Earlier this year I posted a prediction of the minimum arctic sea ice extent, in which I used a simple regression model to predict the average September extent. My final conclusion:

My final estimate for sea ice extent in September 2013 is 4.69 million square kilometres (95% CI: 4.06 – 5.32 million square kilometres).

On October 3rd the National Snow and Ice Data Center (NSIDC) released their estimate of the September extent, which was 5.35 million square kilometres. The linked blog post gives a nice description of the main reasons why the extent recovered, and some of the competing influences on the extent this year. It also explains some of the Antarctic extent’s record gains.

The final extent for September was just 0.03 million square kilometres outside of my 95% confidence interval, or 0.05 square kilometres outside my estimate as posted on the SEARCH September Sea Ice Outlook competition. This means that I only just missed including the observed value in my confidence interval. My prediction in July was 4th closest to the true value, beaten by NOAA, NSIDC themselves, and Barthelemy et al. It was the second closest estimate based on statistical estimation, and was beaten by two model-based estimates. My estimate was also the closest amongst all those that didn’t include the true value in their 95% confidence intervals. The lowest estimate was from Neven’s Arctic Sea Ice Blog, which just goes to show that crowd-sourced estimates aren’t necessarily the best. Watts Up With That were very close to me in June and August (they didn’t submit in July) at 4.8 million km2, but they didn’t give a confidence interval. I think this is the first time in the annual SIO competition that WUWT have come close to the mark, which just goes to show that eternal optimism has its value.

I’m pretty happy with my prediction. I originally planned to do it for the August submission adding July sea ice temperatures, and I think that would have probably bumped up my prediction a little closer to the true value. I also considered doing an ensemble model, using a wide range of different statistical models and averaging the results after incorporating more covariates. I think next year I will try to be more systematic, and submit a prediction for every month using a range of modeling techniques. The key point of my model is that it accurately predicted a very large rebound from the 2012 minimum based on just a few key variables selected without much systematic basis. I think I can do better next year!