Please not another SittingDuck...

ELO has a behaviour similar to Rolling Averages. It gradually forgets the past, so, past battles with retired bots stops affecting the ranking after some more uploads are made.

ELO was designed taking in account a changing environment. i.e. chess players improving their skills over time, new players, retired players...

The changing environment in this case is a bit different from chess though due to how there are many participants in each match, and it's known which participants that may have changed, versus not changed. This extra information could be leveraged.

I haven't looked deeply into it yet, but I imagine that it would be possible to construct a melee scoring system with greater immunity to this type of distortion, by considering the full list of partipants in each battle instead of tossing out information by splitting it into pairings. (Of course, the rumble client doesn't even give the necessary information to explore improved melee scoring systems, because the splitting into pairings is done at the client rather than the rumble server)

The current upload protocol is heavily ELO oriented. Splitting into pairings is how ELO is calculated in games with more than 2 participants.

I'm unsure if it was done that way with ELO specifically in mind... or if it was more "1v1 oriented" with melee as an afterthought.