Tim Sevenhuysen “Mag1c” of ElixersOracle.com talks with Mark Register about the basics of statistics in Esports
Mark Register – How do you decrease noise and increase signal to predict behaviors and outcome?
Tim Sevenhuysen – Yeah that’s always the question. Most of the noise from my experience in League of Legends stats comes from people’s interpretations.
It’s not that there’s data noise so much as the data is what it is but what you think it means changes a lot based on your preconceptions or when you watch the games you see this and therefore you think the numbers mean that.
It’s very difficult to do much that’s predictive in League of Legends which is where a lot of the from what I understand, I haven’t actually read the book that Nate Silver wrote that kind of uses that title or that concept but a lot of that I think is based on predictions which is where he became famous is predicting political outcomes based on all these different polling types, predicting sports outcomes.
In League of Legends it’s very difficult to do predictive things because the game itself, the landscape of the game is changing so often.
I do have one project that I’ve been quite proud of and I’ve been building up over time where I predict the outcome of the game at the fifteen minute mark.
Just saying what’s the status of the game at fifteen minutes, look at the gold differences, the dragon differences, a couple different things in the game, and say how likely is each team to win.
That’s the most predictive I get.
And that’s based purely on historical data over this current season cause the game changes so much season to season, half year to half year, that it’s really hard to look back to the previous year and say it says anything about this one.
But in the past when this situation has come up who has won?
I’ve just tried to distill that model down to the simplest possible pieces, like I said I have gold, I have dragons, and I have map side because historically one side of the map tends to win a little more than the other so that’s in there to control the outcomes and make sure that there’s nothing else going on that goes beyond the state of the game.
But lots of different people when they see my model they ask me, why didn’t you include this, why didn’t you include that, I think this would tell you something and I’ve thought about a lot of those things a lot of those different options but that introduces potentially a lot of noise because if you kill a tower it gives you gold.
So if you’re tracking the difference in gold and the difference in towers now you’re tracking the gold difference potentially kind of twice.
That doesn’t mean it’s going to really throw the model way out of whack because it’s a self contained model it can account for both of those things potentially but you don’t know exactly how it interacts when you got those things that are colinear, they’re both measuring somewhat the same thing even if there’s more to the towers than just the gold.
So you have examples like this where sure you could add more complexity to the model, sure it might tell you something more…but do you know what more it’s going to tell you?
Are you able to actually pinpoint there was this change in the outcome of the model, and I know it came from just this other thing that I added, or was there more to it, is there some other complexity that you’re not aware of.
A lot of what I try to do is just try to keep things as simple as possible.
Let’s start with the things that we think we do understand that are relatively simple that we can actually tell a story about and let’s not try to overreach beyond that and start painting kind of number pictures that we don’t really know how to interpret properly.