Atlanta United Mid-Season Analytical Review

Author Notes: This was written at the halfway point in the season. No information from matches after the June 24 match vs Portland is included. Had to cut it off somewhere.

Also, every bit of data in this post comes to you courtesy of AmericanSoccerAnalysis.com (@analysisevolved on twitter). I want to highlight that upfront because the team at ASA has been producing some amazing content this year covering all of MLS (and more). Personally, I think followers of MLS are a bit spoiled to have a group so dedicated to analyzing MLS based on data and tactical analysis. And importantly, they continue to #evolve the tools and presentation of things there. So, definitely check them out if you have not, or check back in if it’s been a while. On to the Mid-Season Review.

Kings of the South on a Supporters Shield Pace

After 17 games played, Atlanta United sit atop the combined MLS standings, both in terms of total points and points per game. Sometimes a team leads the pack because they’ve played more games or more home games than their closest competitors. That is not the case here. Even after adjusting for the number of home and away games played, Atlanta lead New York City FC and Sporting Kansas City by 1 point at the midway point of the season after balancing the schedules:

A Quick Note on Measuring Underlying Performance and Predicting Future Success

Sometimes over short time periods, results themselves can hide a team’s true quality. In a sea of mostly unreliable metrics for predicting a team’s future success in soccer, points earned to date is actually one of the worst. Goal Difference is slightly better but still bad. Because goals are rare, the ratio of a team’s shots for and against is in fact a better predictor of future points than goal difference. And further, because not all shots are equally likely to score, the ratio of expected goals for and against (which takes into account not only the volume of shots but also an estimate of the quality) is an even better predictor of future success than shot ratios (but admittedly, still not a great one!). This is why you’ll often see a small but growing portion of the soccer world discussing expected goals in the middle of the season. It’s the best metric we have for predicting future goals and thus future points. Analysts (and gamblers) want to look forward, not back. They want to know if a team is likely to continue to pick up results or continue to struggle, or conversely, if the past might not look like the future.

This is not easy, and the results may vary, but let’s give it a shot looking at Atlanta United because .. I mean.. isn’t this what everyone wants to know? Below, we’ll start with combined metrics (attack and defense) first, just to get a feel for where Atlanta currently sits relative to their peers. And then if there’s something interesting on either side of the ball, we’ll dig deeper. And while it’s not perfect, we’ll use one of the better objective tools we have: expected goals. Again, the data is from AmericanSoccerAnalysis who have been setting the MLS world ablaze this season with informed insights on the league and the sport itself.

Atlanta United has strong underlying metrics (but not the best)

Looking at combined metrics for attack and defense, Atlanta ranks 2^nd in expected goal difference per game at +0.74 (just behind league leading Columbus’ +0.78 and ahead of the Sporting Kansas City’s +0.61). This expected goals figure includes penalty kicks at a value around 0.785 per attempt (which not coincidentally is the historical conversion rate for these chances). AmericanSoccerAnalysis recognizes that penalties are few in number but high in likelihood of scoring and therefore may introduce some noise into the predictive power of expected goals over shorter time periods. Therefore, they’ve built a separate xG model that decrements the value of penalties from 0.79 to somewhere around 0.25 (I think of this as trying to award a team for the value of the chance they were on average likely to create before the foul in the box ruined things). The idea is that this is more likely to do a better job predicting a future team’s performance since teams generally don’t produce penalties at high frequencies. This is especially important for a team like 2018’s Atlanta United which has earned 10 penalties through 17 games in 2018 (on “pace” for 20). The highest single season penalty tally since 2015 is 11 (Red Bulls in 2015. Portland and Toronto in 2017), which doesn’t suggest Atlanta will stop earning them, but maybe that it’s unlikely to earn 10 more. Anyhow, when you look across 2018 using this alternate view at expected goal differential, Atlanta is 5^th at 0.52/game behind Columbus (+0.66), Red Bulls (+0.6), Houston (+0.56), and Kansas City (+0.54). And if you put more stock in open play chance creation than elsewhere (perhaps because most of the sport takes place in open play), Atlanta is 3^rd at 0.50/game (1.41 created and 0.91 allowed) behind LAFC and Houston. That is all to say that Atlanta are one of the top teams in MLS, and the underlying metrics support the overall output and results we’ve seen. I believe the team is good. But the 2017 team was good too. To go further, I want to understand the differences.

Expected Goals show Improvements in attack in 2018

A quick compare versus the inaugural 2017 Atlanta United reveals an improvement in these underlying metrics. 2018’s xGD of +0.52/game is a tremendous jump from 2017’s +0.13/game. And, this improvement is mostly on the attacking side of the house. While the value of the chances given up remained mostly constant from 1.26 xG/game in 2017 to 1.25 xG/game in 2018, the value of Atlanta United’s chance creation has increased 35% from 1.29 expected goals created per game in 2017 to 1.75 in 2018. These figures include the lower values for penalties (Five Stripes averaging over 2 xG/game in the more penalty-friendly model). Here is a high level summary of how the chances are being created year over year (measured in expected goals created per game):

We know what’s up with the penalties thing. And the dead ball situations is interesting, but not huge.

I dug further into the open play chances and plotted all of them on shot maps just to see what was up.

It’s not an ideal visualization method for this, but what popped out to the eye was the six-yard box. While Atlanta have played half the games in 2018 to date, they’ve created the exact same volume of open play chances inside the 6 yard box as they did in the entire 2017 season. That’s a rate of 0.9 per game in 2018 vs 0.4 per game in 2017, and that’s important because these shots have expected goal values in the 0.4-0.5 range (you expect to convert just under half of these). You simply are not going to find better open play chances.

For what it’s worth, of the shots created inside the six in 2018, 60% were assisted and 40% unassisted. In 2017, 80% of the shots created inside the six were assisted and 20% unassisted. There is perhaps slightly better scrambling going on in the goal mouth.

If we extend the 6 yard box out all the way to the edge of the penalty area, and dub it the “danger zone,” (as soccer analysts Michael Caley and Mark Thompson - Mark also an Atlanta United supporter) would have us do, the 2018 Atlanta team is creating right around the same number of open play shots per game compared to the prior year (4.0 vs 3.9). These figures include both the six yard box and the area directly beyond it extending to the edge of the 18 yard box. Shots from the deeper area beyond the six have average xG values between 0.15-0.20 and so all in, as a result of shifting some of these chances further forward into the six yard box, the 2018 Atlanta team have created a more potent overall mix of chance creation in the danger zone, with xG per shot of 0.26 replacing last year’s 0.20 per shot rating. I believe this is important.

Perhaps more striking is the increase in volumes of shots from wider areas within the 18 yard box. These chances are historically converted much less frequently (average of 0.07 xG/shot). In 2017, Atlanta created 2.3 shots per game from these areas, and in fact converted them at a high rate of twice the league average. In 2018, Atlanta are creating 3.1 of these shots per game, and have yet to score one (update: after finishing this piece, Miguel Almiron did finally get his goal from the wide angle we’ve all been waiting so desperately for). If they are increasing their open play chance volumes in 2018 (they are: 10.8 in 2018 vs 10.4 in 2017), it is primarily due to this additional shot per game from the wider areas inside the 18 yard box, and this increase is partially offset by fewer open play shots per game from outside the box \ (from 4.2 per game in 2017 to 3.7 per game in 2018). This is mostly a good trade-off, and one that you might expect to pay off with marginal benefits. Unfortunately for Atlanta, they just haven’t enjoyed the increased conversion rates that you’d expect when you more than replace those deep shots (0.04 xG per shot) with the wider penalty area chances (0.07 xG per shot).

The following distribution maps show where Atlanta’s chance creation comes from relative to the league as a whole. Note the substantially larger share near the goal mouth.

In summary, on the attacking chance creation side of things, Atlanta United have increased their xG per game 34% from 1.3 to 1.7. In open play, they’ve increased xG created per game 28% from 1.1 to 1.4 primarily by:

• Increasing the overall quality of their shots (as measured by xG).
• Taking fewer shots from outside the box (half a shot per game decrease).
• Shifting a half a shot per game forward into the six yard box from deeper areas.
• Adding 0.8 shots per game to the wider areas within the 18 yard box.

The remaining increase in overall xG per game has come from earning 10 penalties in 17 games.

Penalty shenanigans aside, that open play output is very good, but should we be concerned that the underlying open play chance creation itself is unsustainable? A couple of additional points to make:

Is there any reason to believe this won’t continue?

First, it should be noted that of the 0.3 (28%) increase in open play xG per game, a significant portion of the increase occurred while Atlanta were up at least one player (maybe as much as half). The basic context here is that 84% of Atlanta’s 2017 xG chance creation output occurred with teams at even strength compared to 72% so far in 2018. Further, the long range historical MLS average is that 92% of xG chance creation comes with teams at even strength. Is there something to Atlanta repeatedly finding themselves in uneven game states? Maybe. But it could also just be a case of small sample sizes. Anyhow, if you take away the increase in xG output per game in advantaged game states, then you’re looking at something like 1.27 xG per game, which would still put Atlanta in historically high company in terms of attacking output from open play. A list of the top 10 xG/game attacking sides from 2015-2017 looks like this:

Atlanta’s 1.27 open play xG/game would put them safely atop recent seasons, bested only by some other 2018 teams (either we’re seeing a flagship attacking year in MLS, or the second half of the season won’t look like the first). You’ll notice I included shots inside the six-yard box in the above list. This is because I’m curious as to whether Atlanta can continue to shift its shots into the utmost advantageous scoring locations. If Atlanta were to continue to put up 0.9 shots inside the six yard box per game, it would stand above all other recent top attacking teams with only 2015 Columbus’ 0.85 shots inside the 6 per game coming close.

So perhaps it’s wise not to expect the underlying chance creation to continue at such a steady clip, but even after adjusting for a few of those worries, Atlanta’s attack seems likely to continue to be good. On the one hand, Nagbe’s absence could hurt this trend, but Tito Villalba coming back should only bolster the attack. On top of all of this, there’s some additional noise coming from a shift in tactics. We’re likely at a turning point, away from the 3-5-2 used for 13 out of the first 17 games of the year, and towards a 4-3-3 / 4-2-3-1 (depending on how tired you are of formation talk).

Sample size warning, but for what it’s worth, here are the underlying expected goal per game outputs for most of the first half of the season in the back 3 compared to the emerging return to the 4-3-3.

Reason for hope: Atlanta should start finishing better.

If we should temper our optimism on the prospects of Atlanta continuing to create such high quality chances on into the second half of the season, then we should also exercise restraint in our dismay over the team’s poor finishing record through 17 matches in 2018. First, here’s a comparison of the finishing overperformance/(underperformance) year over year (goals minus expected goals per game):

2017 was a historical oddity as Atlanta United averaged over a half a goal a game more than the underlying expected goals output. This overperformance was almost entirely accomplished in open play. Conversely in 2018, the team is slightly underperforming in open play. This year over year difference has definitely been felt by the supporters, partly because the baseline of the team’s first year in existence was skewed, and partly because teams that create a bunch of good chances (as Atlanta have done) also miss a bunch of good chances (reminder: this is mathematically true). And missing a bunch of chances sucks, and everyone hates it. Have Atlanta United missed more good chances than they should have? Yea, probably. But the good news is, the rate at which teams overperform or underperform their expected goals rarely persists, while the rate at which they create chances does to a much greater degree. And further the rate at which they create chances is the best indicator we have of their ability to score goals in the future. To illustrate, below is a sneak peak I was afforded at some very cool functionality that’s in development from ASA in their interactive xG charts (here). These are scatter plots showing team averages over the first halves of seasons against their complementary second halves of seasons. The point here is to more easily visualize how stable certain statistics are. Note the very small correlation metrics (r) in finishing but the more significant stability in chance creation as well as chance creation as a predictor for future goal scoring.

So while I can’t be certain, my best guess is that Atlanta will start finishing at a rate that brings their goals closer to their expected goals (rather than 0.2 per game below it). Supporters should find comfort in that.

But is this team different?

I realize the counter argument to all of this is that in the complete history of Atlanta United (2017), the finishing rates have absolutely persisted from the first half of a season to the next. Many were certain that after 17 games last year with the team putting up 1.9 goals per game and overperforming xG by 0.7 per game, that the scoring would come back down to earth, or at least mirror the expected goals output over the remaining games. This did not happen, and in Atlanta’s final 17 games they put up 37 goals on 27 expected goals, an overperformance of 0.6 per game (with a healthy chunk of that overperformance coming with uneven teams on the field – adjusting that figure is difficult but it could be something like 0.3? from even states).

The thing is, to make the argument that there is something special about Atlanta that substantively caused their scoring rates to soar above the underlying chance numbers (and I’ve made this argument before), it would have to have first and foremost *stopped working in 2018* and then go something like this… either… 1) Atlanta’s high pressing and vertical playing style created the kind of 1v1 open chances that are a blind spot for expected goal models (i.e. the xG was too low), or 2) Atlanta’s players were(are?) naturally better at finishing goal scoring chances than historical MLS players (by very wide margins), or 3) a bit of both. The alternative to these theories (the mainstream analytics take) is mostly that a combination of game state and small sample size helped to produce most of the 2017 overperformance.

The problems with #1 are numerous. First, there are other high pressing sides that historically have not experienced the same sort of overperformance in goals and expected goals. Take New York Red Bulls or NYCFC for example. Red Bulls underperformed in the first half of 2017 and overperformed in the second half, and NYCFC did the opposite. Sporting Kansas City, another high pressing side underperformed throughout the year, and RSL (the 5^th highest pressing side by my estimations in 2017) looked similar to Red Bulls. Perhaps more immediate is the fact that Atlanta United have underperformed to begin the season. So, it’s not like there’s a precedent for high pressing teams across MLS to overperform the models consistently.

On number 2, while it is certainly intuitive that different players have different levels of shooting ability and there is some evidence to back this up (that elite high-volume shooters can in fact beat the averages over time), there’s both limited sample size of Atlanta players doing so (a single season in MLS for Martinez and Villalba beating the models, and to a lesser extent Almiron and Asad who also finished above their xG last year) plus the glaring fact that Martinez and Almiron aren’t running so hot in 2018.

All that is to say, it’s not as if Atlanta’s players lost some sort of shooting ability in the offseason. The most likely outcome is probably that over the remaining 17 games in 2018, their goal tally will be reasonably close to their expected goal numbers unless we see multiple weird red card games like... like we have before.

If I had to make a more heterodox argument though, I’m curious about (but do not necessarily endorse) this line of logic:

A. In 2017, having Hector Villalba & Josef Martinez & Miguel Almiron & Yamil Asad in a high pressing transition-based attack was one too many quality attacking players in that system for defenses to defend adequately, and the space it created for each of them was the kind of space that is difficult to capture in the on-ball event data that feeds most expected goals models. But it is this very space that by experience and habit tells each player that a good shooting opportunity has arrived, and so the players took good shots (better shots than the models credited them for if only by marginal amounts).

B. Without Tito Villalba (mostly injured) and Yamil Asad (RIP), the 2018 attack of Josef Martinez & Miguel Almiron & Ezequiel Barco (and Julian Gressel), all of them quality attacking players, and in a more reactive setup, did not have the same sort of space available as in 2017 in which to shoot. The result has been that the team has had to work harder to create similar quality chances, and that these similar quality chances have had to be from traditionally better locations (i.e. closer to goal) to afford the same types of shooting opportunities that the increased space in 2017 allowed. In short, the team may be creating truly similar quality chances this year as last but the statistical models pickup on the more measurable differences in chance quality this year (distance from goal being closer).

C. Last year’s Atlanta United team aggressively disrupted their opponents’ buildups at a rate bested only by the infamous New York branch of the global pressing coalition that is Red Bulls. This year due to the loss of Yamil Asad and Carlos Carmona and the lengthy absence of Tito Villalba, the team has mostly sat back in a back 3 / back 5 setup and executed very well on a strange combination of calm buildup in possession and countering from deep in transition. Based on my calculations, the 2017 team allowed their opponents to pass the ball from their own half 15 times per defensive action compared to 18 times per defensive action this year. I expect this to change over the remaining part of the schedule. First because Villalba has returned and the team have returned to the “shades of Bobby Dodd” 4-3-3. But also, because as I glance at the remaining run-in, I see games against DC United (X2), Montreal, Colorado, San Jose, and New England. I’m not saying the high press on its own generates chances that consistently beat the expected goal models. I’m wondering if it is another factor into this question of spacing and shot selection. More than anything, I’m excited (and anxious) to see how a “new-look” attack might do for a team already sitting relatively comfortable atop the standings.

Please do not sell Miggy.

There’s a lot of knock on questions I’d like to tackle next on the topic, but they’ll have to wait.

Conclusion

Not only has the team put up the best results in the league so far this year, there is reason to be confident (if you weren’t already) that this is not a mirage — that this team is in fact, good... that in a sport where anything can happen (at this point last year, FC Dallas and LA Galaxy looked like the 2nd and 3rd place teams in the West), the underlying metrics support that the team will continue to get good results. Interestingly, Atlanta United is in a situation where the team’s play style over the second half of the season might look very different than what we’ve seen to date. While this will always add some uncertainty, there are reasons to believe the team will continue to create good chances and limit the chances of its opponents. Further, there is reason to believe Atlanta United will find the back of the net at a more consistent rate than we’ve seen to date. Update: LOL, like they did against Orlando City.