Untitled

FROM THE PITCH TO THE ALGORITHM

Hosts: Ian Renshaw · Keith Davids   |   Guest: Ben Teune (WTA / Victoria University)   |   Duration: ~64 min   |   Released: 3 July 2026

A common complaint about the Constraints-led Approach to coaching is that it is not theoretical — it's practical. There is not a clear theoretical rationale for constraints manipulation. It’s based on guess work. There are too many variables to manipulate. In this episode, Ian Renshaw and Keith Davids sit down with Ben Teune, sports scientist at the Women's Tennis Association (WTA) and PhD graduate from Victoria University In Melbourne, to explore whether data analytics can help to give coaches a principled toolkit which they can use.

Ben's PhD research programme was embedded inside the Western Bulldogs AFL club in Sydney for four years. A subsequent publication from some of his data — written at midnight in what little spare time a full-time WTA job allows — builds on that work to ask: Can machine learning quantify constraints interactions? Can data analytics inform coaches how representative their training designs actually are? Can clustering algorithms classify practice  drills by variability level? His answers, backed by method and data, are positive as we summarise next.

01  The Turning Point Isn't Always Theory. Sometimes It's a Drill outcome Going Wrong on a field.

Ben's shift from the area of strength and conditioning into skill acquisition came not from reading a research publication, but from standing on a pitch side-line watching his team lose 0-6, looking at a one-on-one contest drill, and taking a close look at the data. A 12-week sprint program might improve a player's running speed by 0.1–0.2 seconds. But, in practice, if players learn to  use perceptual skills to  read the play earlier, using anticipation and making better decisions to move and position themselves in more sophisticated ways, they could make the most of their S&C work. That thought led Ben to a great topic to research on a PhD programme. The implication for practitioners: the entry point for ecological thinking is often an honest question — what is actually limiting performance here?

→ Clip 01

02  You Already Have the Data. You're Just Not Framing It With a Skill Acquisition Lens.

Ben is emphatic on this point: the GPS data, event data, and tracking data that sports scientists routinely collect are mostly used for physical load monitoring — and not much else. But, the same datasets contain the raw material to quantify constraints interactions, measure representativeness of practice designs, and categorise  performance variability. The gap isn't technology or data — it's the theoretical interpretation of the data. When sports scientists start asking skill acquisition questions of existing data, the picture can change suddenly. The publication under discussion was written as a demonstration of what those applications methods might look like. So,  practitioners can apply the findings in their own performance context and support environment without needing to replicate Ben’sdata.

→ Clips 02, 06

03  Representativeness Can Be estimated by a Number.

Ben proposed that Farrow and Robertson's (2017) specificity index provides coaches with a proportional value to capture the overlap between the performance constraints observed in competition and those constraints which are designed in practice. This comparison of values can provide coaches with an understanding of how close their task designs are to the constraints of competition. Ben's publication extends this idea using Association Rule Mining — a machine learning method that identifies how frequently constraints appear together and compares that co-occurrence pattern between training and competition. The result: a quantified discrepancy that tells you not just which constraints are missing, but which combinations of interacting constraints are missing. As Ben puts it, a single missed kick in a sport like AFL is almost never about one constraint — it's usually about being tackled, in a specific pitch zone, having just received the ball, in under a second. Failing to consider any of those important task constraints impinging on player performance at that moment and a coach may have only partially replicated the performance challenge in practice.

→ Clip 03

04  Constraint Interaction Is the Unit of Analysis. Not Individual Constraints.

One of the most common errors in constraint manipulation is univariate thinking: change one thing, observe the effect. But constraints interact, as Keith Davids has highlighted from Karl Newell's original model. Ben's machine learning approach — using association rules to model constraint co-occurrence — makes this interaction directly visible. Coaches could stop asking “what happens when I add pressure?” and start asking about inter-related features of performance “what happens when pressure emerges alongside changes in spatial position, time-on-ball, fatigue levels and score state?” The data analytics method of machine learning can hold those combinations of interacting constraints simultaneously in a way that a coaching eye may struggle to track across an entire session.

→ Clips 02, 03

05  Clustering Tells You Whether Your practice tasks Are Actually Different — Not Just Designed to Be.

Ben describes a counterintuitive finding from his PhD: coaches who designed visually distinctive practice tasks often produced behaviourally identical interactions. A practice activity of game  may look different on paper — different setup, different rules — but the constraint interactions emerging within them may be statistically similar. Clustering algorithms applied to event data exposed this finding. The implication: if a coach is trying to create variability across a session by designing tasks that look different, they may be unaware of the effects of similar interacting constraints within them. The data determine whether the athletes are actually encountering different task constraints, not the practice  design manual.

→ Clip 04

06  Overloading Produces Two Solutions. Only One of Them Trains What You Want.

Overloading involves imbalanced numbers involved in team practices — say, four attackers against three defenders. These tasks are often designed to see how sub-group within a team exploits the availability of an extra player in a phase of play. Ben's overload  study at the Western Bulldogs found that when AFL teams had a numerical advantage, players converged on one of two strategies: exploit immediately (find the extra player, finish fast, low total disposals) or circulate (pass around, probe, eventually work through a defensive organisation). In a game context, the numerical advantage is often brief — a recovering defender closes it down within seconds. If practice tasks are training the second strategy, you may be conditioning exactly the behaviour that loses the contest. The data made this visible. The coaching response was analysed as a time constraint on the practice task — you must finish the rep in X seconds — which reshaped the team's solution selection.

→ Clip 05

07  Problem-Driven Data. Not Data-Driven Solutions.

Ben phrases this as the single most important mindset shift in sports analytics: “data-driven solutions” is backwards thinking. It starts with data and then constructs problems to match it. Problem-driven data starts with coaching questions — what do we need to improve, decide, or understand? — and then asks what data can help to respond to these questions. This distinction matters practically because “data-driven” organisations often produce more problems than they solve: more metrics, more analysis, more noise. The ecological approach aligns naturally with problem-first thinking because it already positions the performer-environment relationship as the unit of interest, not the data stream. Information derived from performance data may inform a coaching team’s work in co-designing contextualised practice tasks with individual athletes and sub-groups.

→ Clip 07

08  The Department of Methodology Is the Structural Answer to “Coach vs Data.”

Neither Keith Davids nor Ian Renshaw are arguing that coaches should become data scientists. Ben is clear on this: he didn't know how to code at the start of his PhD. The solution isn't to up-skill every coach in machine learning — it's to build a support structure where specialists contribute through a common language. The Department of Methodology concept in Ecological Dynamics gives each specialist — physio, analyst, coach, scientist — a defined contribution to an overarching performance framework that doesn't require mastery of every other domain. The common language is applied from key concepts in Ecological Dynamics that are relevant to particular challenges, issues and problems that can be worked on in performance preparation and athlete development. The common framework that guides this collective activity of support staff is the performer-environment relationship. When that structure exists, data stops driving coaching in practice and starts informing it.

→ Clips 06, 08

09  When Coach and Data Agree, You Get the Best Decisions.

Ben references Ian Graham's How to Win the Premier League, his book on Liverpool FC's analytics operation: in recruitment, the best outcomes came when the data and the coach converged independently on the same player. Neither overruled the other; both were inputs to a shared decision. Ben argues that this principle extends beyond recruitment to every domain of high-performance practice — variability planning, practice task design, load management, constraint design. The data highlight what the coaching eye misses or cannot track at scale. The coaching eye surfaces what the data cannot model. Neither is sufficient alone. The combination integration of both influences can lead to the emergence of  optimal decisions in athlete and team support.

→ Clip 08

10  The Landscape Is About to Change. AI and Optical Tracking Will Bring Competition-Level Data Into Practice.

Ben's view of the near future: optical tracking at every training session, AI coding all events automatically, rich behavioural data as standard rather than exception. This is not in the distant future — it's starting now. The critical variable is not whether the technology arrives, but whether practitioners have the theoretical framework to enable them to use it. Using an Ecological Dynamics lens, the same data could guide precise constraints interactions and manipulation, rather than generate more physical load metrics and recruitment percentiles. Ben's publication is partly a pre-emptive argument: here are the questions we should be asking when the data arrive, and here are the methods that can help us resolve them.

→ Clips 02, 06

CLIP 01   00:01:51 → 00:05:01

The Turning Point: Watching a Drill and Doing the Maths

Ben describes the moment he pivoted from strength and conditioning into skill acquisition. His team lost their first six games. He was watching i drills and analysing the data.  He calculated that the return on a 12-week sprint program might be an improvement of  0.1–0.2 seconds gain running speed. However, if a player learns to use their perceptual skills and is able to read the play early and make better and faster decisions they could supercharge their S&C gains, getting in position even faster. That thought led Ben to focus on skill learning in his honours project, then to a PhD at Victoria University embedded with the Western Bulldogs. Ben’s story is a good one for those working in sport science support roles: asking, “what is actually limiting performance?” can lead to a journey into ecological thinking and frame research questions informed by rich data collection.

WHY IT MATTERS

Ben's background is a useful framing device: he came through the traditional S&C pathway, not from ecological dynamics theory. His switch to an ecological focus on skill learning happened through applied reasoning about what actually moves performance. Coaches with similar backgrounds will recognise the logic immediately.

BEN TEUNE  "If they read the play quicker, if they position themselves better, they're going to get way more gain in that moment than they would from a 12-week sprint program."

CLIP 02   00:09:37 → 00:13:11

The PhD: Practical Questions, Machine Learning & Constraint Interaction

Ben describes his PhD and tis focus on applying a CLA as trying to answer four practical questions to potentially enhance the Western Bulldogs performances. How can we quantify constraints better? How can machine learning model constraint interaction? How long should we run a drill? And what does a single constraint manipulation — an ‘outnumber’ — actually change players’ perception-action skills and their ? The thread connecting all four is the same: analysing data through the theoretical lense of ecological dynamics. Perhaps, the  most useful question that Ben’s work can answer for coaches is how does changing one variable in practice impact the ‘whole system’ -in essence how does it change individual and team co-ordination?. Machine learning can track combinations that no human eye can hold across an entire session.

WHY IT MATTERS

The PhD's design mirrors the coaching problem: you're embedded, you have data, and practitioners are asking you practical questions. The methods Ben applied are not exotic — they're tools (time-series analysis, clustering, association rules) already in use in commercial analytics. What's new is applying them with a skill acquisition lens.

BEN TEUNE  "If you're trying to implement CLA or thinking about constraints, the big challenge is: there are so many constraints and they're all interacting. How am I going to know, if I change this, what the outcome is?"

CLIP 03   00:17:37 → 00:21:18

Giving Representativeness a Number: Specificity Index + Association Rules

Ben explains how Farrow and Robertson's (2017) specificity index — a percentage comparison between competition and training constraints — can be extended using Association Rule Mining. Ben is addressing one of the most important questions for coaching – what transfers from the practice ground to the competition? The original index looks at individual constraints. Ben's extension looks at constraint co-occurrence: how often do these constraints appear together in competition, and how often does that combination appear in training? The result is a discrepancy score that tells you how closely a practice task preserves the perception-action coupling of the actual performance environment and consequently, which constraint interactions are missing from practice, not just which individual constraints. Ian connects this to the red-amber-green zones in the representativeness dial from Renshaw, Davids, Newcombe & Roberts (2019) book — Ben's approach gives that dial a quantified basis rather than a coach's estimate. The key point: looking at one constraint at a time is failing to understand that constraints are always interacting and leads to “missing half the picture.”

WHY IT MATTERS

This is the most practically actionable method in the paper for coaches who have event data. The RLD dial is already a known concept in CLA communities. The extension to constraint co-occurrence is new and directly addresses the multivariate nature of real performance situations.

BEN TEUNE  "It's not like you often think we miss the kicks when they're being tackled — it's when they're being tackled and they're in this position on the field and they've also just received the ball in under one second. If you only put one of those constraints into your practice task, you're missing the rest of the context."

CLIP 04   00:27:25 → 00:29:04

Functional Variability by Cluster: Which of Your Drills Are Actually Different?

Ben describes how clustering algorithms can classify drills by variability — grouping them by the behavioural similarity of the interactions happening within them, not by their visual design. A high-variability drill and a low-variability drill might look quite different on paper but produce statistically similar constraint interactions in practice. The data method exposes this. Ben frames it as a task selection problem: when coaches are building a session, they want to know which drills are genuinely different from each other so coaches can control the inter-task variability deliberately — choosing high-variability drills when you want exploration and lower-variability when you want consolidation. Ian extends this to small-sided game design for juniors and Rian Crowther's spin bowling work.

WHY IT MATTERS

Session planning is currently a largely intuitive process. This method turns it into a data-informed one. Coaches who believe they've designed variety into a session may find that their players are encountering the same core interactions repeatedly — which either validates the choice or exposes a design gap.

BEN TEUNE  "Even though you've got different drills, are the interactions that athletes are having actually variable — or are they doing the same actions? Even though you set the drill up differently, the actions are still the same."

CLIP 05   00:34:33 → 00:36:57

The Outnumber Problem: Two Solutions, One Right Answer

Ben describes his final PhD study: measuring what players do differently - in a common practice task used by many invasion game coaches - they create  games where one team is given a numerical advantage. Ben found that at the Bulldogs, when a team were given an extra player, it produced two distinct strategies — either the team exploited the opportunity immediately (fast, few disposals, drill finished quickly) or they circulated (more passes, more time, gradual progression through the problem). However, as Ben noted, in a real game, the advantage is fleeting — a recovering defender closes it within seconds. The coaching implication is clear: if the players are selecting the second strategy more often in training, they are being conditioned to adopt a strategy that fails to exploit the temporary advantage in competition. The solution Ben identifies is a time constraint on the drill: players must complete the task in under X seconds, inviting players to find the fast solution.

WHY IT MATTERS

This is a clean example of data informing constraint manipulation: the data reveals the solution distribution, the coach interprets it against game requirements, and the constraint change is motivated by evidence. Not guesswork. Ian's reference to Rick Shuttleworth — “if you give a team one more player, you're giving them the answer” — underlines how well this connects to existing CLA thinking.

IAN RENSHAW  "They either go bang and exploit it immediately — which you'd think is the right solution — or we just pass it around and then we lose that affordance of having the extra player."

CLIP 06   00:38:13 → 00:41:35

How a Non-Data Scientist Can Start: The Applied Path

Ian asks a direct question: what do you say to a coach who loves this idea but doesn't know what machine learning means and can't code? Ben's answer was that neither did he, at the start of his PhD. He picked it all up during the PhD because he had t as his background as an applied S&C background meant that he no prior coding experience. Ben makes a great point for coaches and sport scientists: adopting an applied first approach is an advantage, because you think about problems practically. The research paper was written specifically for sports scientists and analysts who already handle tracking and event data — it demonstrates techniques they could apply to their existing datasets without a data science background. The goal is to move them from “I have data” to “I have a theoretical and practical rationale and a method for using it.”

WHY IT MATTERS

The entry barrier to analytics is real but lower than most practitioners believe. Ben's self-taught path is an important message for those practitioners who may feel they do not have the right background for further study: curiosity and applied reasoning are more important than a statistics background. The tools are increasingly accessible, especially as AI coding assistants remove the need to write code from scratch.

BEN TEUNE  "The best thing I could do was look at a thousand kicks and then bring the coach the ones they need to see. They've got five minutes — thanks, here, look at this film."

CLIP 07   00:45:48 → 00:47:17

Problem-Driven Data vs Data-Driven Solutions

Ben names the dominant framing in sports analytics — “data-driven solutions” — and inverts it. The correct frame, he argues, is problem-driven data: start with the coaching problem and then find the data that answers it. A data-first approach creates more problems than it solves because it generates metrics that have no clear coaching question behind them. Ian connects this to Carl Marshall's work in rugby union — where physiologists were controlling how much a prop could run in the final minutes of in practice based on load quotas, not on what might be needed in the last minutes when a game was in the balance. Effectively letting the data override the need to train players to go deep in critical moments. Keith adds the ecological grounding: the problem is relational (performer-environment), not a data artefact. Analyst and coach should be converging on the same problem from different angles.

WHY IT MATTERS

This is the most quotable moment in the episode and the most transferable idea. Any practitioner who has sat in a performance meeting drowning in metrics with no clear question attached will recognise it immediately. The discipline of starting from the coaching problem is something organisations can build into their analytics workflow from day one.

BEN TEUNE  "We should have problem-driven data, not data-driven solutions. The data should be there to solve a problem that already exists."

CLIP 08   00:58:59 → 01:00:10

When Coach and Data Agree: How to Win the Premier League

Ben references Ian Graham's book on Liverpool FC's analytical operation — specifically the finding that the best recruitment decisions were made when the data and the coach arrived at the same player independently. Not one overruling the other; both converging. Ben argues this principle applies beyond recruitment to every domain: variability planning, constraint design, load decisions, drill selection. Ian connects it to the AFL draft — where historical data suggests top picks are not reliably the best long-term players, partly because the human eye and the data are often operating on different timescales and different question frames. The underlying point: neither the data nor the coach is sufficient alone. The optimum is integration, and that requires both sides to speak a common language.

WHY IT MATTERS

This is the closing argument of the episode: analytics doesn't replace coaching judgement, it complements it. The structure that makes this possible — a shared framework, a problem-first culture, and analysts who understand skill acquisition — is what Ben's research paper is ultimately trying to build toward.

BEN TEUNE  "The best decisions always got made when the data and the coach agreed — when they both independently said the same thing. That's when they ended up with their best outcomes."

CONSTRAINTS COLLECTIVE

constraintscollective.com  ·  Bridging Research & Practice