Updated.
I've also started to poke around with the prediction algorithms again. Doing a field test today: the algorithm suggests that there should be between 1400 and 1746 posts today. The low number of posts on the 29th is weighting the data weirdly, but it's actually not affecting the prediction too much.
I've run the algorithm over the data for about a year and a half, and its margin of error is about +/- 225 posts, and an average of +/- 11% off. That's not really great yet. I'm still working on refining it; I'd really like to see the margin get down to +/- 100 posts.
Right now your algorithm is averaging the past
N daily differences to predict today's daily difference. I suggest adding a couple low-hanging-fruit improvements to see if they help. Statisticians call these control variables and data scientists call it feature augmentation.
Using the Sheets functions MONTH(date) and WEEKDAY(date), you can put the raw dates into categories. Work out the average effect for all dates in that category, subtract out those expected effects, do your prediction on the residuals, then add the expected effects back in. There is some art and some science to this.
I would expect the MONTH category to affect the average posting volume, but the WEEKDAY category might affect the average posting volume or the average difference.
Once you have one or two controls/features working, you'll find it much easier to test-drive others that might occur to you (release dates, FOTF replies, whatever).