Month: April 2016

You know the part in The Big Short where Michael Burry (Christian Bale) is sitting there at his desk, explaining to an irate investor that the housing market is guaranteed to crash, it’s just that no one knows it yet? And the fact that it has never crashed in modern times has nothing to do with his certainty that this future crash is now inevitable?

I feel that way about stats and data in football.

I believe with utter certainty that stats and data will play a huge role in the future of the sport. This is despite knowing that my certitude makes me sound like a bit of a loon to parties that approach this subject with some skepticism.

Today I’m going to explain why I carry this certainty about the future, but it requires the audience to shed one big misconception most people seem to carry about the game.

Football is not unique and unrelated to other sports.

Football actually bears reasonable similarities to basketball and both forms of hockey, such that certain ways of analyzing those sports are easily adaptable to football. And this goes beyond stats – German coaches have long consulted with elite field hockey coaches about defensive tactics, and Pep Guardiola includes legendary water polo player Manuel Estiarte in his coaching staff. Even now, researchers like Luke Bornn are taking what they learned applying spatial statistics to SportVU data in the NBA and seeing what they can learn from football's tracking data.

Yes, football has its own idiosyncrasies and you need to understand the game at a high level to get the most out of your analysis. No one intelligent disputes that. But the fact that we've seen massive revolutions in how other sports are analysed that have lead to changes in how the sports are also played means we should expect a revolution to hit football in the future as well.

Chasing Perfection

I have been reading Andy Glockner’s book recently, which catalogs and explains the NBA’s analytics evolution, and it continually amazes me how much it parallels a movement still in its infancy in football.

“There is way more money involved in [the league] today than even ten years ago, and teams have to work harder and harder to find and maintain competitive edges. How they are doing so varies wildly from team to team, and heavily involves state-of-the-art technology to try to move ever closer to solving an impossibly complex and nuanced sport.”

Is that quote about basketball or football? The NBA or the English Premier League? It could be either, right? Except in the Premier League there are bushels full of competitive edges sitting in easy reach of anyone who knows where to look.

Another thing I believe for certain is they won’t stay that way for long. Spending money now to obtain the low hanging fruit and discover new ones also gives a team a head start in what will assuredly be a brain race at some point down the road, and more importantly, will likely yield huge dividends in terms of points, money, and potential titles now.

That’s the thing that I think the analytics movement in football may have gotten wrong through absolutely no fault of anyone involved. The stats guys developing new ideas and doing the work often think of it as, “how do we apply stats to football to learn new things?”

That’s technically correct. However, it misses the major point.

The real goal for the analytics movement in any sport should be: how do we discover and deliver new competitive edges?

Stats and data are a very useful tool in doing that, but it’s a big tool box.

Another thing Glockner highlights in an early chapter is this piece by Chris Ballard discussing the new stats movement in the NBA, circa 2005. What amazes me is not that most of the names mentioned are still quite prominent in NBA front offices, but instead how bloody young they are.

Sam Presti is 29. Sam Hinkie is 27. Celtics “Senior Vice President for Operations” Daryl Morey is… 31.

Anyway, Glockner’s book is excellent, especially if you read it with an eye that it may be foretelling the future of a football analytics movement that has yet to start across most of Europe.

Statheads Are The Best Free Agent Bargain in Baseball

FiveThirtyEight is a mixed bag, but their sports stuff is still generally pretty good. This piece, which examines the expansion of “numbers-savvy front-office staffers over time” is excellent.

“Although the analytical gold rush began before the period we examined, hiring has accelerated at an almost exponential rate over the last few years.”

One of the main takeaways from the article is that baseball teams are spending more and more on stats dorks because they provide a dramatically bigger boost to win totals on a per dollar basis than many free agent signings. Part of that is because baseball's player market has become more efficient over the years thanks to improved use of stats, but a bigger part comes down to basic economics.

They estimate a five-man analytics team costs about $350,000 per year, which still lags behind the minimum salary for a single player.

The takeaway: It paid to invest in analytics early. Teams with at least one analyst in 2009 outperformed their expected winning percentage by 44 percentage points over the 2012-14 period, relative to teams who didn’t — an enormous effect, equivalent to more than seven extra wins per season. 

Even the minimum estimate of two extra wins per year would represent a return roughly 30 times as efficient as spending the same amount on the free-agent market.

One more thing that really struck me out of that piece and that I feel is hugely applicable to football.

Although the big-budget Boston Red Sox were also one of the first teams to demonstrate that an analytics department could help win a World Series, a number of low-payroll, small-market teams — including not only the Moneyball A’s, but also the Rays, Indians, Padres and Pirates — were among the first to form quantitative departments and develop systems to house and display statistical data. It made sense: The more pressing a team’s financial imperative to stretch every dollar and wring out every win, the more likely it was to try a new approach.

How can teams compete with the traditional giants beyond just spending more money?

• Apply the marginal gains.
• Make consistently better decisions than other teams.
• Play more efficient football.
• Recruit better coaches.
• Recruit better players.
• Make fewer mistakes in the transfer market.

Find. The. Edges!

Baseball and Basketball are hugely different sports. In fact, they are more different from each other than basketball is from football. And yet in both of these areas we have seen teams dramatically ramp up spending to get smarter faster than the competition.

Why? Because it helps them win more.

This WILL happen in football.

The only questions are how long it takes before it happens in scale across not just England, but European football as a whole, and which teams will lead the charge and reap the rewards as early adopters.

--TK
mixedknuts@gmail.com

The soccer Pythagorean probably never had enough life in the first place for it to be considered dead but its torpid state was given a little rattle last week when Soccermetrics put up a post evaluating EPL teams by their Pythagorean points expectations to date.

For those not familiar, the Pythagorean Win Expectation was developed by Bill James while working as a night watchman at a bean cannery. That's not a joke. James himself is more or less the reason Liverpool fans have the term 'Moneyball' to misuse. Anyway, he found that a baseball team's win percentage was equal to:

The exponent has since been adjusted to 1.8 but the formula tracks remarkably well with actual results. Similar formulas have since been developed for basketball and football (to avoid confusion, I'll stick with the American convention where 'football' is 'soccer' and 'football' is 'football').

How is it useful?

If you're a baseball GM at the All-Star Break (roughly midway through the season) and you're wondering wether to start moving players and dumping salaries, or if you should instead go add another bat to make a run at the playoffs, simply calculating your team's Pythagorean win probability is a strong indicator of how much the team can expect to win (or lose) before even making any personnel decisions.

There is room for legitimate debate on its usefulness in soccer with it's 38 game schedule—MLB has 162 games, the NBA has 82—but even in the NFL where there are only 16 games, it is still useful predicting year-on-year performances.

As Football Outsiders noted: "[NFL] teams that win a minimum of one full game more than their Pythagorean projection tend to regress the following year; teams that win a minimum of one full game less than their Pythagoerean projection tend to improve the following year."

So even for smaller game-per-seasons sports where you might not get reliable and meaningful results until the season is over, it can still be a useful tool in evaluating something like, "Should we fire Roberto Martinez?" But soccer's attempts to fit win expectation into a similar formula to have been a little less tidy. There are examples here, here and here, the latter being the Soccermetric version, derived by Howard Hamilton. The formula itself is below.

It's not anything a casual fan would use.

In one of his posts Hamilton goes so far as to claim that a more conventional Pythagorean for soccer simply isn't possible. Even some attempts that are decidedly simpler generally abandon the thing that probably gave James the impetus to name it the 'Pythagorean' in the first place—that every term had the same exponent (2 in his formulation).

That's not just a nice aesthetic characteristic.

There's a bit in that formulation which conforms to basic intuition about sports. Specifically, that if a team scores as many points as it gives up over the course of the season, it will likely end up winning as many games as it loses. For now just think about a binary win-loss game, like basketball or baseball.

There's no tying in baseball. That .500 winning percentage is built into James' formula. If Runs Scored (RS) is equal to Runs Allowed (RA), and we have a uniform exponent, then without even plugging any numbers in you can see that it's x/2x, which is .500. Even among the "simpler" models linked above, best-fitting soccer Pythagoreans end up having three unique exponents. That looks something like this ('goalsaway' should almost certainly be 'goalsagainst'):

That's Martin Eastwood's formula. Let me be clear: I have used this formula in the past and it generally gives totally acceptable results. But think about a weird situation where a team scores, oh, exactly 19 goals and gives up exactly 19 goals. They win half of their games—freakishly having 19 1-0 and 0-1 results—to collect the maximum number of points possible, 57.

By the Eastwood Pythagorean, Team 19-19 projects to pick up 69.0 points. Sure, that's an outlier of a season in a couple of ways but something something Leicester. Point is, that's a non-trivial 12-point miss on what we'd think should be the most easily predictable outcome (on average). That and there is no chance of remembering 1.22777, 1.072388 and 1.27248 if you're working on the back of a napkin.

So we want a Pythagorean that's easy to use, easy to remember and we would like for it to fit a the basic intuition about scoring and results. Oh, and we want it to accurately measure the thing we're trying to measure: which teams over- or underperformed. Really we just want something that looks like the original. Baseball has one, basketball has one, even other football has one. Why can't soccer? It can. Mostly.

There are two tricks we first have to use. The first isn't really a trick. It's more of the obvious response to the fact that there are draws in soccer. And because of that, winning percentages are much lower in soccer than baseball or basketball. So instead, we set up to predict the percentage of points taken from the max available. Again, this is a pretty self-evident and not-at-all-original workaround. The other attempted James-like soccer Pythagoreans I came across also did this.

The second trick is a little less obvious.

We have to pretend time stopped somewhere around 1994. While England was an early adopter of awarding 3 points for a win in the early 80s, it wasn't until its usage in the 1994 World Cup that the rest of the world finally decided to play along. By 1995 it was pretty much the standard everywhere.

Prior to that though, teams received two points for a win and one for a draw. If we reconfigure the league tables to look like something from the 80s we can build a single exponent Pythagorean that quickly satisfies our third condition (fitting intuition).

Now if a team scores as many goals as it gives up, it'll probably pick up half of the total points available. (2*W + 1*D)/2* Number of Matches = Percentage of Points Taken = GF^c/GF^c + GA^c It's the Jamesian Pythagorean, even if we don't yet have an exponent (or even know if we can get one; spoiler: we can). If GF = GA, again we end up with x/2x and our equation predicts we'll pick up half the total points available. Under the old 2-points method, there were 76 max points available. Say you scored 50 and gave up 50.

Maybe something bizarre happens and you drew all 38; or you won 10, lost 10 and drew 18; or won 15, lost 15 and draw 8. As long as you win and lose the number of games (you necessarily have to draw the rest), you'll pick up 38 points. Half the max total. As for the other requirements. Yeah, it's pretty simple. Once you do the math you get an exponent of 1.2 (this is very much in line with others' results). So our Soccer Pythagorean is:

Again, this assumes that wins are still worth just 2 points. How well does it work? Pretty well. Using the 2010-2011 through 2014-2015 tables from La Liga, Serie A and the EPL (I skipped the Bundesliga so that we had uniform 38-game leagues in the sample), we're off by 2.74 points.

That's root mean squared error, which, if you're not familiar is just what it reads as. So, take the predicted percentage of points (the results of our Pythagorean above), multiply that by the total points available (76).

That gives us the predicted total points for a team's season. Now take the difference between the predicted points and the actual points (the ones teams collected in real games). That gives us 300 predictions and measures of how far off we were from those predictions. Here's what that looks like in some raw R output.

The 'Pythagorean_%' column is what we calculate from the above formula. 'Available_Points' are the old school max available points (2 * 38 = 76). Multiply those together to get our 'Predicted_Points'. 'Adjusted_Points' are what we get when we take actual, real world points and convert them to 2-point wins (or Pts - W). Subtract 'Adjusted_Points' from 'Predicted_Points' and you get the 'Difference', or your errors. For RMSE, work backwards; grab all your errors, square them, take the mean, then take the square root of that mean. That's where the 2.74 comes from. Going back to Hamilton's model he says that the RMSE for his soccer Pythagorean ends up somewhere between 4 and 5 (although for the single season he does on that page he gets 3.81; for the one from last week it's 4.5). Eastwood does 10 seasons, all of the EPL, and gets 4.08.

Another way to look at it is that in our 300 seasons, for 90 of them our prediction is within +/- a single point. So we're doing much better, right? Well, even though the formula 'works' with a 76-point season, time actually moves forward pretty relentlessly and current, actual seasons are 114-point seasons. That leaves us with two problems. First, we're going to have to figure out how to map the 76-point season to the 114-point season.

That almost certainly will increase our errors. Not even almost, it does.

Second we're going to lose some of the elegance that makes pretending time stopped so delightful in ways that has nothing to do with aging. Take the first problem first. Say a team accumulated six points in the 2-point world. They could do that by winning three games, by winning two games and drawing two, or by winning one and drawing four. In the 3-point world those translate to nine, eight and seven points respectively.

There is no way to perfectly translate from one to the other. But it's close. Just an ordinary least squares regression gives us a good formula for translating from the past to the present.

This is what our actual 300 team seasons look like compared to their values converted from 76-point seasons to 114-point seasons (our R-squared here is .992 but because this is derived from our sample, it's not immutable and a larger set might give us something slightly different).

Now after we transform our predictions to conform to the passage of time, we get a RMSE of 4.35, which is not terrible.

It's not great but 61 of all our predictions (just over 20%) are still within +/- one point. We're not entirely done, though. There's a way to take a James-like Pythagorean and convert it to a linear formula. You can read the paper here, but it involves a Taylor series and taking derivatives and some math you probably don't want to deal with.

Putting aside all the calculus you've forgotten, there's an easy visual explanation to show why we can approximate something so non-linear with just a straight line. This works better with basketball, where the Pythagorean exponent is 13.9. The plot below is of the win percentage as a ratio of PF and PA.

The best teams in basketball outscore the mean by about 110% and the worst teams underscore it at 90%. So the ratios are actually narrow around 1. If you notice, the curve is pretty close to a straight line over that range. After you ignore all the math in the paper linked above and just scroll to the bottom you find the linear approximation to the

James Pythagorean is:

Again, we have to turn our two tricks. We use percentage of total points taken instead of win percentage. And we pretend that wins are still worth two points. Gamma is our Jamesian exponent (1.2), and R_ave is our average runs, which makes no sense in soccer. They don't score runs, they score goals. Use the average goals scored in a season across all teams.

We still have to update to 3-point totals, but because we're running one linear formula, then dumping the results into another linear formula, we can just perform a little substitution and combine them to get a single linear formula. Once you do that, voila (GD is our goal difference, if that's not obvious):

Pretty cool, and simple and nice and it's based on a win being 3-points. It does have one huge caveat¹. It's kind of meaningless until the season is over. Just look at it. Say you play your first game of the season and win it 1-0. That formula says you'll have about 53 points. You'll actually have three. Hell just from last week to this week in the EPL, the RMSE across all 20 teams dropped from well over 8 (admittedly terrible) to a little over 7 (thanks, Stoke, for not even bothering... twice!). This season is weird, though. Fortunately, just eyeballing the GD, points and remaining schedules, a few of the biggest errors have a decent chance of coming way more in line with their predicted points. Except Everton. The Gods are using Everton for their own amusement.

Overall the fit on completed seasons isn't too bad. Our RMSE is 4.7. Not great. Not terrible, but we're not really doing substantially worse than any of the other options (although our mean absolute error, which doesn't punish you quite so much for the large misses on the ends, is just 3.57). Of our 300 team seasons in the sample, 92 are +/- 1 point; so about 30% of the time our predictions are within a single point of a team's actual points. Also, we are neither over or under-predicting as 152 of our errors, almost exactly half, are above zero. For a 38-game season (less than half the NBA, and not even a quarter of an MLB season) that seems surprisingly decent.

Moreover, it turns out it's almost entirely Spain's fault we're not fitting better (and probably measurably so). Here are our eight worst misses by over-prediction: Barcelona, Barcelona, Real Madrid, Real Madrid, Manchester City, Siena (?), Barcelona, Real Madrid. With a smaller coefficient (13.9 vs. 1.2), the soccer line is approximately linear over a wider range than in our basketball illustration above, but not one wide enough to accommodate scoring 110 goals and giving up just 21 with any precision. Stupid Barca. On the other end, the worst under-predictions are almost entirely one man's fault: Paco Jemez. I love just everything about Jemez and Rayo, except for what his attacking philosophy does to this model.

If you think it's better to lose 9-1 than to stop attacking at 4-0 and minimize the GD damage, you're going to engineer anomalous point totals relative to goal differences. And of the five biggest misses, Rayo is responsible for three of them. Math doesn't like heretics. So the idea that we can't do a simple soccer Pythagorean, I'm going to go ahead and disagree with that one. You can make a decent single-exponent Jamesian Pythagorean for soccer with acceptable results. You do have to cheat and make some bad assumptions about space time, but even after correcting for that, we're in the same neighborhood error-wise as things that are far more complicated.

Moreover, we can reduce our two-step Pythagorean to a single, one-step linear approximation. With nothing more than a team's goal difference, you can do better than a back-of-the-envelope job finding if a team under- or over-performed. That's kinda cool. Admittedly, if you're just typing a single line of code into a stats app, then calling any of the other, more complicated functions is in some sense no more or less arduous than what we've done here.

And if you really want to shave off that extra .5 of RMSE, then knock yourself out. But I've got almost no chance of recalling some of the other formula off the top of my head or being able to derive them in a pinch. But our simplest formula? It's very, very close to: (2/3 * GD) + 52. I can commit that to memory easily. And even if I forget it, as long as I can remember 1.2, I can rebuild it in just a couple of minutes. I'm probably biased (read: definitely), but something I can easily access is slightly more useful, even if it's still not totally elegant. @bertinbertin ¹

Okay, maybe the caveat is a little less huge. Danny Page suggests the following, making me feel incredibly stupid for not thinking of it. The tweet is here. But the meat of it is: "Regarding the huge caveat what if you included games played? (0.677(GD)+52.39)*(GP/38)" I hadn't. And doing so, quickly brings results more in line with, well, all the other results. There is also a spreadsheet with his results here.

I can watch a team for weeks without processing much until I know something more specific to look for. I'm not a guy who can watch 15 minutes and determine everything a team is trying and able to do. That's why I nose around in the stats like a old man reaching for his contacts. When I do find one of interest, that often makes watching games clearer and more enjoyable.  Gladbach of recent vintage never made much sense to me until I found out no one in Europe was better at completing passes in the mess around the opposition goal. Then suddenly their entire style of play was easier to pick out and the beauty of Raffael more fully appreciated. Ingolstadt were just another ugly mid-table side until I saw their opposition passing numbers and now enjoy watching their team-wide pressure, Torino a side I would never watch until I saw stats showing how uniquely they pass in their own half, Rayo just a hipster name on a page until I saw their ridiculous high-press, soft-belly defensive numbers, etc, etc. Often times a single number can illuminate an entire team and make me watch in a new way. With 8 teams remaining in European competition and all in action this week, I found something each was near the top of Europe in to help explain their season and to provide a focus to watch the next two weeks. We start with the Europa League...   Europa League Villarreal-Liverpool   No one forces opponents to play out to the wings more than Villarreal. It's not a one year fluke either, last year they led Europe by an even greater margin. At this point we can say that this is at the core of manager Marcelino’s philosophy on how to play. He wants to force teams to the more controllable, less dangerous parts of the pitch where he gets a free extra defender in the sideline. It’s Schmidt-ball in their own half (Leverkusen funnel the ball to the sidelines when teams are building up and then crank up then swarm in the restricted space).   It’s worked beautifully this season as Villarreal have a 6-point edge on Athletic Bilbao in the race for Champions League and their defense to thank. Only Atletico and Barcelona have allowed fewer than Villarreal’s 31 goals conceded. courtesy of Managing Madrid's blog.   This sets up a fascinating battle as Liverpool (and Klopp in recent years) play to the center more than almost every other team in England (and Dortmund did). Alberto Moreno might have even more space than normal ranging up the wing, can he make something of it without leaving an opening for Villarreal?   Liverpool   There is tons more on Liverpool here but for now I'll point out how stingy they have been since Klopp took over. Only Bayern and 2 Italian teams in Juve and Lazio have been less generous with dangerous completions allowed (and Italian teams in general see fewer across the board). Villarreal will probably not see more than a handful of dangerous chances (they are outside the top 10 in time spent in front of goal in Spain and dead last in shots), so they will need to take advantage via a Bakambu goal or a Liverpool individual mistake. With at least one of Skrtel, Lucas and Toure seemingly sure to start Liverpool fans can surely rest easy.   Sevilla vs Shakhtar   It’s the shot quality derby as 2 of the teams who have led the way in Europe as far as avoiding those derided outside the box pot shots will match up. In both the Champions and Europa League, Sevilla and Shakhtar have been right at the top. Skenderbeu fans, I'm sorry but your team sullied the storied name of the Europa League with their performance this season, read this and try again. In La Liga, Sevilla are #1 again, with a similarly low rate. Many stat-heads should enjoy the decision making on display in this tie, though I'm semi-ambivalent for the most part about long bombs. Variables like pressure, passing options, angle, and ball movement are so quickly changing and influence a possession so much that I find it difficult to think most 5% shots with some non-negligible chance of a rebound/corner are disastrous. Plenty of times your possession is just out of options or the best other one is a difficult pass to someone way on the wing hoping they can play in a cross and ending it with a 6% chance at a goal is optimal.   With nothing else much to add here, I’d like to call attention to the season Metalurg Zaporihzia had in the Ukrainian Premier League. They were outscored 50-7 while nabbing 3 points (all from draws) from their first 16 games. They’ve since forfeited their last 7 due to the club declaring bankruptcy and at this point, it’s going to be tough to climb out of the relegation places. Aston Villa fans, it could be worse.       Champions League Atletico Madrid v Bayern Munich   Bayern stop you by refusing you entry into their danger area, not by limiting danger when the line is broken. The high line compresses play and forces long passes often born out of desperation. Those long desperate balls are mostly harmlessly intercepted by Bayern, but when you actually get through, the chances of a shot are higher against Bayern then anyone else. This has been true for the entire Pep-era. Bayern have a larger “lead” in this stat year after year than any other team in any other stat I’ve compiled. If you get a man with the ball in a dangerous area against a Pep side, he will likely be alone (so without the thought of passing options to improve the attack) and the defense is likely to be stretched to where a shot looks quite enticing. Atletico will likely look to recreate their Barca plan by generating just a few situations where they can get a man with the ball closing in on goal. If they create that situation vs Bayern, they are more likely to get a shot off than vs any other team. And yes, I realize this isn't exactly doing something "well" but it's a hard article to write a title for, ok?   Atletico   Only Juventus nears Atletico when it comes to making opponents struggle with the ball to advance to dangerous areas. It takes on average 114 completions for the opposition to get one in the high-danger zone right around goal. We saw how this worked to perfection against Barcelona in the 2^nd leg: A lot of blue all over the pitch except for the most crucial areas.   It figures Simeone will set his team up that way again against a semi-sputtering Bayern. There is plenty more about Atleti's defensive set-up here.       Real Madrid-Manchester City   Way on the right is not where you usually find teams with Real Madrid's budget but here we are. Only Koln, Everton, and a few Italian teams are softer to play against than a team in the UCL semifinals. Aston Villa and Sunderland are just to the left of them for crying out loud. Such a glaring flaw hasn’t gone unnoticed of course, and in some pockets of the stats-heavy internet it’s probably led to Real Madrid being underrated. I’m even feeding into that here, the temptation to point out bizarre stats like this generally outweighs another boring stat about Ronaldo averaging 8 shots or something insane. Madrid play sloppily, provide minimal resistance to the opposition and basically perform like a quickly-thrown together all-star team. These are often taken to mean they can’t compete with the best, but clearly they can. Tons of ugly flaws can be covered up by the Monster of Shot Volume, which Real Madrid can control to an extent unmatched across Europe.   Man City   It’s become somewhat fashionable to highlight Leicester’s direct long ball style as an adjustment that has fueled their rise to the top of the table but I’m of the opinion that their mix is something special like grandma’s meatloaf. You’re not exactly sure how such a generally disagreeable mix of ingredients turned out to taste amazing, but you aren't going to waste your time trying to replicate it to almost assuredly disappointing results. If you want to please your guests, don’t serve them meatloaf, buy them Popeye’s Chicken, and if you want to be a great attacking team, don’t try to copy Leicester's long balls, get enough people forward so you can create intricate interplays and overloads. Man City, Barcelona, and Arsenal will likely have great offenses again next season as they have the bare-bones right.   If City can get ahold of the ball, there don't seem to be many obstacles in their way to racking up a large amounts of chances in this tie. Real Madrid allow the shortest average pass distance in the danger zone in Spain along with the 2nd-highest completion% allowed in the league. City are up near Arsenal miles ahead of #3 in England when it comes to completion% for in the danger zone. Put all that together with the Monsters of Shot Volume on the other side and this tie has the potential to be one of the best to watch for action-loving neutrals ever. If you are trying to win over a basketball fan to soccer, make sure you tune in with them to City-Real Madrid.   Enjoy the games!

What is a radar?

It’s a way of visualizing a large number of stats at one time. In our case, the radars specifically deal with player stats. Some people also call them spider charts or graphs because they can look like they make a spider web.

Why bother creating them? What’s wrong with tables? Or bar charts?

Hrm, let’s deal with the last questions first. There is nothing is wrong with tables of numbers. My brain loves them, and so do many others.

However, you have to admit that tables of numbers are a little boring. Bar charts are better, but they kind of fall apart when trying to compare many attributes at the same time. Radars allow exactly that.

Why bother creating them? That one is complicated. Why bother making infographics or doing data visualization at all? The answer is probably at least a book long, but the quick response is because people like to look at stats presented in this way far more than they like to look at a set of numbers. Radars invite you to engage with them. They create shapes that brains want to process. People have real reactions, and once you get used to what they display and how they display it, you can interpret them much faster than if you had to do the exact same analysis with a table of numbers.

Many of the shapes created correspond to “types” of players, at least when it comes to statistical output. Pacey, dribbling winger. Deeplying playmaker. Shot monster center forward. Starfish of futility.

There’s a lot more methodology chat in the various articles I have written about on StatsBomb, but I need to explain one very quick thing before I move on to player type shapes and examples.

Radar boundaries represent the top 5% and bottom 5% of all statistical production by players in that position across 5 leagues (EPL, Bundesliga, La Liga, Serie A, and Ligue 1) and 5 seasons of data.  In stat-y terms, the cut-offs are at two standard deviations of statistical production.

In non-stat-y terms, Lionel Messi made EVERYONE look terrible. I know, that doesn’t sound that bad because it’s true, but trust me, the newer way the templates are constructed is better.

The design for these was taken from Ramimo's 2013 NBA All-Star poster. I thought it would be really interesting to apply this to football, and then through testing, became irritated by what Messi made everyone else look like if I just used pure stats output. That's when I added the standard deviations idea, and started playing with different positional templates.

QUICK NOTES:

• The only thing these represent is statistical output.
• If you put players in different systems, it may change their output.
• If you put them in different positions, it almost certainly WILL change their output.
• Age will also change statistical output.
• In short, these are a tool to help evaluate players. Like any tool, they have strengths and weaknesses. In general, I have found it much easier to evaluate players WITH this information than without it.

Explaining Bits and Bobs

This means that all the non-percentage stats in this are normalized for 90 minutes played. The reason you do this is to correct for the fact that some players don’t always play 90 minutes. Players that frequently get subbed on or off will inherently look worse if you look at per game stats than per 90 minutes played.

This is the age the player would be at the end of the season. We will change this soon to season age + birthday.

Why use non-penalty goals? Because penalties are converted at a 75-78% rate almost regardless of who takes them. They are a different skill to scoring goals that are not penalties (some teams have even had goalkeepers as their lead penalty takers), and so we strip them out of the scoring numbers.

DRAWING penalties is a great skill (and will be added to assist stats over time). Converting penalties is a very common one.

Shooting%

How many shots were on target out of ALL shots that a player has taken. This includes those that were blocked.

Key Passes

Passes that set up a teammate to take a shot. These are highly correlated with assists, which are passes to teammates who score a goal quickly after. (Note: This is the same stat as Chances Created. Somewhere along the way Opta made Key Passes only mean passes that lead to shots that are NOT goals and CC is all. Which is weird.)

Through Balls

Opta definition: a pass splitting the defence for a team-mate to run on to. Why do we care? These types of passes are generally considered the single type of passes most likely to score a goal.

Scoring Contribution

Combined non-penalty goals and assists per 90 minutes.

PAdj stands for “possession adjusted” stats. The reason why we do this is because it normalizes defensive stats for opportunity. Think about it this way: If your teammates always have the ball, then you can’t make any defensive actions, and you would look worse in this statistic compared to a Tony Pulis-style team that sits deep and constantly defends.

When adjusted for possession, tackles and interception output becomes moderately correlated with shots conceded and goals against, as opposed to having no correlation without the adjustment. In short, it’s an imperfect adjustment, but much better than not having the adjustment at all.

In the bottom left of every radar is the actual statistical output in numbers for each spoke of the radar. Numbers in green are in the Top 5% of output in that stat for the player population and numbers in red are the Bottom 5%.

Forwards + Attacking Midfielder Shapes

Pure Goalscorer

Elite Creative passer

Wide, Dribbling Playmaker

All Around Super Forward

Starfish of Futility

Bowtie of Sadness

Central and Defensive Midfielders

Pure DM

Heavy Attacking CM

Deep-lying Playmaker

General All Around CM

Fullbacks

Defensive

Attacking

All Around

I broke down fullbacks in detail here.

Center Backs

These were developed later, and to be perfectly honest, they are less valid overall than the other positional templates. I knew this ahead of time, but legendary Scotland, Everton, and Rangers player David Weir - who is also a centerback - asked me to take a swipe at creating these and I couldn't say no. They give you a sense of how a centerback plays, but become tricky beyond that.

I do know that Thiago Silva is pretty fantastic, though.

--Ted Knutson

mixedknuts@gmail.com

As you may have noticed from my last update, I have spent a lot of time thinking about how to apply what we learn from studying football stats. How do you take academic knowledge and make it practical? What feedback do we get from talking to players and coaches about how to better apply these ideas in the future? What simply does not work? I cannot stress enough that the world is just getting started in this area. What we understand and think is correct now may dramatically change over the next 20 years, but you do the best with what you have. Progress is a bumpy road. What follows is a goofy idea I was fooling around with some time ago related to developing player skills in ways that are intuitive, especially to younger players. The concept here is that coaches could introduce this, and coach a couple of brief sessions on proper usage and technique, and then players could then use this as part of their free development time to help improve skills. Confession: It might be dumb for reasons I have never thought about but many of the coaches out there have. Here's the rub...  that's fine! Experimenting with concepts and ideas is one of the most important things new people can add to the game. Plenty of potentially bright ideas won't work out for [reasons], but many of those interim steps might be required for something truly brilliant. The other thing I am fairly certain teams do not do enough of is learn from their training information and data. Game data only takes you so far. Being able to objectively evaluate your own players based on training drills from the first team right down through the academy is a hugely valuable resource that can't be replicated in any other way. The ideas below were me poking around with developing an environment where that became fairly natural. Anyway, here you go... the FIFA Skills Pitch You know the FIFA skill drills that help teach you how to perform certain drills in the game? Let’s adapt those and use them for objective performance and skill tracking. The concept is to create a pitch with different areas designated to different types of skill drills. At regular intervals, coaches/analysts track player performance in the various skills and update the skill leaderboards. Game data only generates a very small sample of information about certain important skills. Example: In order to find out who the best shooters in the world are, it can take from 3-5 seasons of data to tease out whether a player is good versus lucky with regard to shooting skill. For skills like long shots, free kick taking, or corner taking it could potentially take even longer to discover this information. However, by starting to track data in training for skills we care about, we can rapidly overcome the small sample size problem and learn in a matter of months what it might take years of game data to discover. We can also tailor player roles better to personnel who actually have the skills we want. And finally we can start to track player improvement in these areas objectively. The objectives here are as follows:

• Increase player development by posting a clearly defined way to evaluate their skills and development
• Better evaluate our own players by creating data driven metrics that teach us more about player skills than we would learn from small sample sizes of game data.
• Post regular leader boards with player rankings across all levels, so that players are incentivized to compete and improve.
• Eliminate on-pitch arguments about who gets to take free kicks or corners or penalties or whatever (Note: This actually happens!). Point to the data, and that person takes it. Want to take more free kicks yourself? Get to the top of the leader board. This should help foster an environment of continued skill work and improvement at all team levels.
• Increase player buy-in by creating skill “games” directly related to one of the world’s favourite player past times – playing the FIFA video game.
• Decay data over time so that players can move up or down the leader board by showing continuous improvement.

If this pitch gets a high amount of use, you might consider developing it on turf, because the free kick and corner locations especially will get chewed to bits. Free Kick Skill Location How?  For free kicks, pick 5 spots on the pitch (Example: central, 22 yds, Right 20, Left 20, Right 26, Left 26). Add two 1-yard wide targets from top to bottom of the goal on each side (make sure they are color coded different colors, since it makes it easier to remember and for the brain to process) plus elevated wall simulations. 2 times a week, have any potential free kick takers take 2 shots each from those locations and track the data. (The fitness guys cannot POSSIBLY complain about 10 total extra free kicks, right?) (Note: Remember to randomize the locations they take kicks from so that they mimic game situations as much as possible. You never get to take a FK from the exact same spot twice in a row in a match.) If you want to maintain the ability for guys to keep practicing and move up and down the leader board (and the on-pitch FK taker roster) then have data expire after a certain amount of time for leaderboard purposes or weight the newer data as more important than the old. Also look into mixing in actual team practice with keepers and walls and tracking that at a less frequent interval to mimic in-game situations. One of the things we might learn here is that certain guys are just better at converting from certain locations and we can use that for in-game decisions. Crossing and Corner Takers Skill Location Do something similar to the free kicks except for corners. Set up elevated targets (or rubbish bins, if you want) at the different locations and heights you want your corners to land in and track the performances twice a week from each side. (This is a really vanilla example from the game itself, but there's a lot that can be done here depending on what is durable and practical.) Obviously there is some subjectivity to the evaluations - this guy hits it harder, or flatter on corners or whatever - but now coaches have actionable information on player skills, and the analysis team has more info to help evaluate the ability and importance of players in the team. [REDACTED STUFF ABOUT SET PIECES] Penalty Kicks Use the same targets as the free kicks except add a large target centrally as well (Insert important game theory point about slow central penalties here). Alternatively, you can literally just recreate the target setup from the video game. Dribble Slaloms One left and one right, exiting into the very top of the penalty box. From there you get 2 options

• Dribble through the slalom and shoot to one of the traditional targets
• Dribble through the slalom and set up a key pass to a number of very narrow target passing areas (say 1.5 balls wide?). Goes along with the concept of making sure all attacking players keep their heads up and are aware of passes to teammates in the box at all times. (Note: Is there a way that we can vary which targets are read as open vs covered?)
• When available, you can use players and defenders here for more realistic simulations.

Variation: Something like this for wide players:

Hi. I am giving a presentation at Science and Football on May 1st entitled The Death of Traditional Scouting. I am looking for a big splash image for the presentation in the style of a movie poster. My current thinking is something like an old Godzilla/Giant Robot disaster movie, but you people are awesomely creative, so if you think of something better, do eet! Prize: £25 Amazon gift certificate. All entries should be submitted to mixedknuts@gmail.com no later than Saturday April 16th, 2016. Please only submit finished entries and not works in progress or my email and head will asplode. I will then choose the winner and the top 5, and include them in a post on the site here. I will also plug your work whenever I use the image. All the best, --TK

Some of you clicked on this just to ask, "WTF is PDO?" Which is fine - we take all kinds here.

The seeming acronym doesn't stand for anything - it was the online handle of Brian King who created the stat in hockey. The definition of the metric PDO is listed below but Wikipedia actually has a page for hockey analytics, so if you want to know more click here.

PDO – Uhhhh…

I’m just going to copy the definition from James Grayson.

“PDO is the sum of a teams shooting percentage (goals/shots on target) and its save percentage (saves/shots on target against). It treats each shot as having an equal chance of being scored – regardless of location, the shooter, or the identity or position of the ‘keeper and any defenders. Despite this obvious shortcoming it regresses heavily towards the mean – meaning that it has a large luck component. In fact, over the course of a Premiership season, the distance a teams PDO is from 1000 is ~60% luck.”

Now you may have seen an occasional tweet from me expressing displeasure with the use of this particular metric, but I've never actually sat down to detail why I think it's dumb. Today I will do that.

Reason 1) It's Theoretically Flawed

Why? Because it treats all shots as equal.

Here's a clue: All shots in football are NOT equal.

Not close. Look at it visually.

This is from one of many pieces by Michael Caley discussing expected goals metrics and it clearly shows all shots are not equal based on distance alone.

Then you add in the whole headers are a lot harder than shots with feet thing that Colin Trainor did way back when and POOF there goes your theory and your metric, and we haven't even gotten to all the other factors that impact a shot's probability of being a goal.

It's kind of sort of fine in hockey I guess because shotqualityomgwtfbbq, but it's just fantastically dumb to use anything that makes this assumption in football.

If you need an image in your head to help explain all of this in personal terms, picture yourself with a football on a football pitch facing a goalkeeper. You take 20 on target shots at the goal from 20 yards out in the center of the pitch.

You also take 20 on target shots at the goal from 6 yards out in the center of the pitch. Which one of those scenarios is going to yield more goals?

Reason 2) It Combines Attacking and Defensive Conversion As If They Are Remotely Related They aren't. Teams technically have infinite choices in how they attack and how they defend. They don't have to be related at all. Therefore, why would we treat them as if they were?

You can have a normal, straightforward average attack and a league leading defense. Or you can have an attack that consistently creates insane chances and pairs it with a defense that gives up exactly the same.

Or you can... well, anything.

The point is that by combining the two separate phases of play into one metric, you miss out on the signal.

"Hey, this team is overperforming PDO!"

Okay, why?

THIS IS ALWAYS THE NEXT QUESTION, and if it is always the next question, then maybe you can - I DUNNO - treat the two phases separately and immediately jump ahead a step.

"This team is giving up far fewer goals than expected in defense."

Aha, now you have my interest. Tell me more.

"This team brought in an attacking assistant coach in the summer to try and boost the number of goals scored..."

Excellent, let's analyze that.

Wait... no team would actually do that in the current football landscape, but if they DID then this would be a very good thing to analyse.

Reason 3) Every Team Does Not Completely Regress

This is a fundamental nerd point, but the fact of the matter is that every team's PDO does not completely regress toward zero, even across multiple seasons.

Why?

BECAUSE ALL SHOTS ARE NOT EQUAL!

There are systemic reasons why some teams allow far worse chances season after season than others. If a team's defensive structure is such that the average shot distance it allows is from 20 yards instead of 15, your goalkeeper has more reaction time on average to make saves, there are likely more men between the ball and the goal, and the team is almost certainly going to post a better save percentage.

Or if you are a crazy high pressing team that tends to keep the number of opposing shots low, but the trade-off is that when someone beats your press they get awesome chances right on top of your goal, then your save percentage numbers are also going to look weird and are unlikely to regress to anything approaching average.

The same applies for elite attacking systems. Some head coaches have an attack that consistently creates better chances than average, which means their shots are more likely to go in the goal, and the team is more likely to post abnormal PDO numbers that have very good reasons to stay that way. And all of this is before we even touch the impact of super elite or sub-par players with regard to skill. One reason why it may look like teams revert to the mean over the course of many years is because manager or head coach tenures last between 12-15 months on average.

Start tracking these things by head coach tenure (or tracking head coach performance across different teams) and it yields a lot more clarity. A weird PDO by a team might be random variation, but there's a decent chance it isn't and for reasons you care about. Other ways of analyzing team performance would be a lot more insightful and should be examined first instead of simply assigning outliers to the random variation dustbin.

Conclusion

Regardless of its common usage in hockey, PDO is theoretically flawed in football and people need to stop using it. Yes, I know there may be data reasons why some analysts continue to use PDO, but as explained above, we should try to find a way past this at the earliest possible opportunity. Do something smarter that better relates directly to the sport you are analyzing. The good news here is that there is now a giant open space just waiting for a clever person to tell the world what they should be using in place of PDO, and that person could be you!