Big stats post here for playoffs matches:
I've been running a program that grabs scrim/match logs throughout the season and predicts the chance of teams taking rounds off one another for the purposes of match prediction.
The system is pretty simple, using all of a team's match/scrim history this season, it looks at the ratios of rounds won/lost of each team against one another and tries to determine a score that could be loosely interpreted as a 'skill level' for each team. If team A has score a and team B has score b then the chance of team A winning a round follows a logistic curve: 1/(1+e^(b-a)). To illustrate this, look at the following graph:
This allows you to get a sense of not only what the ranking of teams might look like, but also how much better than the other teams they are.
As of posting this, the team scores look something like this:
This would indicate that there is a fair gap between the top 3 teams and the other teams in advanced. This correlates relatively well with teams' positions in RGL, the team that I find most noticeably different is MySpectrumWiFi98-5G who seem pretty under rated by this metric. This might be because of underperformance in scrims, or because the team improved later into the season (something this isn't trying to evaluate).
Because this model is based on rounds, it can predict the outcomes of future games. However, there's a big problem with this as for whatever reason, under the RGL format it pretty much always expects games to go to winlimit. (This is because it's based on historical round duration, and I suspect because longer rounds tend to be cut off by going to timelimit).
In any case, here are predictions for tonights 3 games, which as I post this are probably near the end of the 1st half anyway (I'm a bit late to post this).
In theory these would be different for different maps, just because of the round durations, but because of the winlimit thing, it doesn't make much of a difference which map I choose. The chance of winning a round is probably different for different teams, but I don't know if I can easily estimate this, as the problem goes from having 16 dimensions to 121.