Creating some new metrics using Opta's Big Chance stat - Part 2

It has been a while since I have written anything. This is partly due to life just getting that little bit busier, both at work and at home, but also because, after one too many spillages, they keyboard on my laptop stopped working. This is a piece that I actually had 95% completed prior to the keyboard giving up the ghost on me, so I didn't quite get it done, and in this business your data becomes out of date very quickly! This was meant to be posted soon after I presented at the OptaPro Forum last February, and is a follow up to the last piece I wrote that you can read here.

Recently I posted the slides from my presentation at the Opta Forum, and in this I will look at on one of the metrics that I introduced in the presentation, the Big Chance Ratio (“BCR”). If you don’t know already, Big Chances are one of Opta's few subjective stats, described as “A situation where a player should reasonably be expected to score usually in a one-on-one scenario or from very close range.” Big Chances have only been measured by Opta for 4 full seasons now, and this gives us 80 observations to check the relationship with points, but due to this season not yet being completed and teams getting relegated, we only have 51 observations to test the year-on-year relationship. I wrote in more detail here about how many Big Chances a team gets on average over the season and the rate at which they are converted at here, and there has been little change, with the average team over the past 4 seasons taking 535 shots, with about 13%, or an average of 68 of which are Big Chances.

As I did in my last post, I am going to be using the Total Shot Ratio (“TSR”), which measures the proportion of shots that a team takes compared to its opposition as the baseline to compare the different metrics. Although we have many more observations for TSR, I will use the same period to compare the differences between the metrics so that I am comparing like for like.

Below are the two charts showing graphically TSR’s relationship to points and TSR year-on-year, the R² for each are 0.65 and 0.70 respectively, and it’s against these that I’ll be comparing the new metrics.

Moving on to the BCR, the graph below is the one I used in the presentation to show the relationship between the BCR and points, although with last season’s data also included. I don’t think there is anything ground breaking in looking at the BCR, as it is essentially using the same method used for TSR, but applying it to Big Chances.  I have seen the BCR used by others to compare teams, although as far as I am aware, no one has written about it before to show just how meaningful it can be.

The R² was 0.75 over the 4 full seasons that Big Chances have been recorded by Opta. Just like TSR, the average team has a BCR of 0.5, but you can see from the graph that the range in BCRs is larger, from the 0.3 achieved by Reading two seasons ago up to the 0.77 for Manchester City also from two seasons ago.

So, why do Big Chances, with an average of only 68 per team each season have such a strong relationship with points won? Well, it is partly be a case of correlation rather causation. Teams that are winning tend to be more conservative and sit back, restricting the opposition to more difficult shots, whilst also being able to hit teams on the break to create better chances or be more patient and wait for easier opportunities to come along. As I showed in my presentation, this comes through when looking BCRs by game state, teams that are winning by 1 goal on average have a BCR of 0.53, and this increases with each goal the lead increases by, and needless to say that teams that are in the lead more tend to win more points. However, as shown by Mark Taylor (here), the ability to create better chances can also be an important factor in who wins the game, even if the expected goals for each team in a game equal the same.

But what about repeatability, is there ‘skill’ in a team’s ability to be able to both create and restrict Big Chances? Well, the graph below shows there is a positive relationship between the BCR in one year and the following year, but this is not a strong as for TSR, with the R² for BCR at 0.60 as opposed to 0.70 for TSR. However, as one of the strengths of TSR is that there are a large number of shots, we should remember that by looking at Big Chances only, we have significantly reduced the number of observations, and when you consider this, the repeatability is actually quite high.

There are still a lot of shots left over however, so how much information is there in shots which aren’t Big Chances? I'll refer to these as Normal Chances, and to give a bit of extra detail, whilst Big Chances are converted at a rate of about 38%, Normal Chances are converted at a rate of slightly over 5%. Well, from the graphs below, we can see that the relationship between the Normal Chance Ratio ("NCR") and points is not as strong as for BCR, whilst the year on year correlation is slightly higher, with an R² of 0.55 and 0.63 respectively. As you would expect with Normal Chances making up 87% of all shots, the range is similar to what we see for TSR, going from 0.36 up to the 0.67. The average NCR, as with both TSR and BCR, is also 0.5.

As about 50% of goals are scored from Big Chances, with of course the other 50% from Normal Chances, I thought it would make sense to see what happens if we add the team’s BCRs and NCRs together. As they both have an average of 0.5 across all teams, the combined metric will have an average of 1 so it will also be nice and easy to tell which teams are above or below average. Of course this metric needs a confusing name and acronym, and as it is two ratios based on the quality of chances added together, I’ve called it Chance Quality Ratios Added Together (“CQR+”).

We can see from the graph that the relationship between CQR+ and points is very strong, and has an R² of 0.78. The reason for the strong improvement over TSR I think can be explained by thinking of TSR as a weighted average of the BCR and NCR, and by separating them out and adding them back together, we have given each an equal weighting, which is in line with their average contributions to goals. As Normal Chances make up the vast majority of a team’s shots, then their TSR and NCR will always be relatively close. If a team is more efficient at creating and restricting Big Chances than they are at shots in general, then their BCR will be higher than their TSR, whilst their NCR will lower, however the change in BCR will be larger in absolute terms, and the higher conversion rate associated with Big Chances should in general translate into more points won.

How about the repeatability? The addition of BCR and NCR together also has a positive effect on repeatability, and the R² 0.75 is actually higher than for TSR which is 0.70 over the same period.

To summarise the differences between the four metrics that I have covered, the table below shows the R² for each one. As we know, TSR is a good predictor of points and is repeatable, BCR has a stronger relationship to points than TSR, but is not as stable year-on-year, but by adding the team's BCR and NCR together we have a metric with both a higher explanatory power and a greater predictive power.

I thought it would be interesting to check how teams are performing by these metrics this season so far. The table below shows each teams TSR, NCR, BCR and CQR+, with the ranking in the league for each metric, and the table has been sorted by CQR+.

If we look at TSR compared to league position we can see that, as we might expect, it is performing relatively well, with the majority of TSR rankings within 3 places of the league position. We can also see how the NCR is never more than 2 decimal points different from the TSR, although even these small changes do shuffle the rankings a little. Its when we start to look at the BCRs that we start to see the real differences. Chelsea have the highest BCR at 0.71, on the back of being the meanest defence at conceding Big Chances and creating the 2^nd most (behind Arsenal), which compares to their TSR of 0.61. Moving in the other direction we have Liverpool, who have been good at dominating the shot count in their matches and have the 4^th highest TSR of 0.59, however they are not as efficient when it comes to Big Chances and have a BCR of 0.51, ranking them 8^th.

In terms of the CQR+ metric, and on the back of their high BCRs, Chelsea and Arsenal are the leaders of the pack. Man City have been the most dominate team in terms of TSR this season, but like Liverpool they are not efficient when it comes to Big Chances so rank 3^rd by CQR+. Then come Southampton, Liverpool, Manchester United and a bit of a gap to Tottenham, meaning that the top 7 by CQR+ make up the top 7 teams in the league. Down at the other end, 5 of the bottom 6 teams in the league make up the 5 worst CQR+ teams, so it does seem to be working on first sight, and overall, all but 3 of the teams’ CQR+ rank are within 3 places of where they are in the league. 

That does mean there are 3 outliers however. The first two are West Ham and Swansea, who are outperforming their CQR+, where they rank 14^th and 15^th respectively. For both teams, particularly West Ham, it may well be the case that their numbers are being affected by game states. West Ham have so far spent the 5^th highest amount of time winning in the league this season which is likely to be having some downward pressure on their shot numbers, and whilst Swansea haven’t spent as much time winning, they did start the season very strongly. Going in the other direction, by far the biggest underperformance compared to their CQR+ are QPR, currently sitting outside of the relegation positions by goal difference, yet ranked 12^th in TSR, and 10^th in CQR+ thanks to having the 7^th best BCR in the league. Have they been a little unfortunate, or is it the consequence of playing in very open matches when you are not actually very good? I’m afraid I haven’t seen enough of them to know.

I’m not too sure how this metric stacks up against some of the others out there, particularly the expected goals model, although I did use it to enter a prediction in Simon Gleave's Premier Leagueprediction analysis though, which is kindly updated by James Grayson through out the season to show how everyone is doing, and where it is performing relatively well (by points at least, although not by position). 

However, compared to some of the models out there, it has the benefit of being very simple to calculate. All you need is the total shots taken and faced by each team, and the total number of Big Chances they have taken and faced. Unfortunately Big Chance isn’t as readily available as most other stats, you can get it from FantasyFootballScout where you have to pay a subscription, and I have also recently been directed to the AllThingsFPL website which also has them, although having said that the two sites do show slight differences in the Big Chances for each team here and there. I do know that Big Chances do get reviewed in the week following the game which may account for the differences, but I don’t know which site is the most up to date (I have been using the FantasyFootballScout numbers).

I hope that that I have shown here, as well as in my previous work just how useful the Big Chance stat can be and how we can use it to make some simple metrics. Due to not having a precise definition of what a Big Chance actually is, the lack of detailed information on all the Big Chances, as well as its subjective nature, some in the ‘fanalyst’ community have their doubts about Big Chances. Whilst I agree that there may be cases where a shot is recorded as a Big Chance when possibly it should not, that is always going to be the case when subjectivity is added, however I believe that these will be in the minority, and as we see consistency each season with the number of Big Chances and their conversion rates, they are not having a big impact. I think that we should embrace subjective stats as they can add more context to our analysis, and the benefits can outweigh the concerns. 

The slides from my OptaPro Conference presentation

Opta have just announced that they are now taking submissions for The OptaPro Analytics Forum 2015 (more details here).

Earlier this year I was lucky enough to have been picked to present at their first conference, so I thought it would be a good time to post the slides.

If anyone out there is thinking of submitting an idea I can highly recommend it (although as someone who doesn't really do that type of thing, it was bloody stressful!). I wasn't going to bother last year as I thought it would never get picked, but decided on the deadline for submissions that I had nothing to lose by giving it a go and I would never get the data to do something like that otherwise.

So, if you've got an idea, get it in!

Woolcock opta pro analytics forum with links from TheWoolster

Creating some new metrics using Opta's Clear Cut Chance stat

Having read this from Richard Whitall (here), I thought I’d see if I could come up with any metrics using Opta’s Clear Cut Chance (CCC) that both explains what has been happening in season and that might also be used to predict what may happen going forwards. This would build on the work I have done already looking at how efficient teams are at attacking (here) and defending (here).

I’ll start by adding a team’s Creative Efficiency (CE) and it’s Defensive Efficiency (DE) together, which I’ll call the CEDE Score. As a recap, a team’s CE is calculated by dividing a team’s CCCs for by their total shots, the higher the number the more efficient they are at creating CCCs when they do take a shot. Conversely a team’s DE is calculated by dividing their Normal Chances conceded (ie non CCCs) by their total shots conceded.

The average CE over the previous 3 seasons was 13%, whilst the average DE was 87%, add the two together (in their decimal form) and you get CEDE Score of 1.0 for the average team in the league.  We now have a metric to measure the overall efficiency of a team in terms of creating and restricting CCCs. A team with a CEDE Score of over 1.0 is more efficient than the average team and vice a versa.

As I said at the beginning, I would like to create some metrics that both explain what has happened in matches and that high prediction values. To test their explanatory value I will calculate their relationship with goal difference (GD) and to test their predictive value I will calculate the correlation between the year 1 values achieved by each team with their year 2 values.

Below is the graph of the CEDE Score to GD over the past 3 years. Whilst it would be good to have more, CCC data has only been available for the last 3 full seasons. We can see that there is a positive relationship, a team with the average CEDE Score of 1.0 would be expected to have the same GD as that of the average team, which is zero, whilst a team with a CEDE Score 5% above average at 1.05 would be expected to have a positive GD of about 17 goals. However there is some variance in the data and the R² is equal to 0.35, which is not bad, but I would like to do better.

As for the repeatability of the CEDE Score, with only 2 seasons where we can measure a team’s performance against the previous season, and as we can only look at nonrelegated teams, we have even less data with only 34 observations. As you can see from the graph below, the relationship is somewhat weaker, and the correlation coefficient is 0.42.

To see how this compares to other footy ‘fancy stats’, then if you haven’t already, I highly recommend you have a look at James Grayson’s blog, who has done a huge amount of work on repeatability of various football stats. The reason for the low score could well be due to the relatively small number of CCCs created each season. The average team only takes and concedes 70 CCCs in a season, so small differences year on year can have a big impact.

One of the main issues with the CEDE Score is that, as an efficiency score, it is only a proportion of total shots and does not take into account how many shots a team take or faces. For example Stoke under Pulis were a team that had fewer shots attempted in their matches than average, they attempted to take their shots from as close to goal as possible, whilst reducing the opposition to shoot form distance shots. Over the last 3 season’s their CEDE Score has been above average, and in the 2012/13 season they actually had the 4^th best CEDE Score, however this did not translate into league position.

If you are reading this then there is a good chance that you know about the Total Shot Ratio (TSR), but in case you don’t, TSR measures the proportion of shots that a team has taken. TSR is essentially the go to metric in football if you want to compare teams. It has been shown to be meaningful, repeatable, is simple to understand and calculate. Although you can find a lot more detailed info and data about TSR on James’ blog, I thought it would be useful to see the same graphs over the same time period as what I’m working with for ease of comparison. Over the 3 seasons, TSR to GD has an R² of 0.66, and the year on year TSR correlation coefficient is 0.86.

One of the main issues with regards to TSR is that it treats all shots as equally and does not take into account chance quality. So, what we already have is a very good metric that does not take into account chance quality, and what I’ve introduced is an ok metric that focuses on the efficiency of a team to create good chances for and restrict good chances against. Lets see what we get when we combine the two…

Handily, as the CEDE Score has an average of 1.0, we can use it as an efficiency multiplier with TSR, I’ll call it CEDE’d TSR (TSR^CEDE). The effect should be to normailse TSR so that there are fewer outliers. We can instantly see from the graph of TSR^CEDE to GD that the observations are packed closer to the trend line than the TSR graph. The R² for TSR^CEDE is in fact 0.79, which I think is quite a good increase. 

To give some examples of the effect this has had, last season Tottenham were the top ranked TSR team last season with a score of 0.65, their CEDE’d TSR dropped down to 0.60 which ranked 4^th. It has often been mentioned that Man Utd broke many a model last season, mainly because their TSR of 0.53 ranked them only 7^th, however once CEDE’d it increased to 0.60 and slightly higher than Tottenham’s to rank 3^rd. Ok, this still does not quite predict them having the stellar season that they did, but it is a big move in the right direction

In terms of the year on year correlation, the graph looks pretty similar to that of TSR, and its correlation coefficient is 0.87, which is slightly stronger that TSR’s year on year correlation from the same period.

So, how are things looking this season then? Hopefully the table below is pretty self explanatory, its sorted by TSRCEDE and also lists each team’s CE, DE, CEDE Score and TSR, ad shows the rank of the CEDE Score and TSR as well as TSR^CEDE.

The top ranked team in terms of CEDE Score is table topping Arsenal, which corresponds to what Colin Trainor has found in the first of his mid-year reviews (here) and recent table toppers Liverpool in 2nd. However surprisingly the next 6 teams are made up of what we would consider weaker sides, West Brom in 3rd, followed by Stoke who have a good CEDE Score again, Sunderland, Hull, Cardiff and Fulham. The reason for this in the main I believe is that these teams, in effect, know their limitations and attempt to pack their defences to reduce the opposition to difficult shots, I’ll be interested to see if Colin’s reviews corroborate these numbers when he comes to looking at these teams. Man City only come in 14th by this measure and actually have a CEDE Score of less than 1.0 due to their poor DE.

I have to say that this is unusual, in general it is the better teams that have the better CEDE Score. What has happened is that 4 of the biggest shooting teams have seen their CE fall significantly. As these 4 teams, Man Utd, Man City, Chelsea and Everton, also changed their manager in the summer, we’ll have to wait and see if this is a short term effect of having a new manager bed in, or in fact due to different tactics employed by those managers. The overall effect is that the average CE for all teams this season has fallen to under 11%.

If we look at the TSR^CEDE numbers compared to TSR, we see they are, as you would expect, very similar, however the changes are enough to shuffle the pack a bit. Some of the notable changes are Chelsea becoming the top ranked team from 3^rd in terms of normal TSR, Arsenal go from a Utd like 8^th ranked TSR team to 5^th by TSR^CEDE, and Everton do the opposite and go from 5^th to 8^th. At the bottom of the table we’ve got 5 teams grouped together on a TSR^CEDE of 0.43, with Cardiff and then Fulham adrift at the bottom.

I’ll leave it there for now, but hopefully I’ll have a bit of time to be able to a few other metrics I’ve put together using CCCs.



Sources: 2013-14 season from http://www.fantasyfootballscout.co.uk/, all earlier data from http://eplindex.com/

RACE to Goals Model – The Defence

Prior to the season starting, I introduced the RACE to Goals Model, which you can find here, and I suggest you have a read of that before you continue with this one if you want to have a full description of the different metrics and how they are calculated.

Essentially, I am looking at the same metrics, but this time flipped to a defensive point of view, so the rate of shots conceded, the Defensive Efficiency, and the conversion of chances conceded by each type.

I will describe Defensive Efficiency here though, as it’s calculated slightly differently. Whereas Creative Efficiency attempts to show how good a team is at creating good chances, measured as the proportion of Clear Cut Chances to Total Shots, Defensive Efficiency attempts to show how good a team is limiting the amount of good chances the opposition has, and is measured as the proportion of Normal Chances conceded to Total Shots conceded (%NC). So the higher the number, the lower the percentage of Clear Cut Chances conceded, and the more efficient the defence is.

The benchmark numbers are essentially the same, the slight difference being own goals, and those ‘shots’ by players on the defending team that lead to own goals are also included.

The table below shows how well the teams performed last season against the 4 metrics.

The team that conceded the fewest shots was Tottenham, with only 370 over the entire season, so a touch under 10 shots a game. At the other end of the scale were Reading, who conceded 706 shots, the worst by over 60 shots.

Like with the original article, I feel the raw numbers in the table are a little hard to read, so again I’ll add context and measure each metric as the percentage difference from the benchmark team. From the a defensive point of view, having shots and conversion rates below the benchmark is good, but this is not the case for Defensive Efficiency, so I’ve highlighted this in the table as anything in red as being ‘bad’.

As with Creative Efficiency, Manchester United also had the best Defensive Efficiency, limiting their opponents to only 8.2% of their shots coming from CCCs, with Manchester City being the only other team to have a Defensive Efficiency of over 90%, seeing them perform 6% and 4% better than average respectively. The team with the worst Defensive Efficiency was Newcastle, who allowed over 18.5% of all chances against to be CCCs; however the 2nd worst team, perhaps surprisingly considering how few shots they conceded, was Tottenham, allowing almost 18%, and possibly showing the risk of playing with a high defensive line.

Looking at the conversion rates it becomes clear why Wigan struggled last season. They had by the worst rate of CCCs conceded, in fact at 52.7%, they are the only team over the 3 years of data that conceded more than half the CCCs that they faced. They were also the 2nd worst at stopping Normal Chances being conceded. Reading actually had the best rate when it came to stopping CCCs in the league, but unfortunately for them, when you allow the opposition to create over 100 CCCs in total, you will still concede a lot of goals.

No teams outperformed or underperformed all 4 of the metrics compared to the benchmark. Only 4 teams, the two Manchester clubs, Chelsea and Swansea outperformed on 3 of the benchmarks. Liverpool join Utd, City and Chelsea as the only teams who conceded fewer shots than the benchmark whilst also having a higher than average Defensive Efficiency. Despite conceding the fewest shots, we can see why 7 teams conceded less goals than Tottenham following their underperformance in the 3 other metrics.

Converting the metrics into Expected Goals, we see how badly Wigan performed. Whilst they would have been expected to concede just less than 54 goals from the shots that the opposition had, which was only the 10th lowest, they actually conceded 73 (+19.1 goals more than expected). The other big underperformers were Southampton (+9.7 goals), Newcastle (+9.0 goals) and Aston Villa (+7.9 goals). The biggest overperformers were Everton (-9.5 goals), Sunderland (-8.7 goals), Stoke (-6.4 goals) and Arsenal (-6.0 goals).

In my next posts I will combine some of the attacking and defending metrics together to analyse team’s performances in some new ways, and see how the teams have performed so far this season.

This was originally posted on  EPLIndex  http://eplindex.com/43205/race-goals-model-defence.html

Defending Liverpool's Defence

With the season about to start, I thought I would follow up to piece that I did earlier in the year looking at how Liverpool’s form changed over the season, however whilst that looked at attacking form, this one looks at Liverpool’s defensive form. Again I will look at Liverpool’s performance compared to how the league performed on average, how the top 4 performed, and also compared to Liverpool in the 2011-12 season, as well as having the short term form by having the 6-game moving average. One thing to note is that due to there being fewer observations, for example Liverpool conceded far fewer shots, goals etc., that the graphs show more extreme changes compared to the attacking versions of these graphs

I’ll start by looking at shots conceded per game. Apart from the 18 shots conceded in the first game of the season against West Brom skewing the averages, Liverpool performed more or less in line with the Top 4 teams throughout the season.

In terms of the accuracy of opposing team’s shots, despite the slow start that Liverpool had and perhaps surprisingly, they actually allowed significantly less shots to hit the target compared to the Top 4 teams and the rest of the league over the first half of the season, whilst over the 2nd half of the season, a greater percentage of opponents shots were hitting the target.

Moving on to Opponent Shots Conversion and Shots on Target Conversion, we can see how poorly Liverpool defended and Pepe Reina performed in the opening 5 or 6 games of last season. Basically, Liverpool defended and kept goal more or less like a lower league team when going up against a Premier League side in a cup, but this quickly regressed to the mean, and they performed like a Top 4 team from game 7 onwards (in the moving-average plot, this shows up from match 12). Those first 6 games had such an impact though that the end of season conversion rates were still only in line with the league as a whole.

How do Liverpool, or more pertinently Pepe Reina and Brad Jones, do at keeping out Clear Cut Chances (CCC)? So what is a CCC? It is one of Opta’s few subjective stats that can broadly be described as a chance where the attacker is probably central to goal with only the keeper to beat. So a keeper would hope to either save it, or perhaps attempt to put the attacker off sufficiently that they miss. As I mentioned in the original piece, the conversion rate for CCCs is much more variable than the other conversion rates, this is because in some games there will be few or even no CCCs, which means that both very high and very low single game conversion rates are far more likely, and we see this clearly in Liverpool’s form plot (note that the reason you can’t see the league average plot is because it was the same as the Top 4). Again, Liverpool started poorly, but were better than the Top 4 teams from match 7 onwards, apart from a large peak at match 17 where all the CCCs that Liverpool faced were scored giving a 100% conversion rate. More specifically, it was in fact a 4 match period with the goals coming from Tottenham, West Ham and Aston Villa.

With that in mind, it is interesting to then see the rate at which Liverpool were giving up CCCs per game through the season. Again we see Liverpool started off poorly, giving away on average 1.5 CCCs over the first 10 games, but by match 17, where we saw the 100% conversion rate, the 6-game form had fallen to 0.7 per game. So, it was only 4 out of 4 CCCs conceded in 6 games. As the average conversion rate for all CCCs is around 38%, it is a bit like tossing a coin 4 times and getting 4 heads, so I don’t think we should put it down to poor goal keeping. You’ll notice there is a sharp rise in CCCs conceded from about match 20, but this coincided with an increase with CCCs for Liverpool, and can perhaps be put down to increased attacking leaving the defence more open (Note: Liverpool’s average in 2011-12 was the same as the Top 4’s last season).

Finally I’m going to look at Errors per Game. It should be noted that these are ‘on the ball’ errors, so does not include an error like not marking the run of an opponent from crosses (something that had many Liverpool fans pulling their hair out). Again we see the effect that Liverpool’s poor start to the season had, however it took longer for Liverpool to recover from than compared to the other metrics, but by the end of the season, on the ball errors had almost become non-existent. Over the season as a whole, only Arsenal and Newcastle made more errors than Liverpool’s 36, however if you split the season in half, over the first 19 games Liverpool made 28 errors, over the last 19 games it was only 8. As an on the ball error will often leave the rest of the defence wrong footed, these types of errors tend to have a high conversion rate, and Liverpool conceded 10 goals from the 36 errors they made. If Liverpool can continue to keep the error rate at the level of the 2nd half of the season, then there would be a lot less hair pulled out by the fans this coming season. 

Perhaps it was the tough start Liverpool had, perhaps it was the getting used to Brendan Rodgers system, or perhaps they were just unlucky (probably a combination of all 3), but clearly Liverpool started the season really badly last year. If they can perform defensively as well as they did over the last 30 or so games, they could well turn a few of those losses and score draws into wins, and have a good crack at finishing in the top 4.

I posted this oginally on EPLIndex  http://eplindex.com/37116/defending-liverpools-defence-statistical-analysis-1213.html

RACE to Goals Model: League Predictions

At the end of my introduction to the RACE to Goals Model, that you can read here, I mentioned that I would like to look at how teams performed from a defensive point of view, and to check  how reliable the metrics are year on year.

Whilst I have done those things, I haven't had the time to sit down and write about them. But I have been able to create a model (well, I created a few slightly different versions and picked what seemed the best) to predict this season's league table, so I will at least get that posted prior to the season starting so that I don't let bias from early results get in the way.

The model is a variation of the Shot Dominance model, as coined by @mixedknuts (here), which is itself a variation of the Total Shot Ratio (TSR) model that has been looked at in legnth by @JamesWGrayson (such as this), and a good summary of TSR by @TheM_L_G can be read here. What differentiates the RACE to Goals Model is that it includes the quality of chance that those metrics are missing based on the metrics from my earlier piece, and I hope to go into more detail in later posts.

For the promoted teams I did not have the data available to do the same analysis, so a simply did a regression of goals scored and conceded in the Championship since 2000 for promoted team compared to points scored the following season in the Premier League.

The model predicts that Man City are the clear favourites for the title, and that there will be another very close battle for 4th place, whilst down at the bottom, Fulham, Newcastle Southampton and Norwich could well be in trouble. Of course the model does not take into account any managerial changes or player transfers, which could change team strengths significantly. Personally, I'd expect Newcastle and Southampton to do a little better and be replaced by Sunderland down towards the bottom, whilst common sense would say that Chelsea should challenge at the top, but other than that, I think the predictions are reasonable.

Anyway, here's the predicted table.

Rate of Attack and Creative Efficiency (RACE) to Goals Model

Yes, yet another model looking at the quality of chances and finishing of football teams.

This is something that I had hoped to have finished before last season ended, unfortunately life got in the way and it got delayed. Since then there have been a number of very interesting analyses done, including those by @colinttrainor (like this) and @11tegen11 (like this), which have continued the good work done by @footballfactman (like this), where they have put in a significant amount of work to look at where shots are taken from and what the conversion rates are for shots from those areas. Hopefully I can hang on to their coattails.

Personally, I am far too lazy to collect all that data, so I have let the experts (Opta) decide upon chance quality for me, and I hope to make the model as simple as possible. In my blog so far I have looked at how Liverpool and Tottenham have performed in terms of finishing and creativity, and to add some context, I compared them to the league and Top 4 average. Whilst compiling the numbers, I noticed that the League averages were quite consistent year on year over each of the past 3 seasons, and realised that I could create a theoretical average team that I could use as a benchmark to compare the performance of all the Premier League teams.

I am sure I am not telling anyone anything new when I say that the amount of goals a team scores is essentially dependent on 3 things, the amount of shots they take, the quality of chances they create, and the quality of their finishing, and it’s against these metrics that I will be comparing teams against.

Rate of Attack

This is very simply the amount of shots a team takes, and can be measured on a per game (SpG) or per season (SpS) basis. Yes, I know that not all attacks end with a shot and I am basically just using total shots, but I wanted the model to have a 'racey' acronym, so Rate of Attack it is.

On average, each team takes about 14.5 SpG, or about 550 SpS. Between 9-10% of all shots end up with a goal, and this has been found to be consistent season upon season and across different leagues. For those that don’t know, this is called the Reep Ratio, after an amateur statistician named Charles Reep, who looked at various stats, including the conversion rate of shots, in the 1950s. 

Creative Efficiency (%CCC)

This is a measure of the creativity of the team and quality of chances they have, and this is where I am relying on Opta to decide upon what is a good chance, as I am using their Clear Cut Chance (CCC) for this. A CCC is one of Opta’s few subjective statistics, and whilst a full description is not given, a brief description is given by Opta in their Event Definitions under Big Chance (here) 

“A situation where a player should reasonably be expected to score usually in a one-on-one scenario or from very close range.”

Creative Efficiency (%CCC) is measured as a proportion of Clear Cut Chances to Total Shots.

A team with a high %CCC will, over time, create chances that are easier to score from than the average team. Whilst CCCs make up only about 13% of all shots in the Premier League, they are vitally important, as for each of the last 3 seasons, around 52% of all goals have been scored from a CCC. It should be noted that CCCs include penalties, and whilst I did consider removing them from the analysis as they have their own average conversion rate, I decided to include them for a few reasons, there will be some open play CCCs that will be easier to score from than a penalty, I also think that teams that attack more or are more creative will tend to get more penalties, at least over the long term, and that should be included in their Creative Efficiency, and finally because I want to keep the model simple and with as few adjustments as possible. 

Obviously when you multiply a team’s Rate of Attack by their %CCC, you will get the number of shots which are CCCs. The remaining shots will be what I will call, as I can’t think of a more appropriate term, the Non-CCCs. The two types of shot have their own average conversion rate, and the model analyses the quality of finishing of both types of chance by comparing the goal expectancy (number of chances multiplied by the average conversion rate) to actual goals scored for each type of chance.

CCC Conversion

To give an indication of the average difficulty of a CCC compared to the average shot, it is on average about 4x easier to score a CCC as they have an average conversion rate of just under 38%. It should be remembered though that there is a large range in the probability of a CCC being scored, Sam Green of Opta has said (here) he considers the base probability to start at about 20% and it of course goes up to 100%.

Non-CCC Conversion

The average conversion rate of Non-CCCs is slightly above 5%. The reason why I won’t classify them along the lines of a ‘difficult’ chance is that with the goal expectancy range for individual shots being between 0% and 20%, anything with an expectancy above 10% will still be easier than average.

The Numbers  

Here are the hard numbers I have collected for the past 3 seasons.

 

And these are the benchmark ratios/rates that I have either mentioned or will be using for the theoretical average team.

So, how did each team perform last season? In terms of number of shots, Liverpool lead the way by far with 740 shots over the season, 59 more than Tottenham took, the next best team, and not far off double the amount of shots that Stoke had.

It may not come as much of a surprise to see that Manchester United had the best %CCC, with 21% of the efforts being from a CCC, compared to 18.3% for 2^nd placed Manchester City. To put this difference into perspective, whilst Man City took 98 more shots than Man Utd, they only had 3 more CCCs. Liverpool had 178 more shots, but with a %CCC of ‘only’ 13.6% (still above average) had 17 less CCCs.

In terms of shot conversion, the team with the best conversion rate for CCCs was, yes you’ve guessed it, the team who scored the most goals, Manchester United with 44.1% of them scored. The team with the worst conversion of CCCs was, yes you’ve guessed it, the team who scored the least…oh, it was actually Manchester City, with only 28.9%, I didn’t guess it either. So, City had 3 more CCCs, but scored 17 fewer CCCs, a significant amount.

The team with the best conversion rate of Non-CCCs was Chelsea at 7.4% leading to goals, and this time we do find the expected QPR at the bottom of the pile with only 3.1%.

I’ll admit that the table above is a little hard to read though, we’ve got different units and magnitudes of measurement and its hard to see how well each team is doing overall, so lets add some context and measure each teams performance as the percentage change from our benchmark team.

Now things become a bit clearer. We can see that despite only taking 3% more shots more than the benchmark, Manchester United’s %CCC was a whopping 62% higher than average, which goes some way to explaining why their total shot conversion was so much stronger than everyone else at 14.2%. However they also significantly outperformed both conversion rate metrics, meaning they scored almost 13 goals more than expected if they had average finishing. If they had scored at average rates, their total conversion rate would still have been the highest in the league though at a touch under 12%.

Only 2 teams managed to beat the benchmark for all 4 metrics, Man Utd and Arsenal. Of the other top teams, Chelsea and Tottenham had a relatively poor %CCC, Man City were poor at converting their CCCs, Liverpool were poor at converting their Non-CCCs, and Everton were poor at converting both types of chances.

At the other end of the table, only 2 teams performed worse on all 4 metrics compared to the benchmark as well, unsurprisingly QPR, with the other team being Newcastle. Reading were very good at finishing their chances, its just that they struggled to create any.

So what does this all look like when we convert these metrics to expected goals and how did the teams compare? There were 3 big outperformers, Chelsea (15.7 goals above expected), Man Utd (+12.6), and Arsenal (+10.4) whilst there was 2 big underperformers in QPR (-12.3)  and Everton (-10.1). For those of you who are into your ‘proper’ statistics, I’ve calculated the Mean Absolute Percentage Error for the model over the last 3 seasons as 10% and the Root Mean Squared Error as 7 goals. Its been a loooong time since I studied statistical methods, so I may have used the wrong error measurements, but I think that shows that the model isn’t too bad.

I’ll finish with how my model differs from those I’ve mentioned which look at shot location. I’ll start with the weaknesses. The first is that my model is far less granular as I have lumped the 87% of all shots that are Non-CCCs with the same goal expectancy, which means that the type of analysis that I can do with my model probably can’t go quite as deep as the others. Due to the creative efficiency element, I think the model is only applicable to teams and won’t be able to do player analysis. There is an element of trust in Opta that they are consistent when collecting the CCC data as it is subjective, particularly as we do not know their precise definition, although having read this (here), I think its fair to assume they are consistent. And because we do not know exactly how Opta define CCC, I think it will be very difficult to see how or if the metrics change depending on the Game state as the info of when a CCC occurred is not available. Whilst on average there are just under 4 CCCs per game, so it might be possible by watching the highlights or reading the match reports to figure it out for most games, in some cases however, as shown by @analysesport (here), it would be very difficult. Another issue is the relative lack of CCC data, it is not freely available (you need to pay for a subscription at www.eplindex.com for the data), it only goes back 3 years, and as far as I am aware, there is no  CCC data publicly available for leagues other than the Premier League

The positives are that it is very easy to collect and analyse the data, you only need the number of games played by a team, their total shots and the number of CCCs they’ve had to be able to estimate the number of goals they should have scored. One of the issues with simply using shot location, as discussed by @mixedknuts (here), is that it does not take into account the positioning of the defenders. For instance a player may take a shot in the central area of the box but have 4 defenders and the keeper between him and the goal, so the probability of a goal would be low, equally a player may break an offside trap and have the ball outside the area but be 1 on 1 with the keeper, so the likelihood of scoring would be quite high. This model at least separates out those chances where the defenders are not making a significant difference to the difficulty of a goal being scored, and whilst these only make up 13% of the chances, they do make up 52% of the goals.

Hopefully, if I get enough time, I’ll look at how repeatable these metrics are and if they could be of use for predicting matches and also look at how the teams performed on these metrics from a defensive point of view.   

You can follow me on twitter at @The_Woolster


Data taken from www.eplindex.com