Match analysis has always involved seeing a well-dressed ex-pro move into a studio and state that a player “wanted it more” as an explanation for losing a back-post header…
…until now. Now we have the ex-pros of the future, or their substitutes, in the form of computer science PhD students from Belgium. Entranced by magic, or perhaps just math, the students have created the tools of the future, and coaches have shown a “staggering” amount of interest and investment in the technology.
Beyond the crude xG stats
A lot of data-informed betting started for many using the xG metric. The metric shows that a team should have scored three times while they missed four easy goals.
The issue with standard xG models is that they treat every player like a generic avatar in a video game. They don’t care if it was a world-class finisher or a center-back on his weaker foot attempting that volley from eighteen yards out.
This is the state of betting now, and deep-learning tracking models have recently surged to the forefront. The betting markets take into account shot locations along spatial pressure and body orientation along with pitch geometry and defensive recovery speeds. It is safe to say, if you are utilizing basic box scores to try and find a value in the market, you are far from the cutting edge.
Finding asymmetric edges
This leads to an information ecosystem where bookmakers have extremely accurate odds, but they are still blinded when it comes to making decisions.
The “public” always prefers recent narratives. For example, if a player has a scoring slump for a month, bets will be placed in the market to say the player’s team is winning less, despite the fact that the player is still highly valuable because every fifteen minutes he is occupying two defenders.
That difference between what people “see” and what the data analytics are saying is where the true value is located. If you only use the specific services available to you across the internet, whether they be soccer betting tips for SA punters or the more advanced tracking metrics for games in low-popularity leagues, you are not attempting to outsmart a computer. Rather, you are attempting to expose, and profit from, the general public’s overreaction to poor-quality, basic highlight videos, when there is a large disparity between the metrics and what the data says.
Market inefficiencies in non-elite leagues
Looking back to when data started to become more mainstream in sports, it is clear that the leagues with the wealthiest clubs were the great beneficiaries. Clubs like Liverpool, Leipzig, and Brentford, and Club Brugge hired PhDs in physics and built custom software just to be at a marginally higher level than their competition.
Now, because elite leagues have such a developed data infrastructure, they are pretty much priced perfectly.
Real asymmetry exists further down the pyramid. Deep learning models require huge investments to create and maintain, which means that lower divisions do not receive the same level of real-time algorithmic inspection by oddsmakers. A pricing engine may use generalized mathematical assumptions for lower divisions which may not withstand detailed spatial analysis, particularly when there are crowded mid-week matches.
The tracking data suggests that team press intensity exponentially depreciates after the 60th minute of a third match in a seven-day period.
