Capstone Project: Manager Impact in Football
A three-session team project from data collection to causal analysis
Capstone Project
Manager Impact in Football — from raw data to causal claims in three sessions
The Project
After ten weeks of guided exercises, the capstone project is a three-session team project with one broad research question, messy real-world data, and AI as a teammate.
Research question: Does changing a manager improve team performance? And does the impact vary by manager and team characteristics?
Your audience is heads of data science at sports clubs with a strong background in econometrics. Your goal is to produce a rigorous, reproducible analysis that sheds light on the true impact of managerial changes in football.
You will make decisions, be assertive and transparent.
Why this project
- It is specific (one domain, concrete RQ) but broad (no exact steps given to you).
- You will have to plan, layer tasks with AI, and decide — not just follow a recipe.
- Before you start, we’ll go over the companion page Designing Larger Analytics Projects with AI — it covers the mindsets, self-interview,
agents.mdfile, and the three kinds of tests you’ll use all three sessions.
Task and Scope
Your task is to answer the research question above by collecting, combining, and analyzing data on football matches and manager changes. There will be 3-member teams, each picking a country.
| Dimension | Choice |
|---|---|
| Sport | Football (soccer) |
| League tier | First division only |
| Countries (pick one per team) | Spain, Italy, France, Germany, Turkey, Scotland, Portugal, Netherlands, Poland, Ukraine, Russia, or another you can defend |
| Time horizon | 10+ seasons of historical data |
| Session 1, 2 deliverable | Data, scripts, reproducible repo and READMEs |
| Session 3 | Full project repo + ~12-minute presentation |
Key questions to answer by Session 3
- What is the average effect of a manager change on team performance?
- Which types of managers show larger or smaller effects?
- Which types of teams respond more?
- When a dismissal is more expected as shown by news reports, is the effect any different?
Open questions
- Anything not specified is open. You can consult each other and AI, and the team shall make decisions.
Sessions at a glance
| Session | Focus | Deliverable by Sunday 23:55 |
|---|---|---|
| 1 | Data collection, combination, description | Documented dataset + QA notes |
| 2 | Text → expectations (APIs, scraping) | Article corpus + per-change expectation score |
| 3 | DiD analysis + heterogeneity + presentation | Results, slides, repo |
Session structure (same each time)
Each 200-minute session runs the same shape:
- Intro talk (≈30 min) — key concepts, common pitfalls, decisions to make.
- Team work (≈120 min) — you execute; AI assists; I circulate.
- Group discussion (≈50 min) — share what worked, compare approaches, debrief.
Evaluation
You will be evaluated on:
- Data quality — required elements collected, documented, tested.
- News classification — coverage of changes, validated classification, reproducible prompt.
- Causal identification — appropriate method, assumptions stated and checked where possible.
- Heterogeneity & expectations — meaningful comparisons, expectations-vs-reality done rigorously.
- Reproducibility — somebody else can rerun.
- Presentation — clarity for a peer audience, honest about limits.