Fall Major AI Predictions
Edit: I reran the ratings and simulations on this page with the brand new results from Rule One's last minute 1K tournament. Falcons are up a little. OCE lost quite a bit while Team Secret gained in turn. PWR would still be the favorite in a head to head rematch, but due to their worse seeding, they're more likely to go 0-3 now.
Edit 2: Ratings after day 1 available here. I did not get the video ready in time so I'll be finishing that once I'm back from the major. Sorry.
Edit 3: Retrospective available here.
Edit 4: Video about the AI finally here.
The Rocket League Fall Major in Rotterdam is happening this week, but my AI already knows how it will turn out ;)
What the heck am I talking about? I'm fashionably late with a video that I will
hopefully release on Wednesday or else Thursday after I'm back from the major. For worlds, I made an AI that predicted RL esports matches against JohnnyBoi, Gibbs, Lawler, Virge, and Wayton. I had $50 on each bet and it ended up going 3-1-1 in favour of the AI, netting me $100. The video will explain some of the details of how it works, so I apologize if this article leaves many things open. It wouldn't make much sense to upload the predictions for the Fall Major when it's already going on. Therefore, I'm releasing it now.
Keep on reading to learn all about
- the AI power rankings for all 16 teams
- the top 30 players at the Major
- the odds of each team getting Top 8, 4, etc.
AI Power Rankings
Rating numbers are relative numbers of skill, not absolute. The transparent portion of the bar signifies the uncertainty that the AI has about a teams rating. The opaque part of the bar signifies the conservative skill estimate of the teams. This estimate tends to be more accurate than the raw rating value that you get with the transparent bars. You can mouseover the rating to see the exact number.
The approximate odds in a best of five are:
- 50 rating difference: ~57% win
- 100: ~64% win
- 200: ~76% win
- 300: ~84% win
- 400: ~91% win
Chance of upset goes up for teams with > 100 uncertainty and down if it's a best of seven.
Since the AI ranking is purely results based, I'm not going to discuss every single team. I will, however, explain a couple of things that may look confusing to us mere humans.
Why the heck is GenG not #1 NA?
The AI doesn't care where you finish in a tournament. FaZe, SSG, and G2 haven't lost to low rated teams, and their swiss runs have been more impressive if you just sum them up. GenG is ahead in the playoffs, but we have to account for the difference in ratings at the start of the season. When a new roster forms, the AI estimates the initial skill. It uses the average of the estimated skill of the individuals that make up the new team. This method usually works much better than you might expect. In the case of FaZe, they immediately started out pretty close to the rating they have right now. GenG, however, is the one team where this method worked the worst this season (just like analyst predictions pre-season). Although Chronic and ApparentlyJack were the highest rated players on their previous teams, they had little recent success prior to teaming up. The team has gained around 200 points since their beginnings. That can take a while especially when you're losing to some lower rated teams in qualifiers and swiss.
NA > EU ?
You want a simple answer? Yes.
Oh, you're still reading? NA was already starting to pull ahead at the end of the last season. Their performances at the Winter Major, Spring Major, and World Championship were at the very least a little bit better than EU. Yes, EU always had a top tier team, but the average has been worse. They have a top tier team listed here again in Karmine Corp. Liquid and Oxygen also have the potential to pop off. Their ratings aren't far off; it just appears that way because NA's top is unbelievably close. For example, the first round swiss matchup G2 vs. OXG is predicted 63.7% in G2's favor. Not really a big gap.
Why has the small gap widened to a noticeable one? There are 3 reasons.
1) They've imported 2 full top 10 teams into the region. FURIA was top 6 when moving, and then ApparentlyJack, noly, and crr are all ranked as one of the 23 best players in the world right now. They've imported this much talent while stealing half of that from Europe. When you put it that way, it's not as surprising where the gap has come from.
2) Europe is also not sending its best teams. Don't forget that BDS went 3-0 in the first two swiss stages. They got upset by Quadrant, which shouldn't have happened tbf, but upsets do happen. In the second week, they had to play Karmine Corp in quarters. They played them as close as anyone did that weekend. People expect BDS to win every match, but just because they don't, doesn't mean they're not Top 5 EU anymore. It is very possible that if they had just switched bracket sides with Quadrant that week, they would've been the ones in the grand finals. Then, they would have qualified in the last regional while Quadrant, with their 1-3 swiss, would've been out. EG and G1 are also rated above Quadrant.
3) Those 2 reasons sound pretty great, but I wouldn't be honest if I hid that this last one makes the biggest difference. Europe might be underrated a bit because G1 and Moist signed up for an NA tournament just before the new season began. Playing on high ping, they lost to NA bubble teams which funneled points from EU to NA. I don't remove individual tournaments from the dataset, as all the past results have shown that this just makes predictions worse. Surprisingly, this also applies to tournaments like the Salt Mine Atlantic Draft with mixed rosters. Even in this season's case, Moist's adjusted ratings actually predict the results they've had in EU better than if I just remove the tournament. The difference to most tournaments is the one-sided ping disadvantage. I tried some things like removing just the EU matches of the tournament and pretending that the ping disadvantage makes them 300 rating worse. But I would have to manually flag it for situations like this, and I don't have any large dataset to calibrate it. I don't like solutions that require manual adjustment on a case by case basis, because it introduces human bias. That's why I'm just going to let the AI die on this hill for now. If it turns out to be very wrong, then the AI should still be able to adjust really quickly. I may then reconsider adding some fail-safe in the future. I will put the alternative results at the end of the article, in case you're interested.
James Cheese Good Cheese?
James Cheese has won 18 series in a row against APAC teams. This has netted them about 170 rating points. Undeserved? Maybe, but it should be noted that Kamii's individual rating puts him tied 4th with bananahead in terms of OCE players. That seems like a reasonable skill estimate for Kamii. So if that number isn't ridiculously inflated, then maybe neither is their team's. Even if you do subtract all 170 points, they would still be very close to Pioneers and a threat to any minor region team.
The Club is Overrated
Hey, don't blame me. That's what the unbiased AI thinks. People have been hyping them since the beginning of the season. The problem is that they're in a mickey region, so they should be winning every series in order to even be considered as good as the worst EU team. Now that I'm SAM #1 enemy, let me rephrase that. SAM is top-heavy. The Club lost a series to Backroom Enjoyers in the third regional. The rating difference between the teams was almost 600 points. That's a bigger upset than James Cheese losing to DeToNator. That's bigger than any upset that could theoretically happen at the major. The Club already had a rating similar to Quadrant but they lost it.
I have refrained from posting all 48 players because I do not want to single out players at the bottom and create drama. I want to be very clear about the fact that this ranking is stats based. It's a little more complex than just getting more points on average, but a full explanation could fill its own article so I'm not providing that right now. This rating is correlated with the actual skill of the players, but a correlation can in individual cases be very far off the truth. In most cases, a really good player will be moving to the top of the ratings, but just because someone isn't at the top, that doesn't mean they're bad.
Changes since Worlds
Firstkiller, vatira, BeastMode at the top is just business as usual. The addition of Lj to the SSG roster has caused Daniel's play to shift, and he has risen from his permanent Top 5-10 stay to nearly Firstkiller levels. More additions to the very top come in the from the team that no one quite saw coming. Chronic and Jack have gained 200 and 150 rating respectively since Worlds, and who can deny their quality. In the spring, Team Liquid were more balanced stats wise. Since then Atow has been separating himself more and more, causing his rating to rise at the cost of AcroniK's.
Great losses have been dealt to mainly Moist and Falcons. All of Moist are down almost equally; however, as illustrated in the previous section, they might have taken a bigger beating than deserved due to their high ping losses. They have lost their best player in vatira but the talent on the roster was expected to get better results than they did. Let's see if they can prove their current rating wrong at the Major. Falcons' trk511 was top 3 at their peak. That was gone after their World Championship performance already. The first regional then cost him even more as he didn't perform well while Falcons lost. After winning two more regionals, he's back to the top of his team, but they are all still below where they were at the peak. Comm has lost points in the last regional but also in other minor tournaments played with pickup teams. I would expect him to get back up to top 15 if V1 can perform.
SAM has new top individual contenders with Lostt and Aztromick. They're now at the highest rating they've ever been, overtaking the SAM/NA peers caard, CaioTG1, and ajg in the process.
G2 players seem to always be unseparable close to top. Business as usual for them. The other unmentioned players in the top 30 have also not moved a significant amount.
AI Simulated Major
|Name||Top 1||Top 2||Top 4||Top 8||Top 11||Top 14||Top 16|
This table is the result of 1,000,000 tournaments simulated, each with 40 series. For a total of 40 million series sampled from the AI probabilities, including game differential.
All that just to find that the highest rated team is the most likely to win it all. I'm joking. The interesting part is just how likely each team is to get somewhere, not the order. Though there are some interesting situations where seeding makes a big difference. Look at Liquid having a nearly 10% better chance than Oxygen to make top 8.
Other than that, this table better shows that it's really close in general. No matter who you pick for going 0-3, you're more often than not going to be wrong on the team because everyone has an above 50% chance of winning at least one series. On the other end the, same is true for Top 1/2.
Every NA team has a good chance of winning the entire thing. EU not so much if the AI is to be trusted. Well, if anyone can make a 1% run, it's Moist (0 fan bias here haha). Every NA team also has a 70%+ chance to get Top 8; however, some simple napkin math can tell you that the chance of all 5 making it at the same time is below 33%. The realistic expectation is 4/5.
TL;DR: NA wins.
I reckon most people will doubt that NA has a 74% chance of winning. I'm not too sure about it either, but I'm pretty sure they're favoured. As I said, I ran alternative simulations to make EU look better, and those still end in NA's favour in the most generous case. You can find them on this page.
Go ahead and let me know on Twitter how garbage the predictions are ;) Oh and or anything else that you liked/disliked about this post in general. Anything you'd like to see on the website etc. Got quite a few things planned. Make sure to keep an eye out, and maybe bookmark it. I will announce those also on Twitter and Discord when they come. I may provide daily rating/sim updates during the major if I can get it sorted and if people are interested. And then maybe a short summary after to compare the prediction performance of the AI with and without deleted matches.
Thanks a lot for reading.