Here is the list of the relative strength of teams in the MLS based on their performance in the previous season according to Wyscout data. If it were a simple competition than this would also be our prediction for the league table. Previously predictions like these have had a 80% correlation with the actual league table a year later. So will see how it goes for the MLS:

Rank		FBM Wyscout score
1	Atlanta	73
2	Los Angeles	72
3	Salt Lake	71
4	Toronto	70
5	Dallas	67
6	New York City	65
7	Columbus	63
8	Seattle	62
9	Portland	61
10	Chigago	57
11	Minnesota	56
12	San Jose	56
13	Colorado	54
14	New England	54
15	LA Galaxy	52
16	Montreal	51
17	Philadelphia	50
18	Kansas City	48
19	DC	45
20	Orlando	42
21	New York RB	40
22	Houston	37
23	Vancouver	36
24	Nashville	35
25	Cincinnati	30
26	Miami	19

Many scouts wonder why their advice is being ignored by the higher ups. The reason is that whatever scouting report they have drawn up, their report fails to answer the most important question:

What is the probability that player X is able to contribute to the team?

Time to define “analytics” for nba fans. At least how we do it. 1. Team Analytics – style of play that optimizes your PPP 2. Player Analytics – The probability that a player is able to contribute to maximizing the team PPP

— Mark Cuban (@mcuban) June 20, 2019

The answer is a number between 0% and 100%. This answer is never given in any of the reports or presentation scouts give. That means that the decision makers have to calculate this answer based on the report the scout has provided. Of course, they never do this consciously. Yet, our brain makes these kinds of estimations unconsciously all the time. If a scout does NOT explicitly answer this question, the brain of the decision maker is going to make the probability estimation all by himself. In almost all cases, this estimation will be lower than the players the decision maker prefers himself. That is the reason why even the most successful scouts only have contributed to a handful of transfers. Most transfers happen for other reasons than provided by the scouting team.

It really doesn’t matter whether we are talking about data, video or live scouting. If the final report fails to answer the question about the probability that a player is able to contribute to the team, the decision maker is going to answer that question and probably in a less favorable way.

So let’s look at an example. If you are using Wyscout data as a data scout, how can you then answer this most important question: 

What is the probability that player X is able to contribute to the team?

First you need to build a model that transforms Wyscout data into probabilities. Bayesian networks are most suitable for this job, but there are other ways. We prefer to use Bayesian networks. Second step is to validate your model. For validation we have created a Bayesian network to transform Wyscout team data into team probabilities. We calculate the following four probabilities:

1. What is the probability that a team is going to perform well?
2. What is the probability that the attack of the team is going to perform well?
3. What is the probability that the defense of the team is going to perform well?
4. What is the probability that the passing game of the team is going to perform well?

Here are the results for the Premier League and Watford:

Validation comes from the 89% correlation (R2=80%) between the probability to perform well and the rank of the team. This is in line with this correlation in other competitions. So to be clear: 

1. The probability of Watford to perform well is 38%
2. The probability of Watford to attack well is 37%
3. The probability of Watford to defend well is 54%
4. The probability of Watford to pass well is 47%

The next step is to look at the individual players of Watford. Normally we would look at all the players (except the keeper), but for this exercise we only look at the most recent starting XI:

Again, these stats answer the following four questions:

1. What is the probability that a player is able to contribute to the team?
2. What is the probability that a player is able to the attack of the team?
3. What is the probability that a player is able to the defense of the team?
4. What is the probability that a player is able to contribute to the passing game of the team?

As long as a player has at least one of these four probabilities quite high, he is an asset to the team. Of course, if it is only one category, he is a specialist rather than a generalist, unless that category is the overall category.

Taking into account minutes played we can then calculate the contribution each player has made to the team probabilities of Watford:

The contribution of these ten players is:

Here one can see that although Sarr has quite weak data in Wyscout, his contribution to the attack of Watford is on par to what is expected of him.

One can also immediately see that Pereya is the weakest link. So let’s look at a replacement for Pereya. As this is an example only, I am going to use a replacement who obviously would be better suited than Pereya. The player I am going to use is Liverpool’s Mané.

Here we use our transfer model. This gives the following results:

Let me explain this. First we start with the probabilities of Watford and Pereya we have already seen. Taking into account minutes played, we subtract Pereya from the probabilities of Watford. What this means for Watford is that the probability to perform well remains unchanged, but the probabilities to attack, defend and pass well drop a bit. 

Then we look at the probabilities of Mané playing at Liverpool. As you can see, for all but defense, these probabilities are much higher than Pereya’s probabilities. But in part, Mané is playing well at Liverpool because he is playing together with other great players. That won’t be the case if he transfers to Watford. So we have to take into account that his performance will drop a bit. But how much? Fortunately, we have a Bayesian model to calculate precisely that by taking into account the relative strength of both teams and minutes played. To make it explicit:

1. The probability that Mané is able to contribute to Watford is 87%.
2. The probability that Mané is able to contribute to the attack of Watford is 98%.
3. The probability that Mané is able to contribute to the defense Watford is 5%.
4. The probability that Mané is able to contribute to the passing game of Watford is 52%.

What this would mean for Watford is that their probabilities also go up when we add Mané with his Watford probabilities to Watford as is shown in the final row. With Mané playing for Watford the new probabilities for Watford are:

1. The probability of Watford to perform well is 45%
2. The probability of Watford to attack well is 43%
3. The probability of Watford to defend well is 46%
4. The probability of Watford to pass well is 55%

GIven the correlation between overall team performance probability and rank, we can also see that Watford would rise to somewhere between rank 10 and rank 15 in the competition once Mané is playing for Watford. 
Rational decision makers use these kinds of models to calculate for every player they are seriously considering hiring what the probability is that the player is able to contribute to the team and what this means for the team. Once you have ranked all players according to their probability to be able to contribute to the team, you try to hire the best player available. That is how we were able to transfer Dalmau to Heracles for instance.

This is the kind of work that we are going to teach at the Football Behavior Management summer school at the VU-university in Amsterdam in juli 2020. Due to the current circumstances this will be an online course.

Noa Lang first gotten on our radar January 22nd 2018, almost two years ago. This is his FBM contribution chart at that time of that match:

Compared to his most recent FBM contribution chart, the only difference is that Noa Lang today has a higher transitioning & build up then two years ago:

Given Noa Lang’s high attacking contribution it comes as no surprise that Noa Lang scored a hattrick in the match FC Twente vs Ajax of December 1st 2019.

Bayesian statistics, like FBM uses, needs way less data than before you can draw valid conclusions. That makes Bayesian statistics ideal for scanning youth players for talents. Sergiño Dest is a great showcase in this regard. According to Ronald Koeman, the current manager of the Dutch national team, one and a half a year ago it was not known what a great talent Dest was. Of course, this is not the case. Quite a number of professionals knew what a talent Dest was, even in the early years.

FBM for instance created this FBM contribution chart for Dest in April 2017, two and a half years ago:

An FBM contribution chart that is quite the same as the chart of his most recent full game:

With the exception of the much higher green line representing the probability that Dest is able to contribute to the transitioning & build up of Ajax. Yet, Dest’s increase in performance in transitioning and build up is very recent. A month earlier his FBM contribution FBM chart still looked like this:

It is very hard for players to do well on transitioning & build up (the green line) in our FBM system. When they do, they often become the talk of the town. So what happened with Dest is not that his talent wasn’t recognized, but that he only very recently started to play at an exceptional level. That is the reason why you use Bayesian statistics, like FBM, to keep track of talented youth players and at the same time combine that information with scouts with a proven track record for being good at predicting the progression youth players make. Then you are not surprised or disappointed when a star player chooses to play for a different national team than for the country where he grew up.

As there is an 80% correlation between FBM Team Score and the ranking in the Eredivisie and a 90% correlation between FBM Team Score and points scored in the Eredivisie, we can predict what the ranking of the clubs and the points they score. As teams that risks to be regulated can buy players that make significantly difference in the team’s performance, we can really only predict the ranking and points on January 1st 2020. In principle the same goes for any player still bought before the close of the summer transfer window, but in practice this will make very little difference to our prediction, if at all.

So without further ado, here is our definite prediction for the Eredivisie season 19/20:

Rank	Club	Average FBM Team Score	Points 1/1/2020	Points end of 19/20 season	Average points previous 5 years	Difference
1	Ajax	239.5	43	85	85	0
2	PSV	197.5	35	70	79	-9
3	AZ	160.5	29	57	67	-10
4	Utrecht	160	29	57	61	-4
5	Feyenoord	156.5	28	55	54	1
6	Vitesse	144	26	51	51	0
7	Willem II	137.5	25	49	46	3
8	Twente	133	24	47	48	-1
9	Heerenveen	130	23	46	44	2
10	VVV	121.5	22	43	45	-2
11	Sparta	112	20	39	40	-1
12	ADO	111.5	20	39	38	1
13	Emmen	111	20	39	41	2
14	Groningen	103.5	18	36	36	0
15	Zwolle	101	18	35	31	4
16	Fortuna	96	17	34	33	1
17	Heracles	89	16	31	28	3
18	RKC	84.45	15	30	22	8

Explanation

This is what the columns mean in the above table:

1. Average FBM Team Score. FBM Team Score is calculated by adding all the values of the players in the starting XI to the team score at the end of each match. This includes all the data we have of the team from the previous two seasons and the start of the new season. Older matches have less weight and recent matches weigh more.
2. Points 1/1/2020. This is the predicted number of points that team will probably have January 1st 2020.
3. Points end of 19/20 season. This value is only provided so we can check whether the values are in line with the average of the last five years for each position. Due to changes in the team during the winter transfer period, these values are only a prediction if no team changes anything during the winter transfer period which is of course extremely unlikely.
4. Average points in the previous 5 seasons. This is the average points for the rank not for the club. For example, the champion in the Eredivisie had 85 points , on average over the previous 5 seasons.
5. Difference. This is the difference between the 5 year average and our projected points for the end of the season. This shows how likely it is that we overestimate or underestimate the club. So we are likely to underestimate the number two of the Eredivisie and if that it is PSV, which is also likely, then we are underestimating PSV. And we are probably overestimating number 18, which in all likelihood won’t be Heracles, but one of the other teams just above Heracles.

Notes

1. In our projection the teams score 6 less points than the 5 year average which is less than 1% difference of the 849 total points scored in the 5 year average.
2. Although the model predicts the top two in the same way as most people would do, there is an anomaly in the sense that the number two of the Eredivisie would score significantly less than on average. This could be the case as the favorite teams in the Eredivisie had a rocky start of the season. Nevertheless, it is more likely that our model underestimates PSV currently.
3. In the same light, our model seems to underestimate the upper half of the table and overestimate the bottom half of the table.
4. The other anomaly would be the regulation of RKC with 30 points. Although it is likely that RKC will be regulated, 30 point is quite a high number of points to still be regulated. So it is likely that our model overestimates RKC.
5. If clubs are predicted to score the same or almost the same number of points then it is obvious that the smallest change might affect whether one club is on top of the other or vice versa. For instance, we predict ADO to be on top of Emmen, but it could very well be the opposite.
6. As this is the first season we make this prediction we have to see how clubs that are promoted from the Eerste Divisie do in this prediction. We use a deflator for historical games in the lower league, but even with the deflator promoted clubs do quite well. So we might overestimate promoted clubs. Nevertheless, Twente and Sparta did have a great start to the season.

Please note: this is a preview of our prediction that we will make once the transfer window is closed September 2nd 2019. Yet, we don’t expect much to change between this preview and the final prediction. I will update this article once the transfer window is closed.

As there is an 80% correlation between FBM Team Score and the ranking in the Eredivisie and a 90% correlation between FBM Team Score and points scored in the Eredivisie, we can predict what the ranking of the clubs and the points they score. As teams that risks to be regulated can buy players that make significantly difference in the team’s performance, we can really only predict the ranking and points on January 1st 2020. In principle the same goes for any player still bought before the close of the summer transfer window, but in practice this will make very little difference to our prediction, if at all.

So without further ado, here is our prediction for the Eredivisie season 19/20:

Rank	Club	Average FBM Team Score	Points 1/1/2020	Points end of 19/20 season	Average points previous 5 years	Difference
1	Ajax	236.5	42	83	85	-2
2	PSV	192.5	34	67	79	-12
3	Utrecht	173.5	31	61	67	-6
4	AZ	162.5	29	57	61	-4
5	Feyenoord	153.5	27	54	54	0
6	Vitesse	142	25	50	51	-1
7	Willem II	137	24	48	46	2
8	Heerenveen	134	24	47	48	-1
9	VVV	130	23	45	44	0
10	Twente	115	20	40	45	2
11	Emmen	114.5	20	40	40	-1
12	Sparta	111	20	39	38	-1
13	RKC	109	19	38	41	-6
14	Groningen	102.5	18	36	36	0
15	ADO	97	17	34	31	3
16	Zwolle	96	17	34	33	1
17	Fortuna	90.5	16	32	28	4
18	Heracles	85.5	15	30	22	8

Explanation

This is what the columns mean in the above table:

1. Average FBM Team Score. FBM Team Score is calculated by adding all the values of the players in the starting XI to the team score at the end of each match. This includes all the data we have of the team from the previous two seasons and the start of the new season. Older matches have less weight and recent matches weigh more.
2. Points 1/1/2020. This is the predicted number of points that team will probably have January 1st 2020.
3. Points end of 19/20 season. This value is only provided so we can check whether the values are in line with the average of the last five years for each position. Due to changes in the team during the winter transfer period, these values are only a prediction if no team changes anything during the winter transfer period which is of course extremely unlikely.
4. Average points in the previous 5 seasons. This is the average points for the rank not for the club. For example, the champion in the Eredivisie had 85 points , on average over the previous 5 seasons.
5. Difference. This is the difference between the 5 year average and our projected points for the end of the season. This shows how likely it is that we overestimate or underestimate the club. So we are likely to underestimate the number two of the Eredivisie and if that it is PSV, which is also likely, then we are underestimating PSV. And we are probably overestimating number 18, which in all likelihood won’t be Heracles, but one of the other teams just above Heracles.

Notes

1. In our projection the teams score 6 more points than the 5 year average which is less than 1% difference of the 849 total points scored in the 5 year average.
2. Although the model predicts the top two in the same way as most people would do, there is an anomaly in the sense that the number two of the Eredivisie would score significantly less than on average. This could be the case as the favorite teams in the Eredivisie had a rocky start of the season. Nevertheless, it is more likely that our model underestimates PSV currently.
3. In the same light, our model seems to underestimate the upper half of the table and overestimate the bottom half of the table.
4. The other anomaly would be the regulation of Heracles with 30 points. This is quite a high number of points to still be regulated. Even though there is an 80% correlation between FBM Team Score and league ranking, our model put Heracles at place 12 whereas in reality they ended the league in place 6. So it could be that our model systematically underestimates Heracles. On the other hand with 30 points Heracles still gets a lot of points. It is just that other teams get more points. So it could very well be that the race against regulation could be a very tight race this year with a lot of clubs remaining under threat of regulation for a very long time. If this prediction is an indication for the coming season, it will be a very busy winter transfer window with clubs rushing to buy players to prevent regulation.
5. If clubs are predicted to score the same or almost the same number of points then it is obvious that the smallest change might affect whether one club is on top of the other or vice versa. For instance, we predict Twente to be on top of Emmen, but it could very well be the opposite.
6. As this is the first season we make this prediction we have to see how clubs that are promoted from the Eerste Divisie do in this prediction. We use a deflator for historical games in the lower league, but even with the deflator promoted clubs do quite well. So we might overestimate promoted clubs. Nevertheless, these two of the three promoted clubs did have a great start to the season.

Everyone has an opinion on the quality of the transfers of their favorite club in the Premier League. But can we actually measure successful transfers? Here is the table of successful and unsuccessful transfers in the Premier League. The explanation follows below:

Club	Successful transfers	Unsuccessful transfers	Losses per player	Losses per year	Profit per youth player
Tottenham Hotspur	88%	12%	5.34	1.068	5.87
Watford	84%	16%	3.14	0.628	0
Everton	84%	16%	3.44	0.688	11.48
Leicester	84%	16%	5.37	1.074	2.69
Chelsea	83%	17%	4.74	0.948	4.28
Newcastle	80%	20%	2.8	0.56	0.28
Bournemouth	80%	20%	2.58	0.516	0
Brighton	80%	20%	1.8	0.36	0
Arsenal	78%	22%	7.85	1.57	4.65
Manchester City	77%	23%	8.28	1.656	5.98
Sheffield	76%	24%	0.65	0.13	0.97
Liverpool	75%	25%	10.45	2.09	10.44
Burnley	74%	26%	1.53	0.306	0
West Ham	73%	27%	1.83	0.366	3.46
Huddersfield	71%	29%	1.17	0.234	1.65
Average	70%	30%	3.93	0.79	3.49
Cardiff	70%	30%	1.1	0.22	0
Norwich	60%	40%	3.51	0.702	5.39
Southampton	59%	41%	6.53	1.306	6.26
Wolverhamperton	55%	45%	1.4	0.28	0.51
Crystal Palace	52%	48%	3.23	0.646	8.25
Manchester United	52%	48%	9.12	1.824	2.91
Aston Villa	43%	57%	2.51	0.502	1.9
Fulham	34%	66%	2.17	0.434	3.53

The first column indicates the percentage of successful transfers. Here we mainly mean financial success. We have looked at over 800 players who have left their club in the past five years. The basic idea is that if a club received less money than what they paid for the player then it would be an unsuccessful transfer. The idea being that if he was a success at the club, he would have been worth more.

Of course, there are many exceptions. Especially if the player is playing quite some time for the club. For that reason we used a depreciation formula to decrease the amount paid for the player for each year that he actually played at the club. If a player played 5 years for the club, the transfer would be an automatic success this way/ Loan fees were also taken into account.

The second column indicates the percentage of unsuccessful transfers.

The third column indicates what, on average, an unsuccessful transfer cost the club in the past five years in millions of pounds. The fourth column is this amount divided by 5. This number is basically the amount that the club can spend each year to prevent 1 unsuccessful transfer every 5 years. Adding FBM statistics to your data analysis costs a fraction of this amount. Adding FBM statistics immediately reduces the risk of an unsuccessful transfer because we do our own data acquisition and the FBM approach is completely different than any other data provider. For that reason we are 100% complementary to existing data analysis. With FBM statistics you have another data source to confirm or disconfirm that a player will be a success at the club.

The fifth column is how much money a club has made with the transfers of their own youth players. This is an indication on whether the club ought to concentrate on scouting or youth development or both. As FBM uses Bayesian statistics FBM needs way less data before we can draw well founded conclusions. So FBM statistics is ideal for the youth development program of the club as well as youth scouting.

Some caveats in determining successful transfers

First let me stress that these numbers are an indication. One can always use a slightly different formula to divide transfers between successes and failures. Nevertheless, other approaches will basically show the same picture. 

The second point is that only players that have actually left the club are counted. If you think that there are still a lot of bad players on the roster of the club, then it is likely that the club will have worse numbers the year these players leave the club. The opposite is also true. If a club has just shed it’s dead wood, then they will probably do better in the future. But for now, this is how it looks.

Thirdly, and this connects with the second point, the numbers are relatively small. On average we considered about 30 players per club, with only a few of them unsuccessful transfers. That means that if next year a player with an unsuccessful transfer leaves the club, that it will have some impact on these percentages.

Two examples of how FBM Second Opinion would have prevented unsuccessful transfers

Paul Gladon

Paul Gladon transferred for 1.8M pound from Heracles to Wolverhampteron. Here is how FBM statistics view Paul Gladon in his last match for Heracles (see here for an explanation of how to read an FBM contribution chart).

We think that it is quite likely that Wolverhampteron would have saved 1.8M euro if they had seen this chart and all our other data on Gladon.

Davy Klaassen

Although Klaassen’s FBM contribution chart is a lot better than Gladon’s chart, it still doesn’t justify the transfer from Ajax to Everton, especially if Everton wanted to use Klaassen to support their attack, rather than their defense:

Very telling is that even though the Netherlands won 5-0 Klaassen’s attack contribution hardly rises. This chart would and all our other data of Klaassen would probably have prevented Everton from misspending 24.3M pound.

A huge club like Feyenoord ought to always play for the championship. Nevertheless, the distance between Feyenoord and PSV (currently #2) and Ajax (currently #1) proofs to be to big. In this article I react to the wonderful article “Wat brengt de nabije toekomst voor Feyenoord?” by Enrico Raho. Not to criticize, but to show the differences between a classical approach to football statistics as has been excellently written down by Enrico Raho and our FBM Bayesian approach.

Continue reading “Feyenoord needs to defend their #3 spot rather than go for #2 or #1” →