WEBVTT - Why Soccer Analytics Works Like Volatility Arbitrage Trading 0:00:02.720 --> 0:00:14.000 Bloomberg Audio Studios, Podcasts, radio News. 0:00:18.520 --> 0:00:21.680 Hello and welcome to another episode of the Odd Lots Podcast. 0:00:21.760 --> 0:00:23.880 I'm Joe Wisenthal and I'm Tracy Alloway. 0:00:24.200 --> 0:00:25.040 Racy, I have a question. 0:00:25.280 --> 0:00:27.640 I know you spent a lot of your youth overseas 0:00:27.800 --> 0:00:30.920 my youth. Did you ever go to many baseball games 0:00:30.960 --> 0:00:31.400 as a kid. 0:00:31.920 --> 0:00:34.440 Yeah, so I was in Chicago for a few years, 0:00:34.479 --> 0:00:36.400 so I went to the Cups games, and then I 0:00:36.479 --> 0:00:40.480 was in Japan, and the baseball scene in Japan is amazing, 0:00:40.600 --> 0:00:43.800 like the best vibes of a live sports event that 0:00:43.960 --> 0:00:46.320 I've ever ever witnessed. Her encounter. 0:00:46.440 --> 0:00:48.159 I was just talking to someone about this last night. 0:00:48.159 --> 0:00:51.080 I've always wanted to go to a Japanese baseball record. 0:00:51.159 --> 0:00:54.800 That's like, if we ever do a live show in Tokyo, 0:00:55.440 --> 0:00:57.840 let's try to schedule it during baseball season, because it 0:00:58.040 --> 0:01:00.880 is like I sort of think that's a bucket list thing. 0:01:01.160 --> 0:01:03.600 But anyway, the reason I asked this question is I 0:01:03.640 --> 0:01:06.120 have this really vague memory as a child going to 0:01:06.160 --> 0:01:09.200 see Detroit Tigers games with my grandfather, like probably when 0:01:09.200 --> 0:01:11.400 I was like maybe these memories are probably from when 0:01:11.400 --> 0:01:14.000 I was younger than like six or five, but there 0:01:14.080 --> 0:01:17.160 used to be a non trivial number of people who 0:01:17.160 --> 0:01:20.000 would go to the games and they would keep score, 0:01:20.080 --> 0:01:23.600 and they would write down every single a bat and 0:01:23.640 --> 0:01:25.240 they the outcome of every single. 0:01:25.160 --> 0:01:27.560 Your I've seen that, not in person, but like maybe 0:01:27.720 --> 0:01:28.920 movies or something like that. 0:01:29.000 --> 0:01:31.240 It doesn't really happen anymore, like I never see it. 0:01:31.280 --> 0:01:33.440 There are may be a few, like old timers who 0:01:33.480 --> 0:01:36.080 still has a hobby or habit do that, but it 0:01:36.160 --> 0:01:39.240 was like a non insignificant number of people. And it's 0:01:39.280 --> 0:01:40.959 interesting to me. You know, I've been to a couple 0:01:41.000 --> 0:01:44.880 of soccer games this year. There's no equivalent way you 0:01:44.920 --> 0:01:47.960 could do that, right, because it's like baseball is filled 0:01:47.960 --> 0:01:50.720 with all of these discrete events. The picture, who is 0:01:50.720 --> 0:01:51.920 the picture? Who is the batter? 0:01:52.680 --> 0:01:52.880 Hit? 0:01:53.120 --> 0:01:56.240 Not a hit, single, not a strikeout, walk, et cetera. 0:01:56.400 --> 0:01:58.559 Like what would even be the equivalent to soccer? 0:01:58.720 --> 0:02:01.360 So this has been a long running debate in soccer. 0:02:01.400 --> 0:02:03.960 And I remember when moneyball came out, and you know, 0:02:04.040 --> 0:02:07.160 sports analytics became a big thing, especially for baseball because 0:02:07.200 --> 0:02:09.680 as you point out, it's these sort of discrete events 0:02:09.720 --> 0:02:13.080 that have a lot of statistics embedded in them. A 0:02:13.120 --> 0:02:15.519 lot of people were saying that soccer. You could never 0:02:15.680 --> 0:02:18.040 use data analytics in the same way for soccer, Like 0:02:18.080 --> 0:02:20.919 it's too chaotic, it's too fluid, there's too many variables, 0:02:21.000 --> 0:02:24.440 there's not enough goals. That complaint comes up a lot 0:02:24.480 --> 0:02:26.880 whenever we talk about are we going to say soccer 0:02:26.960 --> 0:02:28.480 or football? By the way, you know what, let's just 0:02:28.480 --> 0:02:31.280 say soccer. Okay, all right, I'll try. 0:02:31.440 --> 0:02:33.800 We can say football. I actually don't feel strongly about 0:02:33.800 --> 0:02:34.079 this one. 0:02:34.200 --> 0:02:39.160 But that said, we do see soccer analytics on the rise, right, 0:02:39.240 --> 0:02:41.800 Like now we get all these stories about like tiny 0:02:41.919 --> 0:02:46.160 clubs that are using data to source you know, specific 0:02:46.160 --> 0:02:51.200 players in very moneyball style. Always we obviously have prediction markets. Yeah, 0:02:51.280 --> 0:02:53.400 people are doing a lot of sports betting, and so 0:02:53.520 --> 0:02:56.840 all the analytics seem to be like becoming more important. 0:02:56.840 --> 0:02:59.040 And I will just say, I've read this crazy stat 0:02:59.400 --> 0:03:02.400 right before we came on from the law firm Morgan Lewis, 0:03:02.560 --> 0:03:05.040 and they were saying, in the twenty twenty six FIFA 0:03:05.240 --> 0:03:10.480 World Cup match based data is going to be something 0:03:10.680 --> 0:03:13.839 like so it's one hundred and four matches generating more 0:03:13.880 --> 0:03:18.440 than ninety petabytes of data, yeah, which is a forty 0:03:18.480 --> 0:03:21.839 five fold increase over the volume produced in the last 0:03:21.880 --> 0:03:24.640 World Cup in twenty twenty two, you know, stunning. 0:03:24.720 --> 0:03:26.840 So this stories is a question, and I know we're 0:03:26.840 --> 0:03:28.720 going to get to this in the conversation, and it 0:03:28.800 --> 0:03:33.000 almost is like a philosophical question, which is, okay, we 0:03:33.040 --> 0:03:35.360 see the game of soccer is like very fluid, right, 0:03:35.400 --> 0:03:37.640 there's no we just said there's a fewer discrete events, 0:03:37.640 --> 0:03:41.720 like in theory, is something that's fluid a series of 0:03:41.760 --> 0:03:46.240 like microscopic discrete events. You know, could you get a 0:03:46.360 --> 0:03:50.160 million frames per second and actually turn it into discrete events? 0:03:50.200 --> 0:03:53.000 Like that is sort of like an interesting question to mean. 0:03:53.320 --> 0:03:56.520 A parallel that I think of in this conversation is 0:03:57.520 --> 0:04:00.280 chess is like all discrete events, and that it's seen 0:04:00.320 --> 0:04:02.520 it's like a very difficult thing to crack. But then 0:04:03.080 --> 0:04:04.440 like over twenty five years ago. 0:04:04.400 --> 0:04:05.920 Now it's like, yeah, of course. 0:04:05.800 --> 0:04:07.800 They cracked it. But then you would think, okay, well, 0:04:07.800 --> 0:04:10.280 like what is the opposite of chess, which would be language, 0:04:10.320 --> 0:04:12.600 And you say, okay, you can't. That's fluid, that's all 0:04:12.640 --> 0:04:15.880 over the place, And yet computers seemed to be understanding 0:04:15.960 --> 0:04:16.839 language pretty good. 0:04:17.000 --> 0:04:21.120 My inner luod Iite says, there must still be a secret, 0:04:21.240 --> 0:04:24.800 like unmodel leable you can't talk when it comes to 0:04:24.839 --> 0:04:28.120 like football, the beautiful game. But you're right that technology 0:04:28.200 --> 0:04:29.600 may prove me wrong very quickly. 0:04:29.800 --> 0:04:33.880 Well, as you mentioned, you know, soccer analytics growing, and 0:04:33.960 --> 0:04:37.240 obviously the interest now is for obvious reasons. There's all 0:04:37.279 --> 0:04:40.000 the betting on the World Cup. People talk about the 0:04:40.200 --> 0:04:43.080 x G of like I don't know a situation or 0:04:43.120 --> 0:04:44.680 a player the expected goals. 0:04:44.880 --> 0:04:47.400 Now we have body posing analytics as well, which. 0:04:47.320 --> 0:04:51.720 Yeah, and when our producer Kale introduced me, I hadn't 0:04:51.760 --> 0:04:55.640 seen them before earlier this year, like the momentum charts 0:04:55.760 --> 0:04:59.680 that show like you know how dominant a team is 0:04:59.680 --> 0:05:02.120 in given moment, you see it going waves and stuff 0:05:02.120 --> 0:05:04.200 like that. So there clearly is a lot of work 0:05:04.240 --> 0:05:06.479 being done. I don't know how much it works. I 0:05:06.480 --> 0:05:10.360 don't know how much like momentum charts consistently predict who 0:05:10.440 --> 0:05:12.840 is going to be the winner. But just this question 0:05:12.960 --> 0:05:17.839 of like the ability to model deeply fluid things the 0:05:17.880 --> 0:05:20.880 beautiful game, like it's an art, right, Like can computers 0:05:20.920 --> 0:05:23.800 actually model this? And then like if they can, what 0:05:23.839 --> 0:05:25.680 does that say about the ability to model a bunch 0:05:25.680 --> 0:05:28.279 of other things? Strikes me as just like a very 0:05:29.320 --> 0:05:31.320 relevant question right now, because it's the World Cup, but 0:05:31.360 --> 0:05:34.839 also relevant question in general about market for markets and 0:05:34.880 --> 0:05:39.320 finance exactly. So I'm curious to know like how it works, 0:05:39.920 --> 0:05:43.279 about how the moneyball revolution, where we are in the 0:05:43.400 --> 0:05:46.080 arc with soccer anyway, I'm really excited to say we 0:05:46.160 --> 0:05:49.200 really do have two perfect guests because they do sit 0:05:49.360 --> 0:05:52.000 right in this space and also at the intersection of 0:05:52.040 --> 0:05:54.320 everything that we're talking about. We're gonna be speaking with 0:05:54.560 --> 0:05:58.400 yours Becker's. He is a soccer analytics consultant, the professional 0:05:58.440 --> 0:06:02.120 soccer Analytics consultant for the decade. As well as Mike Tracy, 0:06:02.240 --> 0:06:05.240 he's the head of risk at Apex Fintech Solutions. He 0:06:05.320 --> 0:06:08.560 used to be a volatility arbitrage trader at Peak six 0:06:08.760 --> 0:06:11.920 and he also does soccer analytics for the club FC. 0:06:12.120 --> 0:06:15.240 So literally the two perfect guests to talk about some 0:06:15.400 --> 0:06:18.440 of these questions that are arising right now. So Mike 0:06:18.520 --> 0:06:20.560 and yours, thank you both so much for coming on 0:06:20.600 --> 0:06:21.200 the podcast. 0:06:21.560 --> 0:06:24.160 Thank you for having us is a great introduction. Yeah, 0:06:24.200 --> 0:06:24.679 thank you. 0:06:24.680 --> 0:06:27.400 You know, Mike, let me start with you question. A 0:06:27.480 --> 0:06:32.040 lot of people I think, have, for good reason an 0:06:32.120 --> 0:06:37.200 intuitive feel that like sports analytics and trading markets are 0:06:37.240 --> 0:06:40.720 adjacent ideas that like, we know this, we know that 0:06:40.760 --> 0:06:44.200 the a lot of the prop shops and market makers 0:06:44.279 --> 0:06:47.720 do both and so forth. And how would you articulate 0:06:47.800 --> 0:06:52.120 based on your career the sort of overlap between the 0:06:52.240 --> 0:06:56.600 skills and techniques you've developed as an arbitrage trader, a 0:06:56.680 --> 0:07:00.760 volatility trader, and the overlap of skills that they're like, Okay, 0:07:00.839 --> 0:07:03.800 this thing that we'll get into that we call soccer analytics. 0:07:04.279 --> 0:07:07.279 Well, soccer is unique because I'd say every aspect of 0:07:07.320 --> 0:07:12.200 the game is a distribution. When you look at on pitch, 0:07:12.800 --> 0:07:16.960 the performance by a team, when you look at performance 0:07:17.000 --> 0:07:21.200 by an individual player, when you look at the seasonal outcomes, 0:07:21.440 --> 0:07:26.120 and this unique component that is promotion relegation. So your 0:07:26.200 --> 0:07:31.680 finances are highly variant year over year. And so there's 0:07:31.720 --> 0:07:37.720 a lot of overlap between volatility trading and working in 0:07:37.840 --> 0:07:43.720 soccer in that you are making highly levered bets on 0:07:44.800 --> 0:07:50.120 often imperfect information. Then that's not necessarily predictive like other 0:07:50.160 --> 0:07:51.280 sports such as baseball. 0:07:51.600 --> 0:07:54.280 Can I ask a very basic question, which is what 0:07:54.400 --> 0:07:57.800 is the point of soccer analytics? So if we go 0:07:57.880 --> 0:08:00.520 back to the market's analogy, we talk about price disco right, 0:08:00.600 --> 0:08:02.200 like you're trying to find the right price for a 0:08:02.200 --> 0:08:07.080 particular asset. With soccer, are you trying to price players, 0:08:07.440 --> 0:08:12.160 improve the training, make more successful predictive bets? I imagine 0:08:12.160 --> 0:08:13.400 it's a bunch of different things. 0:08:13.720 --> 0:08:16.080 Yeah, i'd say everything. I think you said what are 0:08:16.120 --> 0:08:18.280 the things you're trying to do? And I think from 0:08:18.440 --> 0:08:22.920 an investor operator perspective, it's what are you trying to 0:08:22.960 --> 0:08:26.240 not do? You're not trying to get relegated and you're 0:08:26.280 --> 0:08:29.840 not trying to spend thirty million pounds on a player 0:08:29.960 --> 0:08:32.280 that is going to be terrible and you'll have to 0:08:32.320 --> 0:08:34.079 get rid of in two years yours? 0:08:34.120 --> 0:08:36.880 What don't you talk about from your perspective? I mentioned 0:08:36.920 --> 0:08:39.960 your soccer analytics pro actually are a producer. By the way, 0:08:40.040 --> 0:08:42.440 just put in the chat the Bloomberg style guide is football, 0:08:43.480 --> 0:08:45.560 which is injury. Maybe we'll just switch to Yeah, let's 0:08:45.559 --> 0:08:49.440 stay in a company style. You're a football analytics pro 0:08:49.960 --> 0:08:52.080 tracy as like, what is the point? Is it more 0:08:52.120 --> 0:08:55.240 on the team side of things in terms of identifying talent, 0:08:55.280 --> 0:08:58.040 et cetera, or is it more on I don't know, 0:08:58.080 --> 0:09:00.800 I guess like the sort of the best side. What 0:09:00.960 --> 0:09:04.120 is the problem you and your professional capacity are trying to. 0:09:04.040 --> 0:09:07.760 Solve, So it highly depends on the organization. Right, some 0:09:08.040 --> 0:09:12.160 organizations might want to like mic sts, make good hires, 0:09:12.320 --> 0:09:16.040 make sure that they don't lose millions of pounds. Some 0:09:16.200 --> 0:09:20.120 organizations might use soccer analytics to try and improve their strategy, 0:09:20.440 --> 0:09:23.679 but they might look at like in game data and 0:09:23.800 --> 0:09:26.400 try to figure out if there's some optimization that they 0:09:26.440 --> 0:09:29.360 can make based on like where the passes should go 0:09:29.520 --> 0:09:32.960 or where players should be, what play styles make more sense, 0:09:33.040 --> 0:09:36.360 or what yields more expected goals if you will, And 0:09:36.400 --> 0:09:38.960 then I guess the third one is entertainment. There are 0:09:38.960 --> 0:09:41.200 a lot of apps out there, websites out there that 0:09:41.440 --> 0:09:46.280 provide fans with some sort of information, and even back 0:09:46.320 --> 0:09:47.960 in the let's say in the nineties, they would have 0:09:47.960 --> 0:09:51.679 the overlays on the TV where they show number of corners, 0:09:52.200 --> 0:09:56.680 number of yellow cards, and possession percentage. It doesn't mean anything, 0:09:56.800 --> 0:09:59.679 but it's it's interesting. So you can only imagine that 0:09:59.720 --> 0:10:01.760 you can go much deeper with that and people will 0:10:01.760 --> 0:10:03.720 still be engaged. 0:10:03.840 --> 0:10:05.680 Well, if you just say more on tho you said 0:10:05.720 --> 0:10:09.600 it doesn't mean anything, Is that something that has always 0:10:09.640 --> 0:10:12.880 been understood or is this something in twenty twenty six, 0:10:13.200 --> 0:10:15.640 we could say that a few of these stats that 0:10:15.679 --> 0:10:18.480 they like to put on the TV were just extremely 0:10:18.600 --> 0:10:20.959 like low signal data point. 0:10:21.080 --> 0:10:23.720 Yes, so when I say they don't mean anything, I 0:10:23.720 --> 0:10:27.000 don't think those were necessarily predictive of the outcome of 0:10:27.040 --> 0:10:30.080 the match. Got corners might because it's an indicator of 0:10:30.679 --> 0:10:33.200 which team it has the overhand in a game, but 0:10:33.280 --> 0:10:37.239 I'm quite sure most of the others, like possession percentage, 0:10:37.400 --> 0:10:40.280 don't mean all that much. I'm trying to look at 0:10:40.480 --> 0:10:41.240 game outcomes. 0:10:41.520 --> 0:10:44.000 Wait, so we touched on this in the intro. But 0:10:44.200 --> 0:10:47.640 like the perceived wisdom in the sort of early two 0:10:47.679 --> 0:10:51.280 thousands was that soccer was far too complicated to be 0:10:51.360 --> 0:10:56.640 given the baseball moneyball style treatment. What actually changed to 0:10:56.679 --> 0:10:59.000 get us to this point where we're not just talking 0:10:59.000 --> 0:11:01.839 about things like expect goals, but we're also talking about 0:11:01.880 --> 0:11:05.600 like body movements and posturing and things like that. 0:11:05.960 --> 0:11:08.120 You know, baseball had moneyball in two thousand and three, 0:11:09.040 --> 0:11:12.160 and soccer had two things in the twenty ten. So 0:11:12.200 --> 0:11:15.760 in twenty thirteen, Chris Anderson and David Sally published this 0:11:15.840 --> 0:11:19.720 book called The Numbers Game that distilled soccer down to 0:11:20.200 --> 0:11:23.880 more of a weakest length game. And then the other 0:11:24.320 --> 0:11:26.840 thing that happened was I would say we had a 0:11:26.840 --> 0:11:31.880 bit of this revolution on Twitter, which you kind of 0:11:31.920 --> 0:11:36.240 spoke about recently, Joe, where ideas were incubated and it 0:11:36.960 --> 0:11:39.480 I give Michael Kayley a lot of credit for this, 0:11:39.600 --> 0:11:42.400 He's he's been on the pod. But yeah, every Saturday 0:11:42.600 --> 0:11:45.360 in the Premier League you would have six matches and 0:11:45.400 --> 0:11:48.120 then you would wait thirty minutes and Kaylee would just 0:11:48.480 --> 0:11:52.560 post a stream of every expected goals chart and it 0:11:52.679 --> 0:11:57.120 started to stimulate this discussion of does this represent what 0:11:57.160 --> 0:12:00.600 should have happened, what should have been the outcome? I 0:12:00.640 --> 0:12:03.680 think it started to lead us down this path of 0:12:04.360 --> 0:12:08.080 where is x G flawed? It only registers when you 0:12:08.160 --> 0:12:11.360 have a shot, and so there is much more to 0:12:11.440 --> 0:12:17.400 the game than simply which shots occur. There's possession and 0:12:18.040 --> 0:12:22.319 the threat of each possession. There's match momentum, and there's 0:12:22.480 --> 0:12:28.160 changing styles. So it's evolved from on ball data to 0:12:28.800 --> 0:12:32.360 tracking data to now we're going to that deeper layer 0:12:32.520 --> 0:12:36.680 of body pods and what movements can prevent or create 0:12:37.040 --> 0:12:38.120 opportunities to score. 0:12:38.440 --> 0:12:42.720 I think what goes heading out with that is when 0:12:42.760 --> 0:12:46.079 I started ten years ago, soccer football was perceived as 0:12:46.200 --> 0:12:48.760 two complex There were twenty two players, there was a ball. 0:12:49.160 --> 0:12:53.600 People are doing interesting analytics research and applied research already 0:12:53.760 --> 0:12:57.480 in basketball, and what was said as well, there's five 0:12:57.520 --> 0:12:59.959 players on each team, so it's a lot less complex. 0:13:00.120 --> 0:13:03.400 There's more games, and we have more data, and there's 0:13:03.400 --> 0:13:07.840 more scoring, so you could derive metrics a lot easier. 0:13:08.360 --> 0:13:09.080 And I think the. 0:13:10.559 --> 0:13:15.400 Evolution in just how AI was applied in data availability. 0:13:15.559 --> 0:13:18.840 So going from like Mike said, only on ball events 0:13:18.880 --> 0:13:21.920 where we know which player made the pass, which player 0:13:21.960 --> 0:13:23.480 made the shot, but we always have to say, well, 0:13:23.520 --> 0:13:25.160 we don't know where all the other players are because 0:13:25.200 --> 0:13:28.640 those are not recorded. And now we have the ability 0:13:28.720 --> 0:13:33.280 to do make highly complicated artificial intelligent neural nets with 0:13:33.440 --> 0:13:35.959 this positional tracking data where at ten trains per second 0:13:36.000 --> 0:13:37.960 or twenty five points per second, we know where all 0:13:37.960 --> 0:13:40.040 the players are, and we know where the ball is. 0:13:40.600 --> 0:13:42.840 And then additionally, at some point we will get the 0:13:42.880 --> 0:13:45.720 full body posts we know where like the whole skeleton 0:13:45.720 --> 0:13:48.560 of all players are. Basically like that sort of went 0:13:48.600 --> 0:13:50.680 hand in hand and it felt like it was inevitable. 0:13:50.720 --> 0:13:53.760 But also the game is just more complex, so it 0:13:53.880 --> 0:14:11.839 needed more compute, it needed more data, needed more knowledge. 0:14:13.880 --> 0:14:17.679 Tracy Mike mentioned that you know, just looking at a 0:14:17.720 --> 0:14:19.480 shot on goals, so again to tell you so much. 0:14:19.480 --> 0:14:22.600 For example, it doesn't tell you if the referees are 0:14:22.600 --> 0:14:25.080 going to revisit a call made a minute before the 0:14:25.120 --> 0:14:28.040 shot halfway out on the other half of the screen 0:14:28.160 --> 0:14:29.320 and take away the goal. 0:14:29.520 --> 0:14:31.200 This was going to be my question, which we were 0:14:31.200 --> 0:14:34.560 talking about before the podcast recording, But like, how do 0:14:34.600 --> 0:14:39.320 you factor in let's say, the occasional randomness of the 0:14:39.400 --> 0:14:43.160 game and perhaps some erratic trying to be diplomatic here, 0:14:43.440 --> 0:14:47.800 erratic decision making by referees and footballing bodies. 0:14:48.480 --> 0:14:53.880 I think it's control the controllables. That's a fixed parameter 0:14:54.240 --> 0:14:58.080 within the game. Is that uncertainty that you can't control 0:14:58.440 --> 0:15:01.600 and it's hard to predict. There is human error, obviously, 0:15:02.320 --> 0:15:07.120 and it's impossible to isolate when that's going to happen 0:15:07.120 --> 0:15:09.240 in a match, and how it will you just sort 0:15:09.280 --> 0:15:09.920 of play the game. 0:15:10.280 --> 0:15:13.200 You also don't look at individual events per city. You 0:15:13.280 --> 0:15:16.920 might from an analysis perspective, but if you're building a 0:15:17.000 --> 0:15:21.120 basic predictive model for football outcomes, you might just take 0:15:21.280 --> 0:15:25.840 all the scoreboard results from the last ten years and 0:15:25.880 --> 0:15:28.320 then all those things cancel out, Like if one team 0:15:28.360 --> 0:15:30.720 has a red guard somewhere you wouldn't even know from 0:15:30.720 --> 0:15:33.600 these models, but you can build some interesting, let's say 0:15:33.640 --> 0:15:37.320 rudimentary predictive models with just the outcomes. 0:15:36.880 --> 0:15:41.200 Of games out of curiosity. Has var changed football analytics 0:15:41.200 --> 0:15:43.600 at all or presented new opportunities for data. 0:15:44.080 --> 0:15:47.640 It presents new opportunities for data because that is an 0:15:47.640 --> 0:15:52.000 event that happens. If a player is fractionally offside because 0:15:52.040 --> 0:15:55.920 his hand is ahead of the last defender. That doesn't 0:15:55.960 --> 0:15:59.360 take away from the fact that he still got into 0:15:59.600 --> 0:16:04.160 a good opportunity to position, possessed the ball, turned and 0:16:04.280 --> 0:16:08.080 struck the ball in the net. And often that data 0:16:08.280 --> 0:16:12.400 is nullified because of the var But in theory, it's 0:16:12.400 --> 0:16:15.160 something that you should consider in your data set. Within 0:16:15.240 --> 0:16:18.840 the raw data itself, there's tons of data that that 0:16:18.880 --> 0:16:24.080 you could add or sensor out. You know, yours mentioned 0:16:24.120 --> 0:16:28.440 red cards, and a lot of our models we censor 0:16:28.560 --> 0:16:32.280 that out of our data set. There's a there's an 0:16:32.360 --> 0:16:38.160 infamous match from three years ago between Chelsea and Tottenham 0:16:38.320 --> 0:16:44.400 where Tottenham went down to nine men and their response 0:16:44.800 --> 0:16:48.800 was to play a very high line, meaning they put 0:16:49.160 --> 0:16:53.560 all of their defenders up near midfield and tried to 0:16:53.600 --> 0:16:58.920 catch Chelsea off sides, and as one would expect, Chelsea 0:16:59.040 --> 0:17:02.760 proceeded to score or multiple goals later in the match. 0:17:02.920 --> 0:17:08.160 And one player in particular, Nicholas Jackson, had three goals 0:17:08.600 --> 0:17:11.359 in that one match, and when you look at his 0:17:12.080 --> 0:17:16.119 seasonal outputs for that entire season, that three goals was 0:17:16.119 --> 0:17:20.120 probably around twenty percent of his total goals. So when 0:17:20.160 --> 0:17:24.720 something like that happens, when you have an irregular game state, 0:17:25.480 --> 0:17:29.679 it is wise to sort of manipulate your data and 0:17:29.800 --> 0:17:32.199 remove that to give you a full picture of what 0:17:32.240 --> 0:17:35.000 does this game look like at an equal game state. 0:17:35.240 --> 0:17:38.000 Oh, that's interesting. So it's not like that game in 0:17:38.040 --> 0:17:42.880 particular was a rich fountain of data. It's important to 0:17:42.960 --> 0:17:46.600 sort of like recognize that the data from this particular 0:17:46.720 --> 0:17:52.560 game is not going to be particularly predictive about other games. 0:17:52.600 --> 0:17:55.040 And therefore, you a guy who scores three goals in 0:17:55.119 --> 0:17:57.480 that match is probably not going to continue to score 0:17:57.880 --> 0:17:59.040 three goals the game for the rest of it. 0:17:59.040 --> 0:18:02.520 There's a rich fountain of data. Okay, So at the 0:18:02.600 --> 0:18:06.119 end of the season, when you see people analyzing the player, 0:18:06.320 --> 0:18:09.639 they often analyze their season and what do they do 0:18:09.720 --> 0:18:13.160 per ninety minutes of football, And so you have this 0:18:13.680 --> 0:18:18.560 highly skewed data set by some really poor data. Yeah, 0:18:18.560 --> 0:18:22.000 and so there's a lot of data mining involved in 0:18:22.040 --> 0:18:26.480 the process of building a model when you evaluate a 0:18:26.760 --> 0:18:29.040 team and individual players. 0:18:29.280 --> 0:18:32.080 So here's a question I have, and you're talking about 0:18:32.160 --> 0:18:35.080 using neural networks, and it's like, eventually, like you know, 0:18:35.119 --> 0:18:38.639 we'll have the compute to like have the position of 0:18:38.720 --> 0:18:43.199 every player's body flash maybe you know, hundreds of images, 0:18:43.520 --> 0:18:46.679 frames per second and so forth, and then you have 0:18:46.760 --> 0:18:50.320 feeded all into a model, and then we learned something 0:18:50.359 --> 0:18:53.200 about who's more likely to win. But one of the 0:18:53.240 --> 0:18:57.199 things that happens in a lot of other domains, and 0:18:57.240 --> 0:19:00.000 here I'm thinking about like chess or go, for example, 0:19:00.520 --> 0:19:03.080 is that you can have these models that are extraordinary, 0:19:03.440 --> 0:19:05.240 but they don't speak English, or they don't speak any 0:19:05.320 --> 0:19:08.480 human language, and so the transmission of like what was 0:19:08.600 --> 0:19:11.840 learned from these events is something usable by say a 0:19:12.000 --> 0:19:15.800 coach who's thinking about strategy or a general manager who's 0:19:15.840 --> 0:19:19.520 thinking about player selection. Talk to us about, like, when 0:19:19.600 --> 0:19:22.680 you think about these machine learning models that can't really 0:19:22.720 --> 0:19:26.760 communicate their findings in any way in language that humans 0:19:26.760 --> 0:19:29.960 can understand, how you sort of bridge that gap to 0:19:30.200 --> 0:19:33.800 where this is useful information for a team or a manager. 0:19:34.160 --> 0:19:37.320 This is generally be understood I think in the last 0:19:37.600 --> 0:19:40.639 couple of years or the last eight years, as the 0:19:40.680 --> 0:19:44.040 biggest promm you can build these models, these neural nets 0:19:44.040 --> 0:19:47.720 already exist. The main thing is the translation, like you said, 0:19:47.800 --> 0:19:52.320 from model outputs to coach, and I think in the 0:19:52.400 --> 0:19:54.520 last couple of years most most teams have found is 0:19:54.560 --> 0:19:57.440 that they need an expert analyst and data analysts to 0:19:57.520 --> 0:20:01.480 do this conversion or the translations, where the coach isn't 0:20:01.520 --> 0:20:04.000 fed the data directly. Sure the coach is fed just 0:20:04.160 --> 0:20:06.960 the information that the analyst finds from the data, and 0:20:07.000 --> 0:20:09.880 that could be in the end. That generally still boils 0:20:09.880 --> 0:20:12.399 down to having video clips, so you can use the 0:20:12.440 --> 0:20:14.600 data to find video, and then you can show the 0:20:14.680 --> 0:20:18.119 video to the coach, which then from bottom up you 0:20:18.200 --> 0:20:18.760 have this. 0:20:19.000 --> 0:20:22.280 Approach that makes sense. But the part about okay, we're 0:20:22.280 --> 0:20:25.000 going to use the data to find video. So all 0:20:25.000 --> 0:20:26.919 of this makes sense. You have the data special ist 0:20:26.920 --> 0:20:30.320 who translates, you have the video so that there's something tangible. 0:20:30.760 --> 0:20:34.000 But talk to us about that specific step where the 0:20:34.119 --> 0:20:37.840 data analyst sees some sort of model output and then 0:20:37.960 --> 0:20:40.840 is able to use that to find the relevant clip. 0:20:41.160 --> 0:20:44.160 To show the coach something potentially instructive, because that seems 0:20:44.200 --> 0:20:45.000 like the hard part to me. 0:20:45.440 --> 0:20:47.919 Yeah, So the model output could be many things. It 0:20:47.920 --> 0:20:51.640 could be outputs from a classification model that says, in 0:20:51.640 --> 0:20:55.000 this given ten second or fifteen or twenty second window, 0:20:55.640 --> 0:20:57.760 this team played in this sort of build up, or 0:20:57.920 --> 0:21:00.480 they have this type of structure. It could also be 0:21:00.800 --> 0:21:03.160 a little bit more advanced where you can simply say, well, 0:21:03.200 --> 0:21:05.840 in this instance, we had a high probability of conceding 0:21:05.880 --> 0:21:08.160 a goal. And that could be just from an expected 0:21:08.200 --> 0:21:10.320 goal shot if you're looking at only a cent data. 0:21:10.359 --> 0:21:13.199 But you could also have model outputs from something we 0:21:13.280 --> 0:21:16.240 call an expected possession value model or an expected tread model, 0:21:16.240 --> 0:21:18.760 where you measure the chance that the teams going to 0:21:18.760 --> 0:21:21.560 score and let's say the next thirty seconds or the 0:21:21.560 --> 0:21:24.280 next possession, and then you can find this fight there 0:21:24.680 --> 0:21:27.760 and either measure when your team is likely to concede 0:21:27.920 --> 0:21:29.920 or measure when the other team is likely to score, 0:21:30.080 --> 0:21:32.560 and you can use those kind of signals to boil 0:21:32.600 --> 0:21:33.320 it down to video. 0:21:33.680 --> 0:21:35.719 Yeah, this is something I wanted to ask. Actually, so 0:21:36.080 --> 0:21:39.399 you mentioned speed just then, like, what is the actual 0:21:39.520 --> 0:21:42.120 latency that we're talking about since we're using all these 0:21:42.119 --> 0:21:46.280 market Yeah, jeez, Like, are we talking about an insight 0:21:46.440 --> 0:21:50.439 that's actionable within seconds, like you're going to sub a 0:21:50.480 --> 0:21:53.960 player on after your model splits something out like live 0:21:54.080 --> 0:21:57.600 during the game. Or is it more realistically that you're 0:21:57.680 --> 0:22:00.680 reviewing the model and the analytics a d a game 0:22:00.880 --> 0:22:04.520 and sort of tweaking. I guess when things have calmed down. 0:22:05.119 --> 0:22:08.040 Apparently most of this does not happen live, So most 0:22:08.040 --> 0:22:11.040 of it happens either pre match or postmatch. 0:22:11.080 --> 0:22:12.399