close and the median of predictions quite good. In one difficult year, statistical predictions perform slightly better; in another difficult year, modeling based predictions perform slightly better. But that might reflect that the statistical predictions tend to expect more decline than the modeling ones do. Looking at Figures 1 and 2 in our Geophysical Research Letters paper, you’ll see what I mean (and anyone can write me if they’d like a copy).

For those without access to GRL, there should be a more public-friendly article in the next few days, coming out in the ARCUS publication, Witness the Arctic (I’ll link it here). The new article contains something new that our GRL piece did not — a look at two office pool competitions (NCAR and NSIDC) that also tried to guess the minimum ice extent for each of the past 5-6 years. Results from these informal predictions show the same pattern we see in the SEARCH results, of distinctly easy and hard to predict years.

[Response: I’m very curious to find out who won the office pool competition!]

• L Hamilton

“Sea ice prediction has easy and difficult years” has just been published (free) in Witness the Arctic. Although I can’t seem to paste the link from my phone.

• Lawrence, easy and hard years only means that anomaly against the general trend is low or large. In short, this means that models have no predictive power if they cant catch this effect.

Still I am thinking it is possible to create a composite index that would perform better than the existing method. However, I would need the data to prove this.

9. Aaron Lewis

My guess is that the Arctic Sea Ice will be gone in 8 years. And it will mostly go in 2 or 3 horrible and terrifying ice loss events such as the summers of 2007 and 2012.

Just which years the ice goes, does not really matter. Albedo feedbacks are kicking in, and permafrost / clathrate CH4 sources are winding up.

For all longer term planning and policy the ice will be gone, This is one of those cases where there is some uncertainty about the near term, but the long term is settled.

The real question is: “How stable is the GIS?”. With large amounts of liquid water in the firn, and fjords under it, my guess is not very, once the sea ice goes. I expect to start seeing GIS ice loss events within 2 decades. Whoops, I am already wrong. We have seen several large calving events, starting in 2009.

Watch minute 64 of Chasing Ice. That is the GIS starting to go down the drain.

10. Paul

Curious about your choice of a quadratic. It bothers me because the limits are so wrong. [see Alex Dessler’s classic “On the impending disappearance of pluto” – http://photos1.blogger.com/blogger/7438/3478/1600/Plutoart.gif ]. Why not chose a function with 7.5 million km^2 at t=1970 and 0 km^2 at t=inf?

[Response: Keep in mind that I’m never extrapolating more than 1 year, on which time scale the difference between a quadratic and, say, a sigmoid curve isn’t enough to be notable. Others have applied “physically plausible” statistical models to estimate when Arctic sea ice will be all gone, but I have little confidence in extrapolation of curve fits beyond the briefest time span.]

• Ernst K

Tamino,

Your last chart goes to 2018 for the quadratic and Lowess extrapolations, so not quite “never”.

[Response: My purpose was to show that the two models do not diverge very quickly.]

The model I used (in a comment above) goes from 7.44 in 1951 to 0 at infinity for the reason you mentioned, but it makes almost no difference for the 1 year forecast (4.14 for me vs 4.13 and 4.15 for Tamino).

I also experimented with having an extra parameter for a minimum ice cover at infinity. Any value between 0 and 3 million km2 gives almost identical fits, but obviously drastically different long term forecasts. There’s simply not enough information in the historical record to make a good statistical forecast of the longer term fate of the ice.

• Ernst K

Sorry! Only the first part of the above comment is meant for Tamino, the rest is in reply to Paul.

• Horatio Algeranon

Technically speaking, if one assumes “a function with 7.5 million km^2 at t=1970 and 0 km^2 at t=inf”, one is no longer “extrapolating” to estimate future values but “interpolating.”

• Gerg

On what basis would you assume zero at infinity … and, presumably, non-zero at not-infinity? It’s not at all clear why year-to-year September sea ice extents should follow some neat decay process, sigmoid or exponential. Infinity is a long way. Without some further understanding, why reject the simpler alternative that the thing just winks-out? (Some mechanistic model projections do show a more gradual decay, but so far the record of those is rather poor.)

I’m with Aaron. A quadratic extrapolation of the PIOMAS all-months anomaly trace hits the September zero in 2020 … six years, not eight.

11. Thanks to Lawrence and Yvan for the links and the material. This suggests to me the limitations of existing ‘logical’ predictions methods; inevitably, probabilities dominate absent a fuller understanding of the causes of internal variability. Over a sufficient timescale, this will produce average results which probably meet statistically significant criteria, but on a year-on-year basis, predicting the unexpected is naturally more difficult.

What is the progress on understanding the teleconnections between global weather/climate patterns and Arctic Sea Ice conditions? Are we any closer to developing a ‘formula’ which permits at least a prediction of a given season’s ‘predictability’? Lots of other questions, but these will do for now…

12. Someone noted above extent doesn’t correlate well with anything much at all. I also agree what happens in August is key. That is why I never make a prediction till early July. Even then, it’s a fool’s errand, really. My predictions for 2011 fit much better for 2012, but I did no formal post on beyond posting in various places I thought new records likely. For 2013, I was even less formal; I parried a bit with Paul Beckwith on Facebook, concluding very little chance of new records.

Also, while extent is important for long term energy abroption in the Arctic, I think thickness and colume tell us much more about the state of the ice. If I recall, volime fell pretty steadily between ’07 and ’12 even as most of those years’ exrents were higher than ’07 and ’12.

I’m not a math guy. I look at trend, of course, but more so I look at the extent of what I call “cottage cheese” ice. I.e., if you use MODIS images you can look at the ice directly and see areas of round floes with grayish much in between. Where it looks white with primarily large diamond-shaped floes I assume rhe ice ismthicker. These observations correlate well with ice thickness graphs.

There’s a lot of cottage cheese this year.

Next, how are the winds setting up? Virtually every time I look at the winf graphs, here are high rates ofmice movement direcrly toward and through Fram Strait, creating along tail of ice down the East Greenland coast… like in 2007. In lower ice loss years that tail is shortened, generally.

Nexr, what is the Arctic Oscillation doing? It was mostly positive last winter, keeping air locked up to the north. This spring, however itmhasmtrended more to neutral, then significantly negative in June. If it stays negative, that should push estimates lower.

We also have an El Nino developing. Loking at graphs of El Ninos, they correlate fairly well with lower ice volumes inmfollowingnyears, primarily, but I’ve a sneaking suspicion this one will push some warm water/air close enough tomthe Arctic to enhance melt a little, then a lot next year.

Right now, I suggest two scenarios. If winds continue blowing ice to the Fram, an early and strong El Nino develops, the AO is primarily neutral to negative, then we hit maybe around 2010, possibly new second-lowest.

If none ofmthe above is true, the cottage cheese will probably still bring us 500k to 1M less in extent than last summer. I expect vulume will be in the 3-4M range.

• J Garland

One major problem with extent is that each grid square in which there is 15% or more ice coverage is counted as covered and each grid square with less than 15% is counted as noncovered. Therefore winds/currents have strong effects as they spread out and compress ice fields.

The real measure in the end would be ice volume or mass, but this information is far harder to get or doesn’t go back in time very far (PIOMAS/Cryosat).

13. toby52

pinotgraves,

By “beat 5m km^2″ I mean will be “less than 5m km^2″.