With the summer break giving many analysts time to refine their models and peruse the transfer gossip, we have been hard at work developing a new metric for evaluating players across Europe’s top five leagues.
There are many issues in trying to assign ratings to players based on in-game statistics such as tackles, passes, dribbles, key passes, big chances etc… These issues stem from many of these stats being interrelated which leads to either double counting of certain metrics or over/under evaluation of a player’s ability. This second issue is most evident when analysing defenders and goalkeepers. Simply counting attempted and successful tackles is fool’s gold. For example, in last season’s Premier League, Erik Pieters attempted the most tackles (106) with a 76.4% success rate (81). Even a casual football fan could tell you that Erik Pieters is not the best defender in the league. The problem is that better teams do not need to rely on their defenders making tackles as much as they control more of the ball. This issue can be fixed by adjusting these defensive stats by each team’s possession numbers, but there are still a myriad of issues in (mis)using in-game statistics.
That being the case we decided to take a different approach in rating players. By looking at a player’s contribution to his team’s goal difference for the time he was on the pitch, we can analyse in real terms the effect he has had on his team. This is not a simple +/- goal difference rating, instead we have incorporated factors such as minutes played, the strength of his own team, and the strength of the opposition.
Thanks to data provided by Stratabet, we have looked at Europe’s top 5 leagues, the Championship, Bundesliga II, and Eredivisie and have rated each player that set foot on the pitch for the 2016/2017 season.
Player ELO (pELO) is a metric that rates players for their score line contribution based on a modification of the ELO system.
ELO ratings were originally developed for rating chess players, with the basic premise that higher rated players would win more games (in chess, a player rated 200 points better that his/her opponent would be expected to win ~70% of games). In the years since, it has been adapted for many sports including MLB, NBA, tennis, and of course football. The beauty about ELO is that it takes team strength into consideration. For example, let’s take a look at Liverpool’s match against Middlesbrough at the end of the 2016/17 season. Liverpool had an ELO rating of 1713 going into that game and Middlesbrough had a rating of 1396. Liverpool had an advantage of 317 points. Liverpool won the match 3-0 and their ELO increased by just 5 points, while Middlesbrough’s decreased by the same amount. Had Middlesbrough won, Liverpool would have lost considerably more (-27 points), and Boro would have gained 27 points. There are further adjustments made for home advantage and goal importance. The goal that makes the match 1-0, or 2-1 is far more important than the goal that makes the match 5-0, and this is adjusted within our ELO model.
We decided to take a similar approach with our pELO model. Players facing off against teams with much higher/lower ELO would be awarded or penalised points accordingly. However, as we are dealing with individual players rather than a team, a weighting was applied to better reflect an individual player’s contribution to his team’s final result. Here is a look at one of the Premier League’s starlets of last season – Kevin de Bruyne.
Each player began the season with a base rating of 100. As the season progressed each player’s rating rose and fell depending on their team’s goal difference for the time they were playing. The use of 100 for every player presents an issue, however. By giving each player an equal base rating, the model assumes that every player in the league is equal. This is of course not true. After a few games (and certainly towards the latter parts of the season) these ratings begin to reflect the “true” hierarchy within the Premier League. We decided against taking a Bayesian approach in favour of letting the ratings develop organically over the course of a few seasons, so there is still further this model needs to go before we see the natural order of things as August 2016 is currently our year zero. As things stand, here are the top 20 aPELO rated players in the Premier League:
As you can see, the top 20 is packed with Chelsea and Tottenham players, as would be expected considering their finishing positions last season. We can take this a step further though. The above aPELO does not take into consideration team over/under performance, so in order to get a more in-depth view, we have developed a new xG model to take this into consideration.
Expected pELO (xPELO)
In football, the final result often does not tell the full story. We’ve all watched games where one team dominated, getting shots off, maintaining pressure high up the pitch, only for the opposition to break away in the late stages and nab a goal, or an unlucky deflection to guide the ball into the back of their net. You think that the other team didn’t deserve to win, and you may be right. In order to check this, we can take a look at Expected Goals (xG). What is now the bread and butter of football analytics, xG is essentially a metric that calculates the probability of a shot being a goal. There are many ways of calculating this, which you can read about here, here, and here, but the basic premise is the same: the closer you are to the centre of goal, the more likely you are to score. A player that takes 10 shots from the centre of the six-yard box will score a lot more goals than a player that takes 10 shots from the half-way line.
By using xG we can say whether a team “deserved” to win or lose, we can find players who get into better positions to shoot, and we can see which teams or players are possibly over or under performing. As an example, Jamie Vardy far over-performed his xG numbers in Leicester City’s title winning season. Based on that over performance, it was likely that at some point in the future his actual goal output would regress. It just so happened that that regression happened the following season, but that is not always the case. There are some detractors of expected goals – and the use of stats in football in general – and you can find some arguments against xG here, but in our opinion, it is one of the best advanced metrics in player and team analysis that is generally available to the public.
So by looking deeper than the actual final result of a match, we can assess each player’s expected PELO, and as the season, and their careers unfold, check their xPELO output against their aPELO. Here’s KDB’s chart again with his xPELO added:
We can see here that while KDB had a great season with an increase in his aPELO of 37 points, in reality he underperformed. When we look at the xPELO numbers, we find that deBruyne is no longer outside of the top 20. In fact, he’s the highest rated player in the Premier League:
The above chart shows the top 20 players who have contributed the most to their team’s xGD. We can see those who may have overperformed and those who have underperformed. Again, as we’d expect, the top twenty is dominated by Chelsea, City, and Spurs. Alright then, so what?
We can use these ratings in a practical way by seeing how teams line up against each other, and from that develop match probabilities based on those starting line ups:
Here is a look at a match from the end of the 2016/17 season. By averaging each team’s xPELO and aPELO we can derive match %s and compare them with what is being offered on the betting market. Each player has an influence on the match % as of course each player has an influence on his team. Here is a look at another match from the end of last season:
Spurs were favourites for this won and won the match 2-1. But how would things have looked if Ibrahimovic hadn’t got injured and started instead of Rooney? Or if Valencia played instead of Tuanzebe? By using Player ELO ratings, we can see how these changes impact on the overall match probability:
Those hypothetical changes would have (slightly) increased Man United’s chances in the game.
pELO also allows us to assess activity in the transfer market. Where are a team’s weak links and is there a player available in the market that can strengthen that area? Are there any players that have underperformed their pELO and are possibly being undervalued? That’s a topic for another article.
Caveats and Limitations
At the moment, our dataset only has information from the 2016/2017 season. That being the case, as mentioned earlier each player began this season with a base-rating of 100 points. As such, it may take some more time for these ratings to find their natural place. Ideally we would like a minimum of two seasons worth of data as this would make the ratings and match probabilities a whole lot sharper (for our team ELO ratings and match probabilities we used seasons 2000-2004/05 to allow the ELOs to develop and trained our data from seasons 2005-2015/16). Sadly, it is not possible to obtain shot data prior to 2016/17 at the moment, but we’ll kick on and see how things develop in future.
We’ll be providing updates regularly for all leagues we have data for – Top 5 European leagues, Championship, and Eredivisie. These will include tables, match previews and match probabilities, as well as a blog series documenting how we do using PELO against the betting markets.
We would like to develop this even further and would love to hear your thoughts on what we should work on next. (We’re thinking about an under-21s database).
Well that’s all for now, if you’ve any questions you can reach us on the usual channels.
by Peter McKeever and Raven Beale.
This article was written with the aid of StrataData, which is property of Stratagem Technologies. StrataData powers the StrataBet Sports Trading Platform, in addition to StrataBet Premium Recommendations.