It all started with a problem, teams would have decent fixture runs and the defenders would pick up several clean sheets. After a couple clean sheets, you start to take notice and look to add them. However, by this time, the fixture run was almost over and these assets were no longer an option. So the question arose, how do you identify teams performing well defensively in tough fixtures ahead of a good run? The answer, while easy, isn’t simple. It comes down to having a realistic expectation and seeing if teams meet, exceed, or fall short of that. To do this you need to develop some sort of rating system for all the teams, use it to set this baseline performance, and then update it as performance changes. This was the start of the xElo Experiment.
Any chess players or early online gamers may know about the Elo ranking system. In short, each player has a rating. These ratings are compared and an ‘expected’ outcome is generated. Following the match, the expected result is compared to the actual result and the player ratings are updated accordingly. What is great about the Elo system is that the relative strength of the opponent is considered along with the outcome to adjust the ratings. For example, if Magnus Carlsen and I had a chess match, and he beat me, that would be a surprise to no one and his rating would likely not change. However, if I was able to best him, my rating should skyrocket (or at least increase the maximum amount allowed in the system) with his rating dropping accordingly. This works well for chess because outcomes are simple: win, loss, or draw. This is also true of football, and the website clubelo.com does have Elo ratings for many leagues and predicts probabilities of outcomes. While this is a great resource, it isn’t exactly applicable to the problem. This is because I don’t care about wins/losses, I want to know the number of goals scored so I can predict clean sheets.
This is where we take the Elo system and start to get a little creative. Elo uses expected outcomes of 0 to 1, with 0 being my chances against Magnus and 1 (certain win) being his chances against me. I needed to make the goals in a game fit into this 0 to 1 range. I tried a number of strategies to assign numbers to this scale (Clean sheet = 0, 1 goal = 0.3, 2 goals 0.5, etc), but ran into some problems. The biggest issue is that goals in football are pretty rare and are not always deserved. A team can have a wonderful game and allow no chances only for the keeper to pass it to Ings who scores. Should this team be credited with the same result as a team who gave up numerous chances only to have your captain Sterling squander them all? After some analysis and pondering, I thought xG may be the answer.
While goals win games, xG gives a bit more context and cuts down on the variability. It can give a better summary of the number and/or quality of chances a team has (or concedes) in a match. This seems valuable in solving my problem and so I started looking into it.
The story starts to get a bit more technical now, so let’s save that for another day and skip ahead to some of the key decisions. First, I want to say that the model is still quite experimental. I am happy for you to look at the predictions and help that shape your decisions, but I have intentionally referred to it as the xElo Experiment to help communicate that this isn’t a tried and true model. Anyway, on to the details:
• The model works within a range of 0 to 3 xG, but rarely predicts results close to the limits.
• Teams have both an attacking (xG for) and defending (xG conceded) rating.
• This attacking and defending rating is also split home and away, for a total of 4 ratings per team.
• Classic Elo formulas are used with some minor adjustments for the constant values.
• xG of Penalties is discounted to be equivalent to an average shot in the box
• Ratings change less for results in which either team (or both) received a red card.
How that you have made it through the history and are up to speed, let’s give you some information that can actually help you (hopefully). Below are the team rankings
Manchester City rank near the top of almost all categories. Other notable teams;
• Chelsea at home stand out and have the best home defense
• Wolves rank near the top defensively
• Liverpool are near the top, but don’t stand out like some of the other teams
• Arsenal and Tottenham don’t stand out much from the Little 14 teams.
• You may notice that all three promoted sides have the same ratings, this is just how I set up the model at this time and we shall watch these numbers change as we move forward.
Finally on to the predictions. Since we are at the start of the seasons, I took a look at each team’s opening fixtures compared to the average for the year.
Lots of teams are not too far off their baseline, but Manchester United stands out as a great run to start surpassing Manchester City at the top. Leeds also stands out, but for the wrong reason. Their cheap attacking defensive options may have to wait (or sit on the bench).
Everyone is a little closer to their baseline for the attack. Manchester United and Tottenham are the teams with a bit of a better run and Crystal Palace, Manchester City, and Leeds getting the worst of it. It is worth noting that the West Ham opening fixtures are not too far off baseline, but this is largely due to their first two games.
Now lets get to the good stuff, the GW1 predictions.
Liverpool is the standout here, partly due to their home fixture against a promoted (but defensive) side, but primarily due to the lack of Manchester Clubs. While I speculate that the Leeds’ rating may be lower than is appropriate, it is worth noting that their rating is almost the same as Sheffield United’s defensive away rating. When it comes to defence, home is usually king.
As for other surprises, West Ham and Leister ahead of Chelsea is a bit of a surprise. Everton and Wolves so low in the attacking chart is also worth noting. For defensive returns, the top teams look to be Liverpool, Wolves, Leicester, and Southampton.
However, it is important to note that no changes were made to the ratings over the off season with signings and departures and injuries. I expect some surprises in the early weeks and we shall see how the model does. See you all next week when we can look back on what happened and look forward to Gameweek 2.