Nice work but you do know that correlation doesn’t mean causation, right??
Of course. It also has nothing to do with what’s going on here – this is a simulation using hypothetical data and processes, not an analysis of real data. It’s thus theoretical and descriptive. Correlation != causation does not apply in a simulation, as there are no cofounders or measurement error – the model is fully specified, and any errors are discrepancies between the model and a real system.
- How is the meta impacting your data? (I.e. class imbalance and what not)
The meta shows up in two ways in the rankings. First, you can think of your build as something that shifts your ‘true’ MMR strength – if you are playing a strong meta build, your true strength is higher than if you were playing an average build, while playing a weak build has you perform as if you had a lower strength. If you switch builds and classes from game to game, then your ‘true’ strength is actually changing from game to game as well, which makes your performance and your rating more volatile. I didn’t put anything like that into this simulation, because I’ve done so enough in the past to know it doesn’t really change anything – it makes the system a bit more volatile, and players ratings on average represent a combination of their MMR and the strength of their plate of builds.
The second way it impacts the model (or reality) is by making match outcomes more unpredictable. If everyone played identical builds then differences in skill and in-game performance explain the variance in outcomes, but when builds are involved differences in the match-up advantage also explain outcomes. This means that while a team may be favored to win, say, 70-30 by MMR, they are in reality only favored to win 65-35; in the worst case the build advantage is a coin toss and it drives real win rates towards 50-50. Over the course of a long simulation, this effect doesn’t really mean anything – it lowers the MMR spread (good players have slightly lower MMRs than their true strength, and weak players have slightly higher) by trivial amounts, and really only on the extreme tips where it’s hard to make matches for players.
- New players started with higher MMR than folks who tried to cripple their way to legendary/diamond.
I always throw this into the initial conditions. Under competitive matchmaking, even huge differences between MMR and real skill fix themselves within 100 games or so. The system in use here is not competitive matchmaking however, so it converges slower – really extreme differences in MMR and skill take 200 games or so to correct themselves. These are huge discrepancies, by the way, like having a 1100 MMR player (~25th percentile skill) being carried to 2800 MMR (3rd best MMR in the league) in the preseason – this corrects itself and that player is comfortably back at their MMR, glued to the bottom of diamond after being carried there.
Note that this ends up converging on true ratings faster than they would have during season 1 (and after sufficient games, ends up converging faster than ladder-less matches!) This works because the imbalanced matches, while not very fun, impose a pretty fast, rough sort of the player base amongst the divisions. After it’s done that, the imbalanced matchmaker actually sets up a lot of opportunities for upsets within each division – if you’re underrated given your MMR and division, it’ll feed you a steady diet of matches that it expects you to lose, and pairs you with teammates that are probably higher rated than you, statistically, so the carry effect is that much stronger.
-What are your independent and dependent variables?
There are no indepenedent or dependent variables, as this is a simulation of system dynamics, not a regression.
