I'm hoping to do an ensemble model for August with some automated variable selection, but I'd like to add some measure of the variations in pressure over time - I saw in a previous post something about the arctic oscillation index ... could somebody tell me how to get that data for use in modeling?

Rob, sorry for taking time to reply. You're right, I'm only using summer variables at the moment, and extending the analysis to winter variables means taking a great deal more care in how I approach the modeling. I am going to generate a new prediction using an ensemble method, similar to that used in the Global Burden of Disease studies, using all the available data. I think I should include temperatures from Sep - June, and also the Arctic Oscillation index. I think an ensemble model will be more representative of the truth. I think it's possible that winter data will be less relevant now than it might have been 10 years ago, because the prevalence of FYI means that late summer phenomena are the key drivers of ice change. We'll see!

Hi Rob Dekker, glad you like my analysis. It's a pretty classical modeling method, really, and subject to all the usual terrible constraints of such a method. I'm thinking of trying a more flexible ensemble method for August. I used April and May because I couldn't get June and I just kind of thought that how hot it is now would be the most important temperature measure. I know nothing about sea ice, how it melts, the arctic or, for that matter, the northern hemisphere, so I just grabbed the first thing I had in mind. It occurred to me that I should use January extent instead of last year's september extent as a predictor, so maybe I should try winter temps. This is part of my ensemble thinking. I have a crap-ton of variables: realistically, all the extent and area values for each month since the previous Sep, all the temp vars since Dec, snow cover since Dec, year, plus volume. I know of a technique for ranking models and then combinining them, so I'm thinking of doing a massive ensemble using mixtures of these covariates. Unfortunately I'll probably have to use R to do it (shudder) but now I've prepared the data in Stata it should be easy ... I'll take your suggestion under advisement for August. May the best estimate win!

Hi everyone, long time lurker first time poster. I thought I'd try a predictive model this year. My prediction is 4.69 million square kilometres, 95% CI: 4.06 – 5.32 million square kilometres. Basis: a Prais-Winsten regression model using snow anomaly, temperature data and area and extent data for april, may, June and lagged value from the previous September. Details are here. My model suggests a huge recovery. It barely seems plausible, but my model has shown itself capable of identifying the two previous crashes, with even some skill at five years out. So it could be right! Here's hoping ...