I should confirm Wipneus' and Yuha's comments. Wipneus' graph shows the difference for each day from the average for that day. Since recent years generally have less ice than earlier years, all recent years appear toward the bottom of his graph.

My graph shows the difference from the volume including the linear trend. Therefore the difference shown is computed from the expected ice volume in that year, and any year can appear at a positive or negative position on my graph. As Wipneus said, I could also have labeled my graph as showing the model residuals. In fact, I originally wrote my post using the word "residuals", until I saw that Tamino used "anomaly" for his graph. I changed my terms to match the language more commonly used here.

I showed every year since 1979 on my graph, but I only labeled the most recent years. The lowest line on the graph is for the end of 1981 and the beginning of 1982. If you look closely at the graph of the data and the model, you can see that the ice was far below the model from the summer of 1981 through the summer of 1982. Looking at the anomaly graph, the line for 1981 starts mixed in with the other years. It starts falling rapidly around day 125 (the beginning of May) and is the lowest line on the graph from day 200 (mid-July) through the end of the year. 1982 starts where 1981 ended. From there it rises, crossing the line for 1981 at about day 200 and returning to more typical anomaly levels by the end of the year.

In response to Epiphyte and others, the PIOMAS data is itself the output of a model, and so it is fair to ask if the PIOMAS data contains errors or biases. My response is that it doesn't particularly matter, as long as the bias is consistent. If it is true that PIOMAS overestimates volume during the early Spring in years with low volume, then as long as that overestimate is similar every year, comparing the data for different years is still valid.

Dorsetmetman asked when the model will reach an ice free level. I would expect the model to be reasonably accurate for short term predictions, but even then you need to take the variability of the anomaly into account. The model says this year's minimum will be 5.1 thousand cubic km. But the most recent anomaly was -0.5 thousand cubic km. If we stay at that level, the minimum will be 4.6 thousand cubic km. And the data is three weeks out of date, and we've all seen what's been happening to the ice in that time. If this year turns into a year like 2012, the anomaly at the minimum could be around -2.5 thousand cubic km, in which case the minimum will be 2.6 thousand cubic km. These all are plausible outcomes. Even if the anomaly in September this year is -2.5 thousand cubic km, the anomaly next September could be 0, in which case we'll get to go through another round of claims of a recovery.

I am very skeptical of the prediction accuracy for this model more than a year or two in advance. Using the model to predict when the Arctic will be ice free just isn't statistically valid, and there's no physical support for long term predictions either. (There is no physical support for this model at all! It's entirely statistical.) If you must know, the model currently reaches zero in 2032, 16 years from now. For comparison purposes, I rebuilt the model only using data through 2000, 16 years ago. The difference between the most recent data and the predicted value for that date is 3.8 thousand cubic km. In 2012, the error was as large as 7.15 thousand cubic km. The anomalies for the data fitted by the model can be quite large, but the anomalies for data outside that time period can be far worse.