An anonymized case study from a UK professional football academy environment.

Improve player availability and support longer-horizon decisions (e.g., development planning and retain/release discussions) by combining physical benchmarks with contextual performance signals. The hard part is messy data (gaps, changing definitions, small samples) and making outputs usable for non-technical stakeholders.

• Sustained >90% player availability across high-intensity competition periods via improved monitoring and decision routines.
• Drove an aggregate ~12% improvement across key physical output metrics through data-led periodisation and longitudinal tracking.
• Reduced manual data entry by ~40% by automating ingestion/ETL from multiple external data sources.

• Longitudinal load and exposure (training + match) from wearable tracking systems.
• Availability and medical context (time-loss / restricted training windows).
• Tactical context signals (e.g., xG/xT/VAEP-style aggregates) used as covariates, not as the sole driver.
• Benchmarks and standards by position/age band.

• SQL/Postgres for modeling-ready tables and longitudinal views
• Python/R for feature engineering, modeling, and statistical evaluation
• Dashboards/reports: Shiny/Dash-style reporting surfaces for staff-facing outputs

• KPI redesign: defined actionable "availability" features to replace legacy, fragile KPIs.
• Feature engineering: acute vs chronic exposure, trend breaks, monotonic constraints where appropriate, and standardized benchmarking (percentiles / z-scores) for comparability.
• Modeling: started with transparent baselines (rules/linear/regularized) before evaluating non-linear alternatives.
• Validation: time-aware splits and leakage control; calibration checks to make probability outputs decision-ready.

• Retain/release: combined tactical-context aggregates with physical benchmarks to support objective conversations.
• Load management: translated model outputs into staff-friendly decision cards (what changed, why it matters, what to do next).

• Designed for sensitive, human-centric data: access control, redaction-by-default reporting, and audit-friendly outputs.

flowchart LR
  A[Wearables + availability records + context metrics] --> B[Ingest + normalize]
  B --> C[Feature engineering]
  C --> D[Baselines + candidate models]
  D --> E[Time-aware validation + calibration]
  E --> F[Decision cards + dashboards]

• A decision-support view that combines (1) availability risk/likelihood, (2) drivers/attribution at feature-group level, and (3) suggested actions (load management, monitoring focus, or further review).
• A benchmarking layer for recruitment/development conversations (transparent percentiles and distributions).

Example reporting outputs (anonymized visuals):

• Benchmarking to determine asymmetry status: assets/benchmark_asymmetry.png
• Game/week difficulty context: assets/game_difficulty.png
• HR-integrated GPS training load monitoring: assets/hr_integrated_training_load.png
• Sprint exposure monitoring: assets/sprint_exposure.png
• Benchmarking cardiovascular capacities: assets/benchmark_cardio_capacity.png
• Standardised per-minute match report (apples-to-apples): assets/per_minute_match_report.png

Code, raw data, and internal identifiers are private. This repository documents the methodology, validation approach, and example artifacts without exposing proprietary details.