31 August 2011

Too-early consideration of sea ice estimates

I won't let it stop me, but considering that it is still August as I write, it's premature to evaluate estimates of September's sea ice extent.  On the other hand, it's fun and easy, and some parts make sense even at this early date.

Last May I made, with colleagues, some estimates for September 2011's extent.  As usual, I also included some other forecasts for comparison.  The NSIDC figure for August 29th is:
Where I've added lines in red for the 30 year mean climatology of September extents (6.67 million km^2), the mean monthly September extents if we take a 30 year linear trend (5.31), and lines at 5.0 and 4.5 million km^2.

While not always true, it's generally the case that the September average extent is lower than the end of August level.  Our end (well, 29th) of August value is about 4.8 million km^2.  So we can be fairly confident that the September figure will come in not much higher than 4.8 (5.0 would be high, my intuition says), and probably some tenths below 4.8.  But probably not many -- look how little the curve changes across late August through September.

Some estimates:
30 Year Mean -- 6.67
Joe Bastardi -- 5.5
30 year Trend -- 5.31
Wang, Wu, Grumbine May -- 5.0
Wu, Grumbine, Wang May (December 2010) -- 4.8
Wu, Grumbine, Wang August (June 2011) -- 4.6
Grumbine, Wu, Wang May -- 4.4

... where I've struck through the estimates that are already (very likely) busted. The 5.0 of our first model estimate is probably too high. The 4.4 of our statistical guess is probably too low. Since I said that I thought this was their probable sense of error, I'm not too mortified. The two estimates which Wu took the lead on making both look plausible. 4.8 might be a touch high if the current figure is this. The 4.6 looks pretty good. The parenthetical dates are when the initial conditions for the model runs came from; we're not entirely up to real time for making our experiments.

The striking thing about the estimates that have already busted is that they're climatology (or someone known for making dramatic statements and doesn't like to see them checked, which is not striking but is worth noting if only because forecasts should be verified).  Normally, you expect climatology to be fairly good, not among the first forecasters eliminated.  Also, you expect, if climate is stable, that the mean will be a better forecaster than a trend (which results only from noise if climate is stable).  Instead, the mean is a horrible predictor.  And even the trend is very bad -- even assuming that ice is declining steadily is insufficient.  The ice extent is down even more than a linear trend.

Update:  For looks at the Arctic, including 3 days of sea ice drift forecast from NOAA/NWS/NCEP, see http://www.arctic.io/observations/520//


L. Hamilton said...

Put in terms of simple trends:
* The linear-decline prediction, a September NSIDC mean of 5.23, is already too high. Sea ice is declining at faster-than-linear rates.
* The quadratic-decline prediction, 4.63, looks much better -- but even so might be too high. NSIDC extent in late August fell below 4.7.
* The Gompertz-decline prediction I suggested in April, 4.44, still looks like a contender to me.
But we'll see.

L Hamilton said...

Results are coming in ...

I offered 4 predictions about September ice back in April based on simple Gompertz curves. Here's how those came out so far:

NSIDC extent -- predicted 4.4, observed 4.6
UB extent -- predicted 4.6, observed 4.6
NSIDC area -- predicted 3.1, observed 3.2
PIOMAS volume -- predicted 5.2, observed ___

(I think my PIOMAS estimate is going to end up rather high, but we'll know soon.)

Starting already with a guess about next year, over at Neven's:

And also a new graphic comparing the 5 major extent & area time series: