Decision Technology has built a range of proprietary models for football (soccer) prediction and player analysis.
Dr Paul Jackson is a Project Manager with DecTech and he’s on the podcast to talk about his journey from being the son of a bookie, to a PHD in High Energy Physics and finally his career in sports analytics including soccer predictions and betting.
- How DecTech were deep into analytics well before the Moneyball phenomenon
- Why the model’s inputs are not over-complicated
- Rating players across different competitions
- How they help high net worth clients with their own trading
- Why they are looking to expand to t20 cricket, snooker and rugby
Dave Duffield: G’day Paul, thanks for joining us.
We are betting from all angles but you’re certainly the first person I’ve had on that has a PHD in high energy physics.
Paul Jackson: There’s actually quite a few physicists in the statistical community at the moment it feels like.
Dave Duffield: There’s a crossover there, is there?
Paul Jackson: Oh, definitely.
Dave Duffield: You did the PHD and then ended up in financial services, just tell the listeners a little bit about your background.
Paul Jackson: Yeah, sure. Funnily enough actually back when I was a kid my dad was a bookmaker, so it’s quite ironic to be on the other side, especially as he taught me that in the long run it’s the bookie who always wins.
I did a degree in Physics and then took that on to a PHD in High Energy Physics, and then decided it was time to go into the real world, unfortunately, so I worked for a company called SunGard who are a financial services company. A lot of my friends were quite surprised at the jump from physics to finance, but there’s the use of these transferable skills … So modelling and statistical analysis and developing code. High Energy Physics is probably 90% about writing good code rather than physics.
Then after doing that for a few years, I decided that I wanted a better work / life balance, and so was looking around for an interesting challenge, and found DecTech. The plan was to go and work at DecTech and retrain as a behavioural scientist, but I very quickly managed to get transferred onto the sports analytics team. I’ve been with DecTech for five years now.
Dave Duffield: So you transferred to sports analytics, was that all part of your grand plan?
Paul Jackson: Not initially, I mean I was certainly fascinated to see that DecTech was doing work in this area as well. But I have to say once I did behavioural science the work of the analytics team was clearly much more interesting for me personally.
Dave Duffield: What is that work, what does DecTech actually do?
Paul Jackson: DecTech is a spin out from Warwick University, so they were set up to apply behavioural science to real world situations, basically to help FTSE 100 companies use the insights of the latest behavioural research to effectively sell more stuff. In that first year of setting up, Henry decided that he’d also like to turn his hand at a football model, and out of that model grew the sports analytics side of the company.
Dave Duffield: You might want to talk about that football model, is the origins of that the Fink Tank ratings?
Paul Jackson: Yeah, it’s how the Fink Tank came about. The model that Henry worked with he famously predicted Senegal’s win over France in 2002. And I suspect that it was just more likely that there would be an upset, obviously we’re all about probability. But Daniel Finkelstein, who’s now Lord Finkelstein, heard this on the radio, got in touch and that’s where the Fink Tank was born. That’s continued now, so that’s … What’s it’s 2016 so fourteen years that’s been going now?
Dave Duffield: Okay, so I mean you can’t go into too much detail as to how they’re generated, but they cover all the major English leagues and in the Australian language I suppose you’d say, there’s probabilities for each outcome and those probabilities obviously easily transfer into odds. The people that have backed the overlays, the value based on those ratings, have done quite well?
Paul Jackson: Yeah, absolutely. Obviously we’ve got a very academic background, so when we started to work in football it wasn’t really with an eye on trading in particular. However very quickly the benchmark to test yourself against is obviously the bookmakers, and so over time we ended up discovering that we had a model that seemed to consistently perform quite well.
Then I think it was around 2006, we started to publish our predictions for free online, and in 2008 the Secret Betting Club picked up those predictions and started to effectively proof us even though we were just obviously, we didn’t have a relationship with them at that point. They also independently found that we seemed to be on to something with our model, and continued to follow us until we decided to take the free website down, at which point Pete got in touch to ask if his members could still keep access.
Dave Duffield: I speak to a local sports modeller and I think he said he had seventy thousand variables in the latest model that he’s working on, but with this one, this was a lot simpler and you weren’t looking to cover every aspect of every game, it was using a few key inputs to generate those probabilities?
Paul Jackson: Yeah, that’s it. We’re certainly nowhere near that number of parameters in our model. The model could include a lot more information, and as research internally we do put a lot of work into, partly because the Fink Tank is always throwing up these interesting questions about certain myths in football and interesting potential correlates. We take that, the model that we always published I think we appreciate the consistency, that people know that it’s the same predictions this week as two years ago, because that allows them to make a consistent system off the back of it.
The model that we provide to SBC and that we used to publish on the website has remained pretty consistent and internally we do look at trying to incorporate things. For example, we’ve got a player rating model which we used with some of our clients; and do you know it’s actually incredibly difficult to add additional information that’s statistically significant? I think if we ever found something that we could actually put a cast iron guarantee on then we would update the model that we publish externally. But actually it’s really quite challenging, it’s not just about doing better, it’s about doing sufficiently well that it’s worth making that big change of saying actually now it’s effectively a new model.
Dave Duffield: Is that because a lot of the important and more well known factors are already we would say ‘baked into the cake’? They’re already in the odds available, so you’re trying to find something that’s significantly undervalued?
Paul Jackson: Well, for example the model doesn’t know anything about team lineups, so for example if your favourite two strikers are injured then you’d expect that team to obviously under-perform and the model won’t know about that. However, whenever we looked into this it feels like even if the model is under-reacting to those sort of events, everyone else is perhaps overreacting. Even though the model isn’t taking it into account somehow it manages to still work over the long term.
But that said, I know that there were other people out there who in the past have used our ratings because one of the great things about providing the free ratings was we did get a lot of interesting discussions from people who were feeding them into their own models, and so I think it is possible to do something around team lineup and what have you. We see hints of it in the stuff we do internally, but as I say it’s just really difficult to come up with a solid statistical basis that the bunch of academics that we are would be happy to put our names behind.
Dave Duffield: Right, so something like Head to Head record and how that affects the results of the next match-up, is that something that has … Maybe significance but does it have any value from a betting perspective?
Paul Jackson: That’s a good example actually, because in the last couple of months we actually had a Fink Tank about Head to Head records. We found that although Head to Head records do a reasonable job … If you take, say, I think it was the eight most recent matches, it does a reasonable job of predicting the most likely outcome, but it does a terrible job compared to if you were going to try and use those outcomes in a probabilistic sense and try and use them to generate odds to take on the bookmakers. So I think the answer on that sort of Head to Head stuff is there’s something there, but the information already in the model is capturing it and is doing a better job on top of that.
Dave Duffield: You mentioned player ratings earlier. How do you put a number on a player, particularly when you’re talking about many different competitions and just a different level of skill and class across those different competitions?
Paul Jackson: Yeah, actually comparing across competitions is incredibly difficult, and we use more than one technique, we use several techniques. That’s actually an area that we kept quite close to our chest actually, because I think in the past some of our ideas have moved out into the community, which is fantastic, but as a business that needs to try and make money and pay the wages of the employees I think in the past we’ve perhaps given a bit too much away. But you’re right; comparing players is incredibly challenging across leagues, and I think is one of the things that we can do better than most people. Actually, fairly confident we can do actually better than our competitors.
Within the match itself, so obviously a single match we feel doesn’t provide nearly enough information on a player’s ability, and so although we can provide ratings when we have perhaps five hundred minutes where there’s a very large uncertainty that we would flag up to our clients about that. But over a season, or potentially if we’ve got ideally a couple of seasons on a player, then they do have enough interactions that we feel that we can come up with a good assessment of their contribution. Basically we look at their impact for each of those actions, so for example if someone takes the ball on the halfway line and passes it and the pass is received by one of his team mates in a more attacking position, then we have an estimate of how much more likely that is to produce a goal and we can sum that up over the match and assign the various credits and deficits to players based on whether they were helping their team to score or hindering.
Dave Duffield: So is soccer harder to model because it’s such a low scoring game? I’d imagine NBA Basketball and it’s one hundred points to ninety-eight, you can see all the scoring events and what lead up to those and the various players contribution to that. Is soccer harder because there’s often just one, two or three goals scored?
Paul Jackson: I think, yeah definitely it makes it harder. You start to look for kind of the next steps, so for example the quality of the shots on target, or other interactions that can lead to those shots. I think that adds an extra layer of complexity. There are also other issues as well; football is obviously twenty-two people on the pitch, and so there’s a lot more interactions and in basketball the number of passes that a particular player makes is much more numerous, so you can capture data on those players more quickly. Also, those American sports they also play each other much more often, and so there are lots of reasons why I think football is harder to model.
Dave Duffield: So there’s the media side, and then the management side. Of particular interest to the listeners I’m sure to this would be the trading side, what’s your involvement there in terms of your own trading?
Paul Jackson: We feel it’s important that we also see risk based on the back of our predictions, so we actually do trade ourselves but we only trade on Betfair. Which is interesting, because it’s not part of SBC’s panel, so I feel quite happy about that, that we’re not potentially affecting the quality of their tips. We also analyse the market and we know that we’re not in any way moving Betfair’s market in any detectable manner, as well. I think that’s really important, that if our tips aren’t of a good quality that we suffer financially. But the other thing, as a bunch of academics we love a nice data set, and so it allows us to capture interesting information and also we try various different strategies on a smaller scale as well, to see how they perform in the real world. So that’s our own personal involvement.
Of course, you’ve mentioned we’ve since partnered with the Secret Betting Club to provide our tips to them in England exclusively. In the past we have worked with odds compilers, which has been interesting because most of that was before my time and the culture of numbers and statistics in bookmaking seemed to be just as reluctantly taken on board as within the football teams themselves.
Actually that’s really been interesting in football, because of course DecTech being involved for so many years kind of pretty much predates the moneyball phenomenon getting traction in the press and within the football clubs themselves. We’ve seen as we’ve visited other clubs over time that that initial reluctance to trust numbers is slowly improving and people are becoming more receptive to it, and I think that’s true of the bookmakers as well.
Dave Duffield: As people become more receptive to it, does that mean that the market becomes more efficient and what edges or what values there was previously tends to gradually disappear over time?
Paul Jackson: I would expect that to be true. Although, we’ve often revisited this and it’s very hard to find a trend to suggest that it’s happening, and I don’t know whether that’s because the volume of money in the markets is driven by effectively the punter who just wants to have a bit more excitement when he watches his team. You’d imagine the likes of Star Lizard and the other analytical guys, I know they obviously have huge bankrolls but that said the betting market globally is incredibly huge as well. I’m sure they can move individual markets with individual bookmakers, but perhaps there’s just not enough statistical people out there, and there’s perhaps a little bit for everyone.
I think the other thing that might be worth saying as well is I get the impression that the likes of Star Lizard also to make it worth their while they are very much in the markets where there’s very high liquidity, whereas perhaps on a Friday afternoon when the tips are being sent out, which is when we tend to do most of our analysis against, perhaps they’re just not in the market at that point and so there’s still an opportunity for everyone else to make some money.
Dave Duffield: Just to finish up then, to quote from your website it mentions that you provide “high net worth clients with the tools to help them trade on their own account and can create bespoke analysis to maximize their returns”. I know you won’t be able to go into a high level of detail, but are you able to expand on that at all?
Paul Jackson: Yeah, so this is something that has actually been around … I’ve been at DecTech five years and so this has been going on since before then, and there’s been some very interesting stories. For example, these individuals are often able to negotiate special … With the bookmakers, and then they come to us to get advice, for example, on how to manipulate that special offer to their advantage. Obviously over time, the bookmakers have got more and more wise to people doing this sort of thing. Also just occasionally perhaps one of them fancies something to do with snooker and so in the past I know that we’ve done a little model of snooker, I think this was to analyze the chance of a 1-4-7 for a particular player. They also obviously take our predictions as well, but we don’t do any of the bet placing or bankroll management.
Dave Duffield: You mentioned snooker, and obviously football … Soccer in our lingo. Are they the only two sports that you cover?
Paul Jackson: We’ve done some stuff for rugby in the past, which unfortunately at that time the data set just wasn’t as detailed. So we use Opta data primarily for our player analysis and something similar in rugby just wasn’t possible at that time. We’ve also got a cricket model that works for t20 cricket, and that’s more of a research project for us, I think that’s the sport that we’d probably most like to focus on in the future.
Dave Duffield: Would that have a pre-game focus or in-game?
Paul Jackson: I think probably pre-game would be our focus. But that said, even within football we don’t trade in game, we do have models that work in-game as well. In fact when we were working on one of the major competitions, there was a lovely widget that took the current score and the number of cards in the match and a few other factors and you were able to track over time the chance of your team winning. Which was really quite nice, although we don’t have the data set to compare that to the bookmakers. So it would be interesting to come back to that in the future when we’ve got the data to see how well the in-play version of the model works.
Dave Duffield: Sounds like you’re only limited by access to that data and just how many hours there are in a day?
Paul Jackson: There’s definitely so many ideas that we could work on as a team that we don’t have time for.
Dave Duffield: Excellent, well I really appreciate you coming on the show, Paul. I’m sure it’s been insightful for the listeners, and best of luck for the rest of 2016.
Paul Jackson: Great, thank you. Thank you for inviting me.
Get More Betting 360
Make sure you don’t miss our punting tips to come! Subscribe on:
or you can directly download this episode by right-clicking, Save As Here.