I've been working on a statistical model to try and rank the teams participating in this lan based on recent results. I managed to get some pretty nice results from this using the group stage data and some scrim results leading up to the lan, which I added just to give the model enough data to run.
https://i.imgur.com/msSuZir.png
The estimated team ratings are pretty similarly ordered to the seedings, which makes sense. The top 8 open group teams in general are a bit off (I had less data where they played each other). There's definitely a big gap between the bus crew and the other open teams based on match results, but comparison between open teams and the invite group isn't great. I had to choose some sort of starting rating for both the open and invite groups and I forced it so that the bus crew is on a par with the bottom of the invite group which I thought was a reasonable assumption. This might still be underrating the other open teams a lot.
Another cool thing about this model is that you can use it to predict the odds of different match results. At the moment, it can't tell if teams are better or worse on different maps (you'd need a lot more data than I've used for this), but it can predict the number of rounds that are going to happen on different maps (e.g. Granary tends to result in lower scoring games than Gullywash). I'm going to show some of the predictions for the 1st round of the playoffs and I'll probably post some predictions for the later rounds too. As this first round is all open vs invite teams, the predictions probably won't be hugely accurate for the reasons described above.
Probably the most interesting matchup is Corentin Mas Damage vs The Bus Crew. As mentioned above, I deliberately set this up so it would be pretty close.
https://i.imgur.com/QsPtv41.png
https://i.imgur.com/iMIBKtZ.png
The next closest matchup is Xypher vs top5sons. Because The Bus Crew confidently beat the other top 8 open group teams they played, the gap here is big, so this ends up being a lot less close. Because the map used for the prediction is Granary, there's a fairly high chance of a low scoring game.
https://i.imgur.com/DXpHROI.png
https://i.imgur.com/zyvogIn.png
The 1st vs 16th seed game prediction might be a little disheartening for THE OCEAN.
https://i.imgur.com/4PQyrta.png
https://i.imgur.com/Xp3DVuu.png
I'm going to try and put up predictions for some of the matches that are on stream later today (if I can wake up for them).
I've been working on a statistical model to try and rank the teams participating in this lan based on recent results. I managed to get some pretty nice results from this using the group stage data and some scrim results leading up to the lan, which I added just to give the model enough data to run.
[img]https://i.imgur.com/msSuZir.png[/img]
The estimated team ratings are pretty similarly ordered to the seedings, which makes sense. The top 8 open group teams in general are a bit off (I had less data where they played each other). There's definitely a big gap between the bus crew and the other open teams based on match results, but comparison between open teams and the invite group isn't great. I had to choose some sort of starting rating for both the open and invite groups and I forced it so that the bus crew is on a par with the bottom of the invite group which I thought was a reasonable assumption. This might still be underrating the other open teams a lot.
Another cool thing about this model is that you can use it to predict the odds of different match results. At the moment, it can't tell if teams are better or worse on different maps (you'd need a lot more data than I've used for this), but it can predict the number of rounds that are going to happen on different maps (e.g. Granary tends to result in lower scoring games than Gullywash). I'm going to show some of the predictions for the 1st round of the playoffs and I'll probably post some predictions for the later rounds too. As this first round is all open vs invite teams, the predictions probably won't be hugely accurate for the reasons described above.
Probably the most interesting matchup is Corentin Mas Damage vs The Bus Crew. As mentioned above, I deliberately set this up so it would be pretty close.
[img]https://i.imgur.com/QsPtv41.png[/img]
[img]https://i.imgur.com/iMIBKtZ.png[/img]
The next closest matchup is Xypher vs top5sons. Because The Bus Crew confidently beat the other top 8 open group teams they played, the gap here is big, so this ends up being a lot less close. Because the map used for the prediction is Granary, there's a fairly high chance of a low scoring game.
[img]https://i.imgur.com/DXpHROI.png[/img]
[img]https://i.imgur.com/zyvogIn.png[/img]
The 1st vs 16th seed game prediction might be a little disheartening for THE OCEAN.
[img]https://i.imgur.com/4PQyrta.png[/img]
[img]https://i.imgur.com/Xp3DVuu.png[/img]
I'm going to try and put up predictions for some of the matches that are on stream later today (if I can wake up for them).
when does the team with clockwork play?
when does the team with clockwork play?
How Many Timelines Are There That Faint Win
How Many Timelines Are There That Faint Win
[quote=Iatgink]How Many Timelines Are There That Faint Win[/quote]
[img]https://metro.co.uk/wp-content/uploads/2019/06/SEI_73607257.jpg?quality=90&strip=all&crop=42px%2C0px%2C518px%2C272px&resize=1200%2C630[/img]
2345ywhen does the team with clockwork play?
we play soon dont worry
[quote=2345y]when does the team with clockwork play?[/quote]
we play soon dont worry
tell us a bit more about the model?
tell us a bit more about the model?
zxptell us a bit more about the model?
It's a round based logit model. Basically there is some unknown skill level for each team, where the chance of team a winning a round vs team b is 1/(1+exp((score_b-score_a)/sigma))). The skill levels for each team are estimated by the rounds each team wins/loses against one another. The duration of rounds is also assumed to be gamma distributed (+ some minimum round time). The parameters of this gamma distribution is assumed to be different for each map.
I estimate all of the parameters for the model at the same time using MCMC sampling. You can then forward model different scenarios for games by simulating rounds from these chances and stopping when you reach 30 minutes/win difference 5.
Here are the predictions for the Ascent EU / BIGBLOKELAN game. https://imgur.com/a/s9IPMMh
[quote=zxp]tell us a bit more about the model?[/quote]
It's a round based logit model. Basically there is some unknown skill level for each team, where the chance of team a winning a round vs team b is 1/(1+exp((score_b-score_a)/sigma))). The skill levels for each team are estimated by the rounds each team wins/loses against one another. The duration of rounds is also assumed to be gamma distributed (+ some minimum round time). The parameters of this gamma distribution is assumed to be different for each map.
I estimate all of the parameters for the model at the same time using MCMC sampling. You can then forward model different scenarios for games by simulating rounds from these chances and stopping when you reach 30 minutes/win difference 5.
Here are the predictions for the Ascent EU / BIGBLOKELAN game. https://imgur.com/a/s9IPMMh
Predictions Ascent/Faint https://imgur.com/a/BC7KsG4
Nothing better than seeing all those hours writing reports being put to use
Nothing better than seeing all those hours writing reports being put to use
Ha. Statistics are great and all, but why don’t we really dive deep into the biology of tf2 lan matches? Let’s start from the beginning
The mitochondria is t
Ha. Statistics are great and all, but why don’t we really dive deep into the biology of tf2 lan matches? Let’s start from the beginning
The mitochondria is t
[url=https://imgur.com/a/395HZpH]Ora/top5 predictions[/url]
[url=https://imgur.com/a/4LM22XW]Se7en/Ascent.NA predictions[/url]