My prediction for September average NSIDC Extent is:
5.15 million kmsq +/- 0.64 million kmsq
My method uses April PIOMAS sea ice volume for the Arctic. I had been toying with using only volume for the peripheral seas of the Arctic Ocean Basin (Beaufort through to Laptev) due to most of the increased volume remaining within the Central Arctic. But on balance I have decided to keep the method simple and avoid meddling. Last year undershot severely due to the 'heuristic' or judgement based, adjustment I made. I have made no such adjustment this year, apart from an adjustment for range to balance hindcast range and success rate.
PIOMAS April volume, calculated from PIOMAS gridded thickness data, is plotted against September NSIDC extent to derive a linear fit (R2=0.75), the relating equation
SeptExtent = 0.3278 * AprVol -1.16143
This equation is used to make central value predictions of sea ice extent for each year from 1979 to 2014. The standard deviation of the residuals is calculated, this is then multiplied by a constant, the result is added to/subtracted from the central value of each year's prediction to create the upper and lower bounds of the prediction. The constant is then varied until the success rate is around 80%, i.e. successful 4.5 years. Most years that the model fails, the failure is small.
The hindcast of the method is shown below.
How does this compare to other prediction methods? Looking at last year the predictions were as follows, the September 2014 average NSIDC Extent of 5.28M km^2 is shown as a horizontal red line.
Looking at last year's results: My range this year of 1.28M km^2 puts this prediction near the middle of the distribution of ranges, with 56% of the June 2014 SIPN predictions being below my range this year. Of the successful predictions 7 had wider ranges than my submission this year, 4 had smaller ranges, the smallest range for a successful prediction was the modelling based submission from NASA GMAO.
My prediction includes 6 out of the eight years from 2007 to 2014, which makes it rather a blunt instrument, but this is not unusual. Of the successful predictions last year Kapsch and Bosse have the same inclusion of post 2007 years, only NASA GMAO and Schroeder et al (CPOM) include less (4 years out of eight). Of the total 11 successful predictions 7 have a greater inclusion of post 2007 years in the range of prediction.
So in terms of the hindcast, and comparison with last year (for which my hindcast is successful), my prediction is within the range of what might be expected.
At the end of June I will use my previously suggested method that might predict large summer losses, and the CT Area prediction. Based upon these I might adjust this prediction by selecting a subset of quartiles from the range.