a @StatsPerform #ProForum thread with quick summary and a few slides per presentation 🧵
(full videos should be public in a few weeks)
(full videos should be public in a few weeks)
1: predicting counterattacks before they occur in real-time by Henrick Biermann & co
interesting look at defensive implications of possible turnover when in possession e.g. risk of your current attacking position; what decisions? (e.g. positional shape, tactical fouls 😍)
interesting look at defensive implications of possible turnover when in possession e.g. risk of your current attacking position; what decisions? (e.g. positional shape, tactical fouls 😍)
KPIs: position of turnover, # players behind the ball, vertical compactness
take-homes:
don't lose the ball 😅
don't lose the ball in centre of the pitch
+2 extra players behind the ball
tight vertical compactness
take-homes:
don't lose the ball 😅
don't lose the ball in centre of the pitch
+2 extra players behind the ball
tight vertical compactness
2: ways to break a high press by Bakr Annour & Silvio Matano
like this idea of 'solving riddles' presented by opponents: pre-match preparation to their style, lineup and in-match reaction to their tactics (~game theory)
models: i) RF to quantify importance of own / opposition..
like this idea of 'solving riddles' presented by opponents: pre-match preparation to their style, lineup and in-match reaction to their tactics (~game theory)
models: i) RF to quantify importance of own / opposition..
..features in beating a press
ii) GCNN/graph2vec to generate a vector for comparison of sequence similarity
take-home: pressing intensity, presence in centre, halfspace more important than ball movement
lots of other applications possible with graph2vec expression of sequences
ii) GCNN/graph2vec to generate a vector for comparison of sequence similarity
take-home: pressing intensity, presence in centre, halfspace more important than ball movement
lots of other applications possible with graph2vec expression of sequences
3: (bags of!) KPIs for scoring efficient runs by Caterina de Bacco
effective runs can be useful defensively (e.g. to intercept, to cover space) and offensively (e.g. to receive, to create space for yourself or teammate to receive)
quantifying these five metrics, create...
effective runs can be useful defensively (e.g. to intercept, to cover space) and offensively (e.g. to receive, to create space for yourself or teammate to receive)
quantifying these five metrics, create...
two new KPIs:
i) metres run effectively (i.e. scoring high in at least one metric)
ii) run efficiency %
recruitment applications:
compare run efficiency of two players
examine impact on a *third* player (e.g. will this DM make your CBs defensive runs more efficient?)
i) metres run effectively (i.e. scoring high in at least one metric)
ii) run efficiency %
recruitment applications:
compare run efficiency of two players
examine impact on a *third* player (e.g. will this DM make your CBs defensive runs more efficient?)
4: the story of Ligue 1 2020-21 by Jens Melvang
using some of @StatsPerform's tools - including Power Rankings, Team Shape, Movement Chains, and frame-by-frame pass prediction metrics (xReceiver, xPass, xThreat) - to evaluate Lille, PSG and Monaco at the team- and player-level
using some of @StatsPerform's tools - including Power Rankings, Team Shape, Movement Chains, and frame-by-frame pass prediction metrics (xReceiver, xPass, xThreat) - to evaluate Lille, PSG and Monaco at the team- and player-level
a few findings:
Lille team with highest % of successful, difficult, high threat passing options (44.5%); Jonathon David the player with highest (64%)
Lille team with highest pass option quality
Lille team with highest % of successful, difficult, high threat passing options (44.5%); Jonathon David the player with highest (64%)
Lille team with highest pass option quality
Q&A w/ #TRFC #SWA manager Micky Mellon
“every other formation (than 4-4-2) was designed to beat 4-4-2”
on his new book, The First 100 Days:
- importance of prep before Day 0: idea of what the job is (e.g. start-up, rebuild, sustainer) and how exactly to start
- win games quick!
“every other formation (than 4-4-2) was designed to beat 4-4-2”
on his new book, The First 100 Days:
- importance of prep before Day 0: idea of what the job is (e.g. start-up, rebuild, sustainer) and how exactly to start
- win games quick!
on answering criticism from fans:
- “I pay attention to what’s said on Twitter but not directly” (good advice 😂)
- “[the fans] keep you in a job and they can get you out of one”
- “Don’t lose focus”, “[we] don’t let outside noise affect us or panic us”
- “I pay attention to what’s said on Twitter but not directly” (good advice 😂)
- “[the fans] keep you in a job and they can get you out of one”
- “Don’t lose focus”, “[we] don’t let outside noise affect us or panic us”
- have a long-term plan agreed on by everyone in the building: how are we going to score goals, how are we going to keep clean sheets
- overcome short-term noise and to players in poor form: “get back on track to what makes you a good player for this team in this position”
- overcome short-term noise and to players in poor form: “get back on track to what makes you a good player for this team in this position”
on after the first 100 days:
- Tony Bellew: "[you should] never enjoy training”
- "the best organisations can live in that trauma and suffering [required to sustain performance at top level]"
- Tony Bellew: "[you should] never enjoy training”
- "the best organisations can live in that trauma and suffering [required to sustain performance at top level]"
on using stats at the club:
- "build and challenge processes together"
- get everyone together in the same room, including analysts
- difference in current generation: "we love numbers; if it’s a 2, we want it to be 3” 😂
- "build and challenge processes together"
- get everyone together in the same room, including analysts
- difference in current generation: "we love numbers; if it’s a 2, we want it to be 3” 😂
- use metrics as a motivator for players, e.g. monitor player running and communicating results (e.g. poster with latest stats on gym wall) instantly helped TRFC get more players in the 1000 metre club
- measure improvements in key metrics (e.g. crosses allowed by opposition)
- measure improvements in key metrics (e.g. crosses allowed by opposition)
- explain to players exactly exactly why we’re aiming to improve certain stats, e.g. “do that and we can get the ball in the box more”
- keep monitoring and communicating to sustain performance
“never ever two games that are the same” - biggest challenge is reacting in-game
- keep monitoring and communicating to sustain performance
“never ever two games that are the same” - biggest challenge is reacting in-game
5: a RL approach to scout allocation by Arnav Prasad
scouting allocation has two problems: i) region-level (where to scout) and ii) player level (who to scout)
successful policy balances strategies of exploitation (high value options) vs exploration (uncertain options)
scouting allocation has two problems: i) region-level (where to scout) and ii) player level (who to scout)
successful policy balances strategies of exploitation (high value options) vs exploration (uncertain options)
a multi-armed bandit approach was taken to optimise this balance
simulated rewards ('expected regret') using two existing algorithmic policies - epsilon-greedy (simple) and Thompson sampling (Bayesian) - and a new policy, multi-armed risk-aware bandit algorithm (MaRaB)
simulated rewards ('expected regret') using two existing algorithmic policies - epsilon-greedy (simple) and Thompson sampling (Bayesian) - and a new policy, multi-armed risk-aware bandit algorithm (MaRaB)
MaRaB incorporates our desired level of risk with alpha parameter: high = choose highest value option regardless of risk; low = choose option with best worst-case scenario
great to see probabilistic frameworks introduced and framed in an understandable way re: risk management
great to see probabilistic frameworks introduced and framed in an understandable way re: risk management
6: Luton Town submission by Abhishek Mishra and Soumyajit Bose
use opposition attacking metrics as a proxy for defence to capture otherwise invisible (in event data) aspects of defensive style and quality (~opposition allowed metrics)
use opposition attacking metrics as a proxy for defence to capture otherwise invisible (in event data) aspects of defensive style and quality (~opposition allowed metrics)
defensive presence proxied via KPIs based on convex hull of player's defensive actions (peep deepxg's PATCH): i) passage ratio (% passes completed through hull) and ii) passage value (sum of xThreat from passes through convex hull)
defensive prevention proxied via xG prevented..
defensive prevention proxied via xG prevented..
..(average xG of shots conceded from similar movement chains)
recruitment applications to highlight low-block defenders as well as aspects of defensive style e.g. 😺vs 🐶 defenders
recruitment applications to highlight low-block defenders as well as aspects of defensive style e.g. 😺vs 🐶 defenders
7: how to navigate our of a counterpress with Gerald Lim & Chua Zhi Yuan
pitch control model to continuously quantify counterpressing using three metrics: i) ball-oriented (<10m), ii) man-oriented (5m of players of interest), and iii) passing lane-oriented (~covershadow)
pitch control model to continuously quantify counterpressing using three metrics: i) ball-oriented (<10m), ii) man-oriented (5m of players of interest), and iii) passing lane-oriented (~covershadow)
lots of exploration of scenarios and trends in Ligue 1
- identifying optimal strategies based on maximising own xThreat vs xThreat conceded, e.g. clearance (high risk, low reward) vs attempted long pass (lower risk, high reward)
- identifying optimal strategies based on maximising own xThreat vs xThreat conceded, e.g. clearance (high risk, low reward) vs attempted long pass (lower risk, high reward)
cool applications to opposition analysis e.g. exploiting team weaknesses to pressure and performance analysis re: individual decision-making
• • •
Missing some Tweet in this thread? You can try to
force a refresh
