Overnight I realized that my graphs has a big flaw, the fact that the years are normalized with themselves can induce a lot of autocorrelation. For example, if it is the case that a low september minimum is completely random, that would still increase the ratio of normalized winter extent vs normalized september extent, because the low september extent would drag the average for the whole year down, thus increasing the normalized winter extent. Furthermore, since calculating the normalized extent requires knowledge of the average for the whole year, it is not possible to predict anything with it beforehand.

Therefore, I experimented with instead normalizing the years to their predicted average extent from a linear regression of all the years. Now, it should be truly neutral, and also you can make predictions. Unfortunately, doing this does reduce the correlation a lot,

*but it is still there*.

Now, we can make a prediction for 2020 based on the january value. The high extent compared to the ever-decreasing trend makes this year stand out a lot, the normalized january extent is an all-time high: 1.322. See the red area on the january graph. Will this mean the september extent will be very low like the graph suggests? Or does it mean the correlation will break down? If we trust the graph naively, the expected normalized september

~~minimum~~ average for this year is 0.42, which is

**4.30 Mkm^2** (which is third lowest of all time, behind 2012 and barely 2007), with a lower uncertainty bound of ~2.97 Mkm^2 and a high bound of ~5.33 Mkm^2.