Yer Da was not the only one upset about the delay in releasing the SPFL post-split fixtures. In his digital man cave, B.U.R.L.E.Y. impatiently tapped his fingers loudly on his metaphorical desk, waiting to spit mathematical hot fire about how he thinks your club stinks. Yet, B.U.R.L.E.Y. had to wait to see who was playing who and where to do this. So with Wednesday’s announcement, he could finally doom your team to their fate.
I am going to go ahead and award B.U.R.L.E.Y. a “Not Too Shabby” grade with these picks back in August. B.U.R.L.E.Y. successfully picked both 5 of the 6 teams correctly in the top and bottom six at the split, with only Kilmarnock performing well above B.U.R.L.E.Y.’s expectations. The robot Craig also got the order of 1-4 correct thus far, though he struggled a bit more on the bottom half of the table, having County and Thistle 8th and 9th instead of 11th and 12th as they currently sit. All in all, not bad for B.U.R.L.E.Y. heading into the split.
We can take each clubs updated metrics at this point of the SPFL Premiership season and see where B.U.R.L.E.Y. puts every club in the end of season table. To do this, we follow the same methodology we discuss here and simulate each club’s remaining schedule 1,000 times. We can then take the average points earned for each club in those simulations and project where they will finish come May.
Looking at the top half of B.U.R.L.E.Y.’s projected table, we do not see any movement from where teams sit currently. The most noteworthy thing from the top half would be that B.U.R.L.E.Y. sees Rangers pulling ahead of the pack for 2nd place, as he puts them on average earning 2 more points than Aberdeen come season’s end with the clubs currently sitting even on points.
The bottom of the table sees a bit more movement though according to B.U.R.L.E.Y. Currently, 5 point separate 9th and 12th, so the fight for automatic relegation and the relegation playoff spot are very much still up for grabs. I have recently discussed the metrics behind Partick Thistle finding themselves at the bottom of the table, and B.U.R.L.E.Y. must have read that article intensely, predicting that Thistle will not be able to pull themselves from the bottom.
The robot also sees County stuck to the relegation playoff spot, while Dundee will jump over Hamilton and finish 9th. He also sees St. Johnstone getting ahead of Motherwell and finishing the year 7th. Clearly with how close B.U.R.L.E.Y.’s projected points from the remaining matches are in the bottom of the table, only the Saints and Steelmen can feel relatively safe about not being relegated come season’s end.
Finally, I just wanted to quickly mention how B.U.R.L.E.Y.’s match up with you all and Nate Silver in picking SPFL matches each week was going. In the graph above, you can see that B.U.R.L.E.Y. has a slight 2 match lead this season over you all and 538. I really wanted to commend you all for matching Silver and Co.’s SPFL predictions. Is the increased awareness of football analytics leading to more informed fans? I certainly hope so! Is my twitter account just an echo chamber? Could be, but I still doff my metaphorical cap to you all matching 538’s model in picking SPFL matches this season so far. Let’s beat that nerd Nate Silver!
In the aftermath of yesterday’s 3-0 loss to Zenit St. Petersburg, the internet has had plenty of thoughts about who the blame for Celtic’s form this season. These thoughts range from lukewarm to spicy hot, such as seeing on one Celtic message board say that “Brendan Rodgers is a fraud”. Regardless, Celtic supporters have had numerous ideas on who to blame for the result in Russia last night.
One of those facing the brunt of the blame for both last night and Celtic’s form dipping this year compared to last is Mikael Lustig. The 31 year old Swedish fullback is in his seventh year at Celtic and has become a cult hero for much of the Celtic fan base. However, it seems the Swede has seem his form dip this year. Before this season, the right back was one of the more consistent players for the Bhoys, but from observation it seems Lustig has not played as well as he has most of his Celtic career.
The above map was created by Dougie Wright and shows where the chances created against Celtic have come from in league play. The darker the green is on the heat map, the more chances have come from that area. It becomes pretty clear that teams in the SPFL see Lustig as a weakness in the Celtic back-line and are trying to expose the area he patrols.
And if we look at the numbers for both the Celtic defense and the rest of the league, we see that to be the case. In the above table, we have the number of Key Passes (or the pass that set up a shot) teams concede on the right wing, the percentage of total key passes a team concedes that comes from the right wing, that percentage of key passes from the right wing compared to the league average, the expected goals from key passes conceded on the right wing, the percentage of a team’s xG conceded that came from key passes on the right wing, and how that percentage compares to league average.
As we see, the numbers are not kind to Mikael Lustig this season. Though Celtic have conceded about the average number of key passes from the right wing, they have conceded easily the highest percentage of the key passes coming from the right wing. 28.57% of the shots Celtic have allowed originated from a pass on the wing Lustig usually patrols, which is 12.21% higher than league average.
Celtic have also conceded by far the highest xG in the SPFL from key passes from the right wing at 5.07, which is 37.62% of their total. This is 22.03% higher than the average in the SPFL. The next highest percentage of xG conceded from Key Passes on the right is 17% less than what Celtic have conceded. Teams have clearly pinpointed Mikael Lustig and the right side of Celtic’s back line as the area to attack and these numbers show that they have clearly been able to do that.
Celtic’s recruitment strategy has been one of the more common places most have put blame for the result in St. Petersburg. Some have asked why a replacement for Lustig was not found either in the summer or January transfer market. Brendan Rodgers has been reluctant to play Christian Gamboa this season, meaning Mikael Lustig has played a lot of football this season. Both by observation and numbers have suggested that Lustig has struggled this season. Perhaps the high number of minutes he has played both for Celtic and Sweden has caused this dip of form, but if this continues Celtic will have no choice but to look for a replacement for the long serving Mikael Lustig.
It seems to be the standard now for Rangers to find themselves in the headlines of Scottish football, both for matters on and off the pitch, on a seemingly weekly basis. From memory, we have had Pedro Caixinha sacking, Carlos Pena being sent back to Mexico and his choice of handkerchief, and Josh Windass’ various hand gestures to fans as examples of off field shenanigans surrounding Rangers this season. However, among the headline grabbing antics at Ibrox, Daniel Candeias has quietly put together one of the better campaigns in the SPFL this season.
I discussed the idea of expected assists last season, but thought this would be a good time to have a refresher on this metric. Expected assists applies the idea of expected goals to those creating the chances. Using the same model we use for xG, xA applies that same number to the player who created the chance via pass. So if Daniel Candeias passes to Alfredo Morelos and Morelos takes a shot worth 0.2 xG off of it, Candeias gets an xA of 0.2. This is an attempt to quantify the type of chances a player creates, rather than something like passes completed percentage.
If we look at the expected assist leaders for the SPFL Premiership this season, we see Candeias is by far and away the league leader. Through February 11th, Daniel Candeias has a total xA of 9.06. The next highest player in the SPFL is at 6.08, quite a ways away. He has an xA per 90 minutes of 0.41, the highest of any player who has played at least 700 minutes this season. He has the highest expected assist numbers from free-kicks, open play, and is 4th in xA in set-pieces. Last season, Niall McGinn had the highest xA in the league and he had an xA per 90 of 0.35, so Candeias is averaging a higher expected assists for every 90 minutes than league leader McGinn did last year. All of this is to say Candeias has been one of the best creators on attack in the SPFL Premiership this season.
Looking at Candeias’ pass map from open play this season above, it is no surprise to see most of his contributions have come from that right wing. From that wing, Candeias has found James Tavernier for 10 key passes, Alfredo Morelos 9 key passes, and Josh Windass for 9 key passes. Most would agree those three are Rangers most dangerous attacking threats, so the Portuguese winger finding them so often has certainly helped lead to his success this year. Furthermore, looking at where the average location for those players were when Candieas set them up for a shot, we see all of three of them located in the “Danger Zone”. This is the area in the 18 yard box in between the 6 yard box and these shots have been found to be the most likely to be scored. When Daniel Candieas is setting up the likes of Morelos and Windass, he is finding them in the most dangerous locations on the pitch where they can score goals.
Against Aberdeen at Pittodrie on December 3, Daniel Candeias sets up the winning goal. He times his run well on the right and puts a first touch low cross perfectly onto the foot of Josh Windass, who is square in the “danger zone” of the box and able to finish, leading to a man hug for Candeias from Graeme Murty. Run on the right, cross into a dangerous position where his teammate is waiting and can easily finish a high xG chance for a goal.
Despite the score finishing 0-0, Candeias had a very good match against Celtic in December. We see another example of what he has done so well in that game. His cross finds an open James Tavernier in the heart of the danger zone, where Tavernier’s shot is only kept out by a good save by Craig Gordon. Another cross on the right to a teammate in the danger zone, who forces a good save from the keeper from a high xG chance.
This season, most of the more ardent Twitter debates in Scottish football have been discussing if certain Rangers players are actually good. Never-ending feuds about whether the likes of Josh Windass and Carlos Pena are good players or not are found at various corners of the great time waster known as Twitter. Daniel Candeias thankfully does not draw such hard line opinions. Most seem to know he is valuable to Rangers success this season. However, it is a bit surprising his praises have not been sung at such a level as they have with someone like Alfredo Morelos by Rangers supporters. Both have been key cogs to Rangers attack this season and Rangers will need them to continue their form if they are looking to finish second.
One avenue to try and create understanding about analytics in football is to use them to praise players. Letting a player know he has good stats, what those stats mean on a basic level, and what he is doing to get those impressive metrics can help foster acceptance and further understanding from players. Which brings us back to our pal Kris.
If I told Kris Boyd, “Hey your expected goal numbers of 7.90 are great,” he would most likely look at me as if I had four heads, with two of the heads telling him to switch to a vegetarian diet. However, he might be more receptive to the idea if I said something along the lines of “we have this stat that says you have done a great job at creating shots that are more likely to be scored. If you continue to be able to get shots centrally in the box, you will likely keep scoring goals,” he may be more receptive to the idea. A player may be more interested in analytics if it is includes praise and includes an easy to understand idea to lead to continued success.
This hypothetical discussion with Kris Boyd does indeed contain some of the reasons why he has seen a resurgence on the pitch this season. Boyd has 9 non-penalty goals and 0.56 non-penalty goals per 90 minutes so far this season, only Alfredo Morelos has more in the SPFL Premiership. He also has had good underlying statistics this season, with an xG of 7.90 which is 4th in the league and an xG per 90 of 0.54. These xG stats show his goal scoring has been sustainable for Kilmarnock this season.
If we compare these stats to the same numbers from last season, we see a clear improvement. Last season, Kris Boyd scored 7 non-penalty goals and 0.32 non-penalty goals per 90 minutes. Like this year, Boyd’s goals were about what you would expect based on his expected goals numbers, with an xG total of 7.38 and a per 90 average of 0.30. We see the striker/pundit has already surpassed both his goals and expected goals total from last season and we are only into February. While you may or may not enjoy his TV work this season, Boyd’s goals have been a big part of Kilmarnock’s success this season.
Not only has the Killie striker been knocking in the goals this season, but he has been a bigger part of the Ayrshire side’s attack overall. Looking at his expected assist numbers from last year and this year, we see last season Boyd had 3 assists, an xA of 1.67 and 0.08 xA per 90 minutes. This year while he only has 2 assists, Boyd already has an xA total of 2.58 and a per 90 of 0.16. His contributions to the Kilmarnock attack has improved, with improved shooting and creation statistics this year.
Given that Kris Boyd is 34 years old, this improvement from last year is a bit surprising. Strikers typically regress at that age, something many may have assumed was happening last season with Boyd. So the question therefore becomes what has happened between last year and this year to see this improvement from the Killie striker.
During this campaign, Kris Boyd has been able to get far better shots than he did last season. He is averaging 0.16 xG per shot this year, while he only averaged 0.11 xG per
shot in the 2016-2017 campaign. To simplify this, on average every shot Boyd takes this year has had a 5% higher chance of going in based on where and how he has been shooting. Not only is he taking higher quality shots this season, but he also taking more of them, averaging 3.4 shots per 90 minutes this season and 2.59 per 90 last season. Taking more and better quality shots is clearly one of the reasons for the rise in Boyd’s form this season.
Thirty-four year olds do not usually see those types of jumps in those numbers from one season to the next, so why has Boyd? Well, he has some help shouldering the load at Kilmarnock this season. Last season, Boyd was the only Killie player who played at least a third of the available minutes in the league last season (1140) with an xG per 90 of at least 0.3. Only Souleymane Coulibaly had an xG per 90 of over 0.2 at Kilmarnock last year, and he left midseason. No player who appeared in at least 1140 minutes averaged at least 0.1 xA per 90 or more for Killie last season. Boyd was the focus of the Killie attack last season.
The current campaign must feel like a weight has been taken off of Boyd’s shoulders. Though both playing less minutes than Boyd, Eamonn Brophy and Lee Erwin are both averaging over 0.2 xG per 90, at 0.68 and 0.23 respectively. Both are options off the bench for Kilmarnock that are capable of scoring. Along with other scoring options besides Boyd, the Killie attack also has considerably more options for players that can set up Kris Boyd and his fellow strikers.
As previously mentioned, no Killie players averaged over 0.1 xA per 90 minutes last season. Boyd has already bested that himself, averaging 0.16 xA per 90 minutes, but he is not alone. Along with Boyd, Gary Dicker, Brophy, Jordan Jones, Adam Frizzell, Stephen O’Donnell, Erwin, and Rory McKenzie have all appeared in at least a third of Killie’s available minutes and have averaged at least 0.1 or higher xA per 90 this season (0.20, 0.19, 0.17, 0.16 , 0.14, 0.13, and 0.12 respectively). Boyd is clearly flourishing with more creative support around him.
While he may not be interested in knowing the math behind it, there is no doubt Kris Boyd has been one of the better strikers in the SPFL according to the stats. Analytics skeptical players like we can presume Boyd is may not be interested in regression models and scatter plots, but they are surely interested in becoming better players and they are definitely interested in receiving praise. Praise and recognition for a player can lead to more playing time, improved contracts, and a move to a bigger club. Framing the discussion around football analytics to a skeptical player about how it can improve their career may be a bridge to understanding for them.
Many describe third place teams in the Champions League groups “parachuting” into the Europa League. If we apply this metaphor to Celtic’s performance in their final Champions League group match against Anderlecht, the Celtic plane suffered some turbulence before the crew ejected and headed into the Europa League and Stuart Armstrong was the drunk pilot doing shots of whiskey before take off.
Overly complicated metaphors aside, Stuart Armstrong’s performance in the first half at home to Anderlecht had many Celtic supporters ready to move on from the midfielder. Armstrong is coming off a season last year where he was seemingly the first name on the Celtic team sheet, getting call ups and starts for the Scottish National team, and even leading the Tartan Army to victory (albeit too little too late for the World Cup). Armstrong finished the SPFL season 13 goals and 6 assists, good for a 0.79 Goals + Assists per 90 minutes.
However, as this season began and the good feeling surrounding the well quaffed midfielder started to fade away. Armstrong was on a contract that expired at the end of the season and seemed to be an impasse on a new agreement. Though a one year extension was eventually agreed to, as fans sometimes do, Celtic supporters were not too pleased with a player who seemed to be looking for greener pastures.
This feeling, combined with a reduction in goals and assists has only furthered the belief among some Celtic supporters that Stuart Armstrong is now surplus to requirements. Indeed, Armstrong is on the score sheet less, scoring only 1 goal and 3 assists so far in league play, 0.49 Goals + Assists per 90 minutes in the SPFL. These numbers have some Celtic fans asking, “What is wrong with Stuart Armstrong?”
And the answer to that question is “Nothing, thanks.” We can expand upon that answer and add a “regressing to the mean”. If you are here, you are likely at least somewhat familiar with the advanced stats in football and that they may tell us more about a players performance compared to traditional stats. While Armstrong may not be hitting the heights he did last season in goals and assists, what do these underlying stats such as xG, xA and others say?
If we first compare expected goals for Stuart Armstrong, we do see numbers that have gotten worse so far this season. Last season in 2,168 minutes, Armstrong had an xG per 90 minutes of 0.33. If we compare that to this season, we see a decreased xG at 0.12 per 90 in 818 minutes. It seems so far that the midfielder is no longer the goal threat he was last season.
However, we might need to dig a little deeper into these shot numbers. First, let us address the elephant in the room with Armstrong’s xG numbers last year. Despite putting up very impressive numbers, Armstrong overachieved his xG numbers in his goals scored, scoring 13 goals with an xG of 7.90. Now, some players are able to consistently able to over-achieve their xG over multiple seasons, but so far this season it seems Armstrong is not showing he is able to do that.
Of course, it is still too early in the season to come to widespread conclusions, however there is no shame for a central midfielder not being able to continuously finish at that high of a rate. With his xG total at 1.02 and 1 goal scored, Armstrong is about at where we would expect. In addition to perhaps seeing a regression to the mean scoring goals, Stuart Armstrong has also found additional competition for spots in the Celtic midfield.
While Armstrong was not around the Celtic first team until this time last season, from that point on it seemed he was one of the first names on Brendan Rodgers team sheet. However, this season we have seen the emergence of Callum McGregor, goal scoring threat, as well as Oliver Ntcham’s arrival. McGregor has contributed 5 goals already this season in league play, while Ntcham has also added 3 goals.
This competition has seen Armstrong playing less. Last season, Armstrong appeared in 63.39% of available minutes in league play, while this season he has only appeared in 54.37% of available minutes. He has seen a slightly reduced role so far this season, but while he has not been able to score as much for Celtic this season, he has contributed to the attack other ways.
While Armstrong might be seeing some regression when it comes to goal scoring and xG, his passing numbers suggest he is at a similar if not better level than he was last season. Last season, he averaged 0.24 xA per 90 minutes, while this year he is averaging 0.29 xA per 90. He averaged 1.79 Key Passes per 90 minutes (or passes that lead to a shot), but is averaging 2.4 this season. All of this has lead to a similar Assists per 90 minutes for Armstrong as last season, averaging 0.37 per 90 this year compared to 0.42 last season. He is setting up his teammates just as well when he is playing, he is just seeing them converted at a slightly reduced rate.
Along with stats showing how he sets up shots, Armstrong also seems to be just as vital in Celtic’s attack overall as he was last season. Looking at xSA, which quantifies the pass before the pass before the shot, Armstong is averaging 0.17 xSA per 90 minutes this season, while last season he averaged 0.07 per 90. This is another metric showing Armstrong has been more than a great head of hair this season.
Tuesday night was a rough night for Celtic and their supporters. There is no hiding that Armstrong had a poor game and was subbed out because of it. This game seemed to be the final piece of evidence for many Celtic supporters that Armstrong has regressed and is no longer necessary. However, over this European campaign, who DID have good performances? Ntcham seemed to do much better than Armstrong at home against Anderlecht and set up goals in Belgium, yet his passing was erratic before those goals. The list of Celtic players who get a passing grade in Europe this campaign is a short one and casting Stuart Armstrong aside because of a small sample against superior competition seems short sided.
Yet, Armstrong still seems destined for pastures new soon. He added a year to his contract, but will find himself on an expiring contract at the start of next season. While he his stats suggest a player still able to create for his teammates, we might not be able to expect a double digit goal total from him every year. If Celtic were to get in offer in January or in the summer that is near their valuation of him, it might be wise to sell. If someone offers a transfer sum for a goal scoring midfielder for Armstrong, I would certainly take that offer.
Congratulations Scotland, you have passed Intro to Expected Goals and are now moving onto the advanced class. Most following me know that expected goals are the likelihood that a goal will be scored on a shot. Expected goals is now a term that more and more Scottish football fans are familiar with, understand, and can discuss coherently. Sure, there is the occasional “Yer Da” still yelling about “Goals and Points being the only stat that matters!”, but compared to three years ago, football analytics literacy has grown considerably in Scotland.
However, now that many have the basics down, we need to have a talk about expected goals. On Twitter last week, I noticed there was discussion about the usage of xG and in particular summing xG totals for individual matches and saying things like “(this team) should have scored 2 goals because they had an xG of 2.” Let me first throw myself at the mercy of the metaphorical court, I have created a few different visualizations where a summed xG total for an individual match was present. It is still on the xG maps I publish each week for the SPFL.
I chose to sum xG on the graphics I have posted to try and ease Scottish football fans into xG. With that being said, there are some issues with summing xG for individual matches. Danny Page covers the issues in an article he wrote pretty comprehensively. Danny points out that if you sum the xG, you will miss on on the variance that can occur in a single match. In his article, he says:
Arsenal won 0–3 with a xG scoreline of 0.39–1.49. In these cases, some may say “The right team won” because the xG and real life scorelines match. However, these values are only adding expected goals. But something is missing. Only adding independent probabilities misses half of the story: variance.
A good situation to think of here is a shot with an xG of 0.05. That shot may go in, it has gone in before, but it is not likely. The instances where it does go in is the variance Danny is talking about, but generally it is not a shot that is going to lead to goals often. But let’s say that a team has ten of those 0.05 xG shots, compared to a team that has one 0.50 xG shot. The second team’s shot is much more likely to go in than any of the first team’s shots, but summing the xG in this situation they would both have an xG total of 0.50.
Sometimes those lower xG shots will lead to a win, thus the idea of variance. Typically summing xG over the course of a season variance usually will find the mean. However, anything can happen in one game. Therefore, Danny puts forth that rather than summing xG totals in a single match and making conclusions off that, it would be better to use win percentages based on the xG of each team’s shots and the likelihood of the goal difference for that match based on the xG output, so that is what we are going to do.
To do this, we will take the xG of each shot for a team in a match and run them in a Monte Carlo simulation 1000 times. This is similar to what I do to come up with the numbers for B.U.R.L.E.Y. for the season. With these simulations, we can come up with 1,000 results of matches with the xG results of a particular match and produce how many times each team would typically win and draw, what would be the most common scoreline, and the typical points per game from that xG performance. In addition to seeing the sum of the xG for a match, we will see the team that was most likely to win and what the score would typically be from a match with that xG output.
Using Danny’s xG simulator and taking all the graphics he came up with as a template, I will now be producing these graphics for every SPFL match. Henceforth, these graphics will accompany the xG maps we have been producing each week and will hopefully give some further insight into expected goals. As this now the “advanced class”, please feel free to let me know if you have questions or comments about this!