|
|
I think Noxville did this at some point but it's interesting to use MC runs to check the effect of the format on the expected result. Also it's a nice way to see the approximate likelihood of some individual event happening given the initial assumptions. E.g. Liquid has the highest team ranking, and they end up winning the event 20% of the time. In Noxville's runs all teams also had a > 0.2 probability of not getting top8, even the top teams.
But the order of teams isn't all that interesting. When you just use elo to give the probabilities of individual match results, over a large enough run the average ranking of teams will just be the initial elo ranking right? Then it comes down to whether your ranking system makes sense, and it's hard to get it exactly right for the reasons you brought up too. The Bing team using some player level features like player weekly KDA (not sure if this takes into account pubs too) doesn't seem too sensible to me. It would be very interesting if one had the ability to take into account individual matchups between teams and how those end up skewing the expected results from purely elo based runs, but it's hard to have solid data on those. But I think if I wanted to make a bracket prediction manually the individual matchups play a pretty big role. Evaluating those matchups is pretty subjective though because there typically are fairly few recent games between teams and it more comes down to everyone's own view of how the teams draft and play and how it'll end up working out.
|
Wow that is just so much numbers and statistics. Usually I can understand this sort of thing, but wow this is a lot to take in. I can't imagine how long setting all these simulations up can take. It was definitely a cool read though and I am looking forward to see how your predictions pan out. Thank you for sharing this!
|
Very interesting how you went through the process. Start with ELO, compare to Glicko, adjust ELO since it was better and so on. Reminds me a lot about how one normally works, seeing that process instead of just the results is a lot more reading but also gives more content one can select the interesting bits from.
|
The process could have been better. we were initially going to do a really simple simulation so we could find out how many games each team would likely play, for other compendium predictions.
But we got a bit carried away. For future whenever we change from ELO to glicko, or just change the parameters in the system, we can run the model on previous games and see which settings predict future game results with the most accuracy. So rather than just choosing a setup because it 'feels right', we have mathematical proof it's the most accurate rating system we can think of
We also did a follow up for main event predictions just now.
https://medium.com/@STRATZ/the-international-2017-main-event-predictions-e32616de1a7
looks pretty bleak for lower bracket teams.
|
|
|
|