Since I started working on my soccer team ratings around 9 years ago, my goal has been to make my rating system as predictive as possible.
xG models are popular because they contain useful descriptive information as well as being more predictive than simpler metrics such as goal difference/points/shot ratio.
For a rating to be predictive we need to make as many adjustments as required such that a league average team always rates as league average. A team that plays while losing scoreline may appear to perform better in a metric, a team that’s played lower ranked opposition may appear to be better than they are, a team with a player advantage will appear better than their true level e.t.c. Another consideration is to lower the variance such that our rating system needs fewer games for the same level of predictive accuracy than other ratings.
In this article I am going to give a basic summary of the ideas I have that improve on the predictive performance starting from a baseline of xG.
All these have been the basis for the ideas and adjustments that are used in my team ratings today. I am going to cover these topics in a lot more depth in the future of this site.
- A team gets no credit for dangerous passes. It’s a good sign if a team is making more successful passes into a dangerous area than another if everything else was equal.
- Fixture adjustment – This will obviously be important over smaller samples of games, we need to adjust for teams that have played differing schedules.
- Game state adjustment – A game state refers to the ‘state/scoreline of the game’, i.e. was the home team winning the game for 80% of the match? Was the 2nd half a formality as a team lead 4-0? Was the game tied for 90% of the total minutes? When a team leads a lot of matches and does not have a clear incentive to score more they may appear to perform at a lower level. On the other hand, when a team leads a lot of matches they may find scoring and attacking easier. How do game state adjustments actually work and how do we calculate the direction and magnitude of the adjustment?
- Red card adjustment – If teams play minutes up or down a player that will make them look better/worse than they actually are (as red cards happen fairly randomly)
- Scoring goals gets no credit. Even though finishing is fairly random it is not completely random hence we may want to give a team *some* amount of credit for finishing ability or goalkeeping/defending ability.
- Unassisted shots – Are unassisted shots created through a more random process? Are they less repeatable and should hence be given less credit.
- Shots with very large xG give the team too much credit. I don’t think this is too debatable if we are talking about an unassisted shot that bounces off a players thigh on the goal-line during a scramble following a corner but I still believe this even if we talk about a great open-play intricate team goal that results in an unmissable pass into the goal.
- Team style – Can teams overperform raw xG because of the way they play? Consider maybe Juventus or Atletico Madrid. How could we predict this over/under performance.
- Sample size – If we have to rate a team’s performance to the best of our ability over 3 games should we use different parameters than say, for a 25 game sample? When do we start putting more weight on big chances? When does awarding credit for successful passes stop adding predictive value?
If you would like to get in touch about anything you can e-mail syzygy112358@proton.me or tweet me at https://x.com/SamH112358
Syzygy Analytics 
Leave a comment