I thought I'd throw together what I have
pre-ban.
Unfortunately, my hard drive that contained the original project file died, so I'll have to recreate it at some point.
For an outline of the process, it wasn't anything too crazy.
- I converted the OU usage data (again, 1760 ELO weighted) into .csv format for easier processing.
- I trimmed out all truly 0-usage Pokemon and NFEs, minus a couple (Chansey, Haunter, Pikachu, Clamperl, and Scyther). For whatever reason, I decided that Magnemite, Diglett, and Trapinch should just be trimmed, but maybe when I redo this I'll leave them in as well. This was mostly to filter out ultra-low usage things that I deemed to be statisically irrelevant.
- Then, over the period of analysis (Jan '25 to Aug '25 in this case), I did a trimmed mean, which is the average excluding the top and bottom outliers. I did this (of course) to decrease the impact of massive spikes in usage, attempting to get something more indicative of stable trends.
- Next, the natural log of this trimmed mean (plus 1 to handle usage values near and equal to 0) was taken to reduce the skew at high usage values (Tar, I'm looking at you).
- Then, I used this NTM (Normalized Trimmed Mean (normalized isn't exactly the right term here but oh well)) to create a matrix of distances between each Pokemon. These were hierarchically clustered using Ward's method, and 6 clusters ended up producing the best results by silhouette method, IIRC. It may have also just been what I thought looked the most reasonable qualitatively, I don't recall exactly.
- Oh yeah, "Gap" is the difference in NTM with the Pokemon above. The clusters (CL 6 column) have been renamed with the traditional tier names.
So, with that out of the way, here we go!
January 2025 - August 2025 Usage (1760 ELO Weighted):
View attachment 783168
View attachment 783169
In the spoiler are the images for the E cluster (the lowest cluster) in case you want to see them as well.
Yeah, big ol' table when you include everything down to Castform! Lol. I had to include it as separate images because I think the forum got mad at how massive it is. Not ideal, got a little squished here and there.
Here is a link to the spreadsheet.
I made a graphic that shows each cluster, down to Typhlosion (arbitrary stopping point) in the lowest cluster, the E cluster. All this graphic does is take the data from the table and make it easier to visualize in a traditional tier-slop format. The tiers are ordered, by
Rank.
6 Cluster Usage-based "Tier List":
View attachment 783167
Now, I'm not really claiming this to be exceptionally useful in any way. Just a little fun project for me to see what "natural" clusters/tiers arise from usage statistics. I might edit this post with some commentary, but I'm not a high-level player, so it would likely be less-than-useful. I guess one lighthearted comment I can make is that even trimming out their highest usage months, the Zoomer duo of Registeel and Raikou still make an appearance quite high, lending to the idea that if the trend continues to prove to be stable, they may be worthy of OU. And of course, we expect in the post-speedpass world that Jask, Wak, and perhaps even Smear and Vap will see some drops. I don't predict that Zap will see too much of a usage hit.
Well, that's that for now. I hope this proves somewhat interesting!
