I spent the last 2 months developing a method for profiling coaches with data. Ultimate purpose? Supporting the recruitment.
Check out my introductory piece for
@AnalyticsFC
⬇️ and enjoy this supplementary dashboard:
Have you ever wondered why football managers don't get profiled as readily as football players?
@pwawrzynow
has attempted to redress the balance and talks about his process in this article for Analytics FC:
Can anyone explain? Liverpool DO NOT WANT Tarkowski - but are happy to continue with Van Dijk?
I mean. The eye test puts JT ahead - but the data is embarrassingly in his favour.
I love how Man Utd fans have to have at least one perfectly good player being widely hated at any point in time. Shaw, Lukaku, Pogba, Fred, now Maguire.
Just have finished an actual app for my dissertation. It revolves mainly around networks, graphs etc. and works for single games only but can obviously be expanded or even maybe connected with BSC Explorer. The only thing left is to binge-document it. Wish me luck.
You've already seen the most typical use case of my coach profiling method which is to tweet controversial dashboards and earn likes. Now for some other stuff - let's try to use the metrics in practice and figure out who could make a sensible Peter Bosz replacement:
Felt a little bad I don't have anything cool to post for the euros but also didn't want to torture you with a 200th pass net you'll see today so I spent the yesterday's afternoon (and the evening...aaand a tiny bit of the night) giving birth to this dashboard:
I was dedicated enough to count votes (TOP3):
1. Bruce 29%
2. MOURINHO 23%
3. Allardyce 13%
Hard to find another coach who overperforms this much. This and the game state reliance should be the major topics questioned by the top clubs that might want to hire him in the future
I searched for a hypothetical Roy Hodgson replacement that would:
1. Improve performance levels and maintain consistency.
2. Prefer high pressing (must-have) and keep the possession (nice to have).
Some interesting names came up ⬇️
1/2
Watching yesterday's game, I couldn't shake the feeling we don't really know how to get our counters going. On the other hand, Spaniards pressed and pinned us well so maybe our subjectively poor display in this area was expected? To figure this out, I obviously employed data:
I'm a little sad nobody requested a single most impactful coach in my db - Marcelo Bielsa 2018-2020 (from his time in Championship) so here he is⬇️.
Also I'm logging off for now but later I'll go through the requests that didn't make it and maybe post some more interesting ones.
So apparently Frank Lampard had finally put an end to the debates on how to fit all this new attacking talent in the XI by utilizing 3 left wingers today.
My non-shot xG model is slowly getting better and better. It made me learn a lot about ML in football, evaluation, calibration, all that stuff. Pretty excited about the outcome.
Today I watched data processing for 8hrs, then had my own little fun coming up with some metrics:
You know how with heatmaps we often compare values to the league avg to better highlight unique patterns? Recently I was working on new data viz tools & got the idea to try it for pass nets too. Still needs work but some cool results already - check out De Zerbi's before & after:
Hmm, the defender tackles, pressures, blocks and intercepts a lot in a team that by definition will tackle, pressure, block and intercept a lot because they sit deeper. Adjust stats for the amount of opportunity or don't compare players.
In pursuit of keeping myself sane, this week I took a break from everything with the word "bsc" in it. How was I spending my evenings then?
Long answer:
Short answer:
If you're wondering how's my Saturday, I managed to combine all of my single-game visuals into one huge match report graphic and started developing non-shot xG model by accident. Nothing is finished yet so here's a blurry pic of something. You know the drill.
Oh boy, awfully quiet in here lately. Last time I was on, I posted like a 100 teasers of my match report template but then kinda stopped developing it cause of no time so I might post the (almost) whole thing now as well. No write-up attached so feel free to ask about anything.
I can finally show you what I've been working on for the past 2 weeks. Classifying positions was just an intermediate step which allowed me to aggregate stats and visuals per the system used.
All of this was generated automatically with only raw event data as an input:
The discourse around Tottenham shouldn't be about whether Mourinho had built a good or a bad team but rather whether his approach is suitable for a top team in 2020.
I made an animated pass net from yesterday. There are maybe 10 frames that don't look like something broke.
So many reactionary takes, time for mine. Get DDG, AWB & Ronaldo out of the club and I'll be beyond happy. Non-starters might be difficult to sell but the perfect clean-up in the next window would look like this for me ⬇️ Now get a proper RB to support Sancho, get a proper DM,
I was graced with the access to full DALL-E today for some reason.
"football player mourning after taking a shot with xG<0.5 (shouldn't have been a goal)"
Panenka works: EXQUISITE MENTAL STRENGTH!!!! HOW CALM!!!!! INCREDIBLE TO DO IT IN SUCH MOMENT!!!!
Panenka doesn't work: DISRESPECT!!! BAD DECISION MAKING!!!! HOW YOU COULD DO IT IN SUCH MOMENT?????
Personal announcement frenzy continues. Earlier this month I joined
@acspezia
,
@SEfodbold
& Casa Pia AC (we don't have a name for the group yet, drop your ideas in the comments) as a data scientist. Exciting times ahead!
But there we go. Determining formations for both teams in a game takes now 10 seconds instead of like 1-2 minutes (it was really terrible). First step towards making some cool things with this on a larger scale.
I’m seeing a yet another discourse about the state of public analytics on here but once again, no actionable insights for the people to implement. So I decided to gather my thoughts into a thread but it got so hilariously long, it's a blog post now:
Have to admit I'm somehow more excited about what Brighton does now than I was about what Chelsea will do. Probably because there is no candidate who makes obvious sense and it's a chance to bring someone fun who's been unnoticed by a wider public so far.
W odpowiedzi na niektóre komentarze, Guardiola i Klopp w surowym outpucie byli na 2 pierwszych miejscach wśród trenerów TOP7 lig. Przewinęli się też Arteta, Tuchel, Nagelsmann i Flick, ale chcieliśmy zarzucić trochę mniej oczywistymi opcjami (które też przeszły test AI).
[THREAD]
My new template - Shooting Report - is finally ready for the public! It took a while to develop and test but I believe I've got everything right. It contains four separate maps and a table at the bottom. Let me walk you through it.
Read More ⬇️
I remember a time when United players , managers , executives wouldn’t be seen in their local Italian after a draw at home let alone getting knocked out of Europe. This last week we’ve seen a global tour of F1 , Concerts, Cricket and UFC events. This lot are Tone Deaf!
Even if it's 3 next games, how can they indicate anything in terms of Ole being good enough for what's ahead? All the ultimatums like these scream cluelessness. There is enough data to act.
Ole Gunnar Solskjær completed the training session at Carrington today, Sir Alex Ferguson was there too. Feeling around the staff is the same: he’s staying. 🔴
#MUFC
Manchester United are prepared to give him another chance with Tottenham & potentially for Atalanta/City games.
Warto być z nami od 17:30.
✅️
@Piotr_Domagala
porozmawiał z Jackiem Zielińskim,
✅️ dzięki
@AnalyticsFC
pokażemy ich autorskie profile trenerów Legii i Rakowa,
✅️ znów z
@MarekWasiluk
zajmiemy się naszymi radarami i wnioskami po połowie fazy grupowej.
Prowadzi
@rkedzior
.
Poza tym algorytm niczego nie wybiera. Najpierw człowiek na bazie domenowej wiedzy piłkarskiej, doświadczenia i jasno zdefiniowanych założeń buduje statystyki/modele. Potem również człowiek testuje czy odzwierciedlają one stan faktyczny i czy pozwalają przewidzieć przyszłe wyniki
Do piłki wszedł VAR i setki innych technologii - pomiary zdrowia, kondycji itd piłkarzy to już norma a ludzie się zastanawiają czy algorytm dobrze wybierze trenera… tak czasem się pewnie pomyli ale pytanie brzmi ile razy pomylił się Smuda kiedy oceniał przez pryzmat schodów…
Made some improvements to the model. Now wingers who switch sides are not confused with CMs so frequently.
Also it understands better when players form a group within the same position (like e.g. a pair of strikers) resulting in Muller/Barkley/Duricic not being STs anymore.
Interesting peculiarity we noticed this week while testing some new ideas for team KPIs at
@WislaKrakowSA
. Last season, Man City created the least amount of immediate danger from goal kicks out of all EPL teams.
Immediate = up to 15 seconds
Danger = scoring probability
This is a serious question btw. Obviously you'll have to do with imprecise proxies for things, impossible to directly transform on-pitch data to explicit indicators of mental traits. But there has been this research:
Don't know who needs to hear this and it's not directed at anyone in particular but data viz doesn't become "good" or useful just by existing. It's the underlying information that's key while the graphs are just tools to convey it.
#MUNCRY
Big U-shaped formation from Man United today. No central penetration, terrible spacing, even worse passing angles and more importantly: how the hell are you supposed to counterpress with that structure? But obviously the transfer activity is our main issue.
This is the kind of stuff we've been working on at
@WislaKrakowSA
for the past year. And I think we're slowly reaching the point where we'll be able to share even more cool experiments & research from behind the scenes. Stay tuned!
Personal update. After 2 eventful years, my time with Spezia & Casa Pia has come to an end. I'm extremely grateful for being believed in from day 0 (shout-out to
@StatifiedF
) and everything we managed to build & learn along the way. It was an unforgettable experience! ❤️
This is a great use of data science no doubt. Perfect for media and general research. But if you work in recruitment, nothing will beat the expert knowledge. Reach out to scouts, coaches etc, understand what do they care for in a player & build roles/profiles based on this info.
"Can Man United take a manager who has just been knocked out by Benfica?"
@seemajaswal
,
@petercrouch
&
@themichaelowen
discuss where Erik ten Hag's future could lie following Ajax's Champions League exit 👀
#UCL
Almost finished my next random Sunday project in time for the game. I clustered progressive passes! Again. But this time I trained a general model for all teams at Euros in an initial attempt to figure out what possibilities such universal model can offer.
It's time.
Score's quite high but mostly due to the team quality. The output close to what an average coach would achieve. He's definitely doing an OK job at United all factors considered but is it enough to earn him a new contract given the growing title aspirations?
I have a script that used to take 2 hours to run. Wasn't satisfied with it so I spent half of the day today working on performance. Now it takes 104 hours.
I've graduated last week which means I have more time for my top secret project now. Still some work to be done before the release but here's a cool bit: I used moving average to automate trend recognition in time series data.
I was dedicated enough to count votes (TOP3):
1. Bruce 29%
2. MOURINHO 23%
3. Allardyce 13%
Hard to find another coach who overperforms this much. This and the game state reliance should be the major topics questioned by the top clubs that might want to hire him in the future
Można :) Dlatego trzeba podejść z głową do tego co się chce mierzyć i w jaki sposób. To jest trochę mój konik, od lipca dużo testowaliśmy i badaliśmy, żeby mieć pewność że to co mierzymy znajduje odzwierciedlenie w wynikach zespołu w dłuższej perspektywie.
Pass Clustering, Take 2: I tried to figure out the best way of automatically determining what makes teams unique. Using all passes (instead of only progressive) helped with the small sample issues. The results are pretty cool already - look at Spain and their heavy circulation!
I developed a small project today - clustering progressive passes made both by and against teams. It could add a new level of insight to the opposition analysis and answer not only a question of what do we need to prepare for defensively but also how it's worth to progress.
🔴🎙️
#konferencjaLIVE
Czesław Michniewicz: Ci, którzy nie grali w defensywie, już siedzą w domach. Taka była nasza taktyka na wyjście z grupy.
#KierunekKatar
Varane is clearly the new guy here. He doesn't yet know we don't really use line-breaking passes so there's no need for anybody to even occupy that space there. It takes him like good 5 seconds of confusion before simply laying it off to AWB.
Let this be my current project's timeline.
The main goal is to create a tool that allows for fully automated and highly contextualized metrics & visuals aggregation while also providing an end user with easy-to-use interface.
@NathanAClark
You're granted a 1-month old Antonio Conte that was randomly lying around on my phone. I've got a feeling it's from that time I was doing some testing only on Serie A. It would explain the most similar coaches. Otherwise it's fine.
#LIVLEE
This is actually my first crack at convex hulls & passes combo but it turned out so good, I might as well share.
This kind of plot supplements networks with some cool new information like e.g. that Hernandez covered the full width.
As always, feedback appreciated!
Let's actually have some fun. The final app is nowhere near being finished but since filtering and plotting work fine, feel free to drop your requests for aggregated graphics like these⬇️
(only EPL and Serie A 20-21 teams though, that's all the data I have processed right now)
I stumbled upon some of my (relatively) very old public work while moving phones. Here's a MONSTER-THREAD of some of the more interesting bits judged in retrospect.
Remember the project that turned out too complicated for my bachelor's thesis? Well, it's still not finished and probably will never leave my local machine but I updated the data today to perform my quarterly "who Man United relies progression on" check.