2014 FIFA World Cup Brazil Thread

[azzaman333]

Spoiler :

Okay, I couldn't find data for international matches in the time I was willing to spend on this, but I used the last 5 seasons' Premier League results. So 5*38 matches -- which in fairness is a lot more matches than I'd get for international matches. Also, the teams all play each other, so the model doesn't need to do a whole lot of inferring. The only inferences it needs to make are for teams that weren't in the EPL for the entire period. So it's quite a full set of data. International matches would surely suffer from lack of data, but I think this is decent enough as a Proof of Concept.

Anyway, here are the ratings from a simple regression:

Spoiler :

Team Rating
Man United 100
Man City 100
Chelsea 98.67847251
Arsenal 91.24987148
Liverpool 86.74672761
Tottenham 79.9527655
Everton 77.00189131
Southampton 67.34082524
Swansea 65.57085085
Newcastle 61.48417829
Stoke 60.30450676
West Brom 59.96917222
Sunderland 59.77150928
Fulham 58.71666108
Aston Villa 58.59152678
Crystal Palace 56.37806953
West Ham 55.47980011
Birmingham 55.20307239
Blackburn 53.13381893
Norwich 52.59403121
Bolton 51.5010988
Blackpool 49.78606726
QPR 47.94238609
Wigan 46.36281808
Hull 46.15081066
Reading 45.83081317
Wolves 44.60609349
Portsmouth 41.1694753
Cardiff 37.03001916
Burnley 36.14653848

EDIT: I should say, I set the ratings to be capped at 100. I might change that, cos there's a lot of bunching at the top apparently.

EDIT2: here we go (added a new coefficient, rather than uncapping it):

Team Rating
Man City 100
Man United 99.66432344
Chelsea 98.04183822
Arsenal 88.14217575
Liverpool 82.28388009
Tottenham 73.05871513
Everton 69.14361476
Southampton 56.1279588
Swansea 54.04388897
Newcastle 48.6834998
Stoke 47.55741922
Sunderland 46.44663783
West Brom 46.38355903
Fulham 45.13295989
Aston Villa 44.66287277
Crystal Palace 41.68245462
West Ham 40.91714886
Birmingham 40.54873051
Blackburn 38.27243406
Norwich 37.30300328
Bolton 35.42319069
Blackpool 32.74057874
QPR 30.28019169
Hull 28.56939532
Wigan 28.28378844
Reading 27.70709578
Wolves 25.77749475
Portsmouth 21.73979115
Burnley 16.85879097
Cardiff 15.94182248

It's still worse than an Elo model though. The table below is where my Elo model for the EPL had the clubs at the start of March 2014.

Team Rating
Chelsea 1387
Man City 1377
Arsenal 1310
Liverpool 1302
Tottenham 1222
Man United 1222
Everton 1189
Southampton 1068
West Ham 1046
Newcastle 1016
Hull 961
Swansea 960
Norwich 956
Stoke 950
Sunderland 938
West Brom 920
Crystal Palace 911
Aston Villa 910
Fulham 822
Cardiff 790

Clearly, the Elo model is a lot better at showing the relative strengths of the teams at any given time over a regression model.

 

[Mise]

Better in that it reflects the current league table you mean?

P.S. what do you mean by "at any given time"? Do you mean that in March 2014, Chelsea were the strongest side, or do you mean that, given all results up to March 2014, Chelsea were the strongest overall?

 

[azzaman333]

Ranking changed, not rating. England is in a very tight grouping, where you can find up to ten or a dozen teams separated by twenty or thirty points. They dropped 80 points, so they dropped plenty of ranks. That doesn't change the fact that they still rank hundreds of points higher than teams that got similarly bounced from the group phase, and it will take a decade of getting similarly bounced for them to be rated in the same range as those other teams that get similar results. The difference being a product of wins the current team had nothing to do with.

My intent wasn't to pick on England by the way, they just provide a good example because of their extensive winning history.

I don't see how England being rated about the same as Italy, Switzerland, Ecuador or USA is an unfair reflection of their ability. You don't want the rankings to change too quickly otherwise you get one upset result and everything goes haywire. If they continue to perform poorly, their rating will continue to decline.

 

[Akka]

That would be indeed a valid point if minnows weren't rated as minnows according to the Elo system, but they actually are.

Okay, last time : the entire ELO system is about adjusting the rating over time, but it can only do so if the teams play against each other rather frequently. It's not magic. My point is precisely that the configuration by regions means there isn't enough inter-regions playing for this adjustement to work adequately.

 

[azzaman333]

Better in that it reflects the current league table you mean?

P.S. what do you mean by "at any given time"? Do you mean that in March 2014, Chelsea were the strongest side, or do you mean that, given all results up to March 2014, Chelsea were the strongest overall?

It's a better reflection of team performance in that season, and therefore it's better in its predictions for the remainder of the season. The Elo ratings 3 months out from the end of season are a better predictor for the 2014 end of season results than the regression model up to and including the end of 2014.

By any given time, I mean if I gave you my Elo ratings and your regression model and told you to predict who would win the next round of matches, the Elo rankings would be better.

Okay, last time : the entire ELO system is about adjusting the rating over time, but it can only do so if the teams play against each other rather frequently. It's not magic. My point is precisely that the configuration by regions means there isn't enough inter-regions playing for this adjustement to work adequately.

The Elo (it's not an acronym) system does not require teams to play inter-region matches on a frequent basis. The model is capable of inferring the relative strength of the regions from limited inter-regional play.

 

[Kyriakos]

^Cardiff?

Doesn't Wales still have its own national team? (and doesn't one need a national league for that? ).

 

[Mise]

It's a better reflection of team performance in that season, and therefore it's better in its predictions for the remainder of the season. The Elo ratings 3 months out from the end of season are a better predictor for the 2014 end of season results than the regression model up to and including the end of 2014.

By any given time, I mean if I gave you my Elo ratings and your regression model and told you to predict who would win the next round of matches, the Elo rankings would be better.

Well now you're moving the goalposts... If I do the model just on the 2013/14 season up to the end of Feb, I get this:

Team Rating
Man City 93.96333433
Liverpool 82.75719434
Chelsea 75.3550508
Arsenal 67.56367789
Everton 55.98147084
Man United 51.76895396
Southampton 45.85335431
Tottenham 41.51826666
Hull 38.40090183
Swansea 36.9682555
Newcastle 33.24527476
West Ham 32.68571435
West Brom 30.74512725
Stoke 27.57849063
Aston Villa 25.75336413
Norwich 19.14088412
Crystal Palace 17.58421748
Sunderland 17.17873013
Cardiff 6.891073113
Fulham 4.779796631

I don't see any real difference between the Elo ranking and this regression. Other than predicting the top 5 perfectly anyway (which, admittedly, is easier than the bottom as there is much more movement at the bottom than at the top). If you took the two rankings, can you really say that Elo would win?

Secondly, if you do the Elo rating now, Man Utd will look like a mid-table team. In other words, recent results have indeed caused them to plummet down the table, even though it's fairly likely that last season was anomalous and they'll bounce back this season.

Finally, there are multiple ways you can cut this. Elo has an appeal because you don't throw out any data consciously; there's no parameter in there to say that this data is relevant and this data isn't. The time weighting, rather, is built into the system: it's a weighted moving average. That's appealing, personally, because it means I don't have any worries about spurious parameters or overspecification. But there are advantages to being able to pick and choose time periods or tweak parameters to the model. For example, by simply weighting each of the past 5 seasons by 5, 4, 3, 2 and 1 (5 for the most recent season), you get a better picture for what next season might look like:

Spoiler :

Team Rating
Man City 98.26118948
Chelsea 86.76467998
Man United 86.69971958
Arsenal 79.42983375
Liverpool 78.68460796
Everton 64.58560067
Tottenham 63.56115317
Southampton 50.4101791
Swansea 47.24765746
Stoke 41.7699479
West Brom 39.521656
Newcastle 38.83586863
Sunderland 38.22767979
West Ham 37.77249916
Crystal Palace 34.9684692
Aston Villa 33.12703986
Hull 31.20240093
QPR 23.49463615
Burnley 15.31285298

(no data for Leicester, sadly)

How does that compare with Elo rankings? Where do you think Man Utd will finish, for example? Where does Elo put it?

In general, I just don't see why everybody seems committed to designing an elaborate scoring system instead of doing what you would do any other time you have some observations and want to make predictions.

 

[Marla_Singer]

How does that compare with Elo rankings? Where do you think Man Utd will finish, for example? Where does Elo put it?

That would depend on the K factor you would attribute (the one defined differently for friendlies, qualifying and world cup games). The higher will be the K, the more weight will be given to recent results.

Here is the Elo formula:


Rn is the new rating, Ro, the old rating, K is the weight attributed to the game, G an index based on something similar to the goal difference log, W is the actual outcome of the game (0 for a loss, 0.5 for a draw, 1 for a win). and We is its expected outcome.

The expected outcome is based on the pre-game rating differential between both teams (from 0 to 0.5 for the lower rated team ; from 0.5 to 1 for the higher rated team).

Here's the formula applied to calculate win expectancy, dr being the aforementionned pre-game rating differential:

In general, I just don't see why everybody seems committed to designing an elaborate scoring system instead of doing what you would do any other time you have some observations and want to make predictions.

Well, during my old classes in economics when I was a student, I learnt that we just can't predict the evolution of a stock rating simply by looking at its past curve. The same would apply here in the meaning that the actual outcome of future games depend on external factors which don't appear in the outcome of past games.

The only thing you could focus on is how myopic you are regarding recent results: if you consider short-term events are decisive, then you should give them more weight, if on the other side you consider what matters the most is the long trend, then you should give them less weight.

 

[Dell19]

^Cardiff?

Doesn't Wales still have its own national team? (and doesn't one need a national league for that? ).

There is a Welsh league. It was set up relatively recently and 5 or 6 of the Welsh teams chose to continue playing in the English leagues whilst the rest switched which makes sense as Cardiff, Swansea and Wrexham would have taken a step backwards if they had moved.

 

[azzaman333]

Well now you're moving the goalposts...

I didn't have data up til the end of 2014, so I stopped where my data ended.

If I do the model just on the 2013/14 season up to the end of Feb, I get this:

Spoiler :

Team Rating
Man City 93.96333433
Liverpool 82.75719434
Chelsea 75.3550508
Arsenal 67.56367789
Everton 55.98147084
Man United 51.76895396
Southampton 45.85335431
Tottenham 41.51826666
Hull 38.40090183
Swansea 36.9682555
Newcastle 33.24527476
West Ham 32.68571435
West Brom 30.74512725
Stoke 27.57849063
Aston Villa 25.75336413
Norwich 19.14088412
Crystal Palace 17.58421748
Sunderland 17.17873013
Cardiff 6.891073113
Fulham 4.779796631

I don't see any real difference between the Elo ranking and this regression. Other than predicting the top 5 perfectly anyway (which, admittedly, is easier than the bottom as there is much more movement at the bottom than at the top). If you took the two rankings, can you really say that Elo would win?

I think for long term predictions, the regression method is better. For match day predictions I'd still back Elo because I've seen that the predicted probabilities of winning according to Elo do actually reflect the long term probabilities of those matches unfolding that way.

Secondly, if you do the Elo rating now, Man Utd will look like a mid-table team. In other words, recent results have indeed caused them to plummet down the table, even though it's fairly likely that last season was anomalous and they'll bounce back this season.

That depends how you define mid-table. Is it Everton and Tottenham? Or is it Southampton, Stoke, Newcastle, etc? Because even their current rating of 1203 (a drop of almost 200 from this time last season) is higher than Southampton, Stoke or Newcastle have ever been rated (since the start of the 02-03 season). It's also more or less the same rating that Arsenal had at the end of the past 2 season. But as I said above, the regression does seem to be a better predictor of long term outcomes. I'd like to see how good it is at predicting individual matches though.

Finally, there are multiple ways you can cut this. Elo has an appeal because you don't throw out any data consciously; there's no parameter in there to say that this data is relevant and this data isn't. The time weighting, rather, is built into the system: it's a weighted moving average. That's appealing, personally, because it means I don't have any worries about spurious parameters or overspecification. But there are advantages to being able to pick and choose time periods or tweak parameters to the model. For example, by simply weighting each of the past 5 seasons by 5, 4, 3, 2 and 1 (5 for the most recent season), you get a better picture for what next season might look like:

Spoiler :

Team Rating
Man City 98.26118948
Chelsea 86.76467998
Man United 86.69971958
Arsenal 79.42983375
Liverpool 78.68460796
Everton 64.58560067
Tottenham 63.56115317
Southampton 50.4101791
Swansea 47.24765746
Stoke 41.7699479
West Brom 39.521656
Newcastle 38.83586863
Sunderland 38.22767979
West Ham 37.77249916
Crystal Palace 34.9684692
Aston Villa 33.12703986
Hull 31.20240093
QPR 23.49463615
Burnley 15.31285298

(no data for Leicester, sadly)

How does that compare with Elo rankings? Where do you think Man Utd will finish, for example? Where does Elo put it?

In general, I just don't see why everybody seems committed to designing an elaborate scoring system instead of doing what you would do any other time you have some observations and want to make predictions.

Elo rankings from the end of last season have Man Utd 6th.


I touched on it above, but a massive advantage to Elo is the easily calculated probabilities for matches, which generally fit the true distribution of results well. It's also very easy to customise the predictions, such as adding factors for home ground advantage, historical bogey teams, midweek matches, importance of the match, etc.

 

[Mise]

How accurate is it at individual matches? E.g. what percentage of matches would it correctly predict the result for? (Win/Draw/Loss I mean.)

 

[azzaman333]

How accurate is it at individual matches? E.g. what percentage of matches would it correctly predict the result for? (Win/Draw/Loss I mean.)

About 53% is what I'm sitting at, 71% if drawn matches are excluded.

 

[Mise]

Yeah that's pretty good. Mine comes out at 50% on historic results (hardly changes if draws are excluded), but you would expect that to be quite high as it's specifically optimised for predicting past results... I doubt it would do as well at predicting future results. The R-squared on GD is 23% for 5 seasons and 33% for 1 season.

 

[Timsup2nothin]

The Elo (it's not an acronym) system does not require teams to play inter-region matches on a frequent basis. The model is capable of inferring the relative strength of the regions from limited inter-regional play.

Of course. When there is inter-regional play points migrate towards the stronger regions, and then circulate within those regions. The problem isn't the limited inter-regional play, the problem is the limited rate of data accumulation overall relative to the rate of change in the teams.

Since I've probably offended the English enough let's use Brazil as an example. Their rating is accumulated from something over 900 matches played. They did not have a single player on their roster who participated in ten percent of them. They only had five 'grizzled veterans' who participated in five percent of them, three of them just barely.

So at least 95% of their rating is based purely on 'tradition' not the current strength of the squad. This led to seeding them and expecting them to win, based on a rating system that gets more input from Pele than it gets from Hulk. The Elo system is terrific for measuring the grand tradition over time. For measuring the current relative strength of squads, not so much.

 

[Zelig]

Of course. When there is inter-regional play points migrate towards the stronger regions, and then circulate within those regions. The problem isn't the limited inter-regional play, the problem is the limited rate of data accumulation overall relative to the rate of change in the teams.

Since I've probably offended the English enough let's use Brazil as an example. Their rating is accumulated from something over 900 matches played. They did not have a single player on their roster who participated in ten percent of them. They only had five 'grizzled veterans' who participated in five percent of them, three of them just barely.

So at least 95% of their rating is based purely on 'tradition' not the current strength of the squad. This led to seeding them and expecting them to win, based on a rating system that gets more input from Pele than it gets from Hulk. The Elo system is terrific for measuring the grand tradition over time. For measuring the current relative strength of squads, not so much.

I'm really not sure this is the case. I don't really feel like doing it, but if you go and calculate England's Elo rating based only on the past five years, I doubt it differs a whole lot from the full Elo rating.

https://en.wikipedia.org/wiki/England_national_football_team_results_–_2000s

Since 2009, England has lost to: Spain, Ukraine, Brazil, Germany (2x), France, Netherlands, Sweden, Chile, Italy and Uruguay.

Lower rated (as of today) teams they've lost to: Ukraine, Italy, Sweden.
Higher rated (as of today) teams they've beat: Mexico, Spain, Belgium, Brazil

 

[Timsup2nothin]

I'm really not sure this is the case. I don't really feel like doing it, but if you go and calculate England's Elo rating based only on the past five years, I doubt it differs a whole lot from the full Elo rating.

https://en.wikipedia.org/wiki/England_national_football_team_results_–_2000s

Since 2009, England has lost to: Spain, Ukraine, Brazil, Germany (2x), France, Netherlands, Sweden, Chile, Italy and Uruguay.

Lower rated (as of today) teams they've lost to: Ukraine, Italy, Sweden.
Higher rated (as of today) teams they've beat: Mexico, Spain, Belgium, Brazil

And how many current players have been on the team for five years? I'm guessing four. Maybe five. There just aren't enough international matches to feed this sort of system and get a conclusion that is well related to the current team because the current team changes much faster than the data comes in.

As a measure of the grand tradition it's great. I'm all for it.

 

[nocho]

So at least 95% of their rating is based purely on 'tradition' not the current strength of the squad. This led to seeding them and expecting them to win, based on a rating system that gets more input from Pele than it gets from Hulk. The Elo system is terrific for measuring the grand tradition over time. For measuring the current relative strength of squads, not so much.

This is just not true at all. Hungary would be still up those charts if it were remotely like this. Of course there is an effect from past results that takes time to fade away, but that's most certainly not a matter of over 40 years (Pele), more something like about 5 years or so. Also, it's kind of normal that results don't fade immediately, that's really the point of a ranking system, isn't it? Otherwise just look at the world cup results for a ranking. Besides, should Holland, for instance, after their 2nd spot in 2010 have dropped to 60th because of a dismal Euro 2012? No, their result in 2014 more or less shows that it was somewhat a fluke negative result. And if 2014 would have seen Holland going out in the group stages, they'd surely be out of the top 20 now. Similarly, I think it's about right that Spain still is in the lower part of the top 10. They just lost "only" 2 games after all (against fellow top 10 teams)... Their real demise (if it exists) should be confirmed in the next few years to warrant a much lower rating. But it surely won't take decades for its rating to plummet if Spain were to lose everything in the next few years. So no, "95% of their rating being tradition" is nonsense, I'm sorry to say.

That being said, it is true that the relatively few measurements make the rating less trustworthy, but that similarly applies to any ranking system you might conceive based on some statistical model.

 

[Quackers]

In 2018 UEFA is introducing a new qualifying format. All 54 nations will be involved. In the new system each group will be determined by quality. So Group A will have the best teams on the continent.

According to the proposed format,[1][7][8] the 54 UEFA national teams will be divided into four divisions. In the top division, teams will compete to become the UEFA Nations League champions. Teams will also compete for promotion and relegation to a higher or lower division. The UEFA Nations League will also be linked with UEFA Euro qualifying, providing teams another chance to qualify for the UEFA Euro finals tournament.

I'm really looking forward to this and we will get regular updates about the true natural standing of the top European nations. I'm sure England will be below Germany, Spain and Italy but after that I can see us battling it out with France, Portugal and the Dutch for 4th place

 

[kochman]

Two things.
One, where did you see this?
Two, you're crazy if you think England is better than the Netherlands (or France) or that Italy is better than the Nethers...

In fact, Italy isn't that good, and will be without Pirlo and Buffoon, while still not having a decent striker (hint, Balotelli isn't).

 

[Timsup2nothin]

So no, "95% of their rating being tradition" is nonsense, I'm sorry to say.

That being said, it is true that the relatively few measurements make the rating less trustworthy, but that similarly applies to any ranking system you might conceive based on some statistical model.

Actually, it's mathematics. Most people who try to ignore mathematics do end up sorry, so you are ahead of the game. Almost 95% of the matches that Brazil's rating is based on did not involve any current players. Pele played in 92 matches, Hulk has played in 38, therefor Pele does in fact have more influence on their rating than Hulk does, though Hulk clearly will have more impact on their next game.

That being said, we agree in the main. I have been saying all along that any effort to model international match results is pretty much pointless due to lack of timely data. That's why I don't bet on qualifiers, or even group games. By the round of sixteen there is real data that applies to the actual current teams.