I haven't looked at the source data to verify but could it be that the huge number of new RPG's created in 2017 actually cut off the older data (because this wouldn't be present in the first 100 pages, especially if there was an increase via bots?). Would it lead to a risk that all but the most longstanding RPG's would factor into your data?.

The 2017 spike is interesting.  One of my first rules of data is that if something looks off then it probably is.  However, indications of a bot invasion don't hold true either as if the data is predictable it would show a similarly sharp decline and that hasn't happened.  It looks like there has been a gradual decline since 2017, and whatever caused the 2017 spike also removed a lot of your earlier data from the sample.  I'd be inclined to run the script on 'all' data (think the site goes back to 2013) and see what patterns emerge.  I'd then be inclined to see if I could strip out the bot topic starters (maybe any post of 1 post only?) to see what the data looks like.

Another interesting thing to look at as a fellow data 'enthusiast' may be to identify topics started by board.  You mentioned that we can't measure RPGs run offsite or via pm but instigated / coordinated from here but you can measure the potential change in the number / distribution of 1-1 topics / interest checks.  If their variance changes differently from the other topics it may show a trend and a way of estimating the 'known unknown' of RPGs via pm.

It's terribly interesting!