Sports Analytic Skills

Site: walrusquant.github.io/sports-analytic-skills · Docs: Getting started

Standalone agent skills for sports analytics and modeling.

Install the skills, point your agent at your data, and use focused guidance for EDA, time-safe features, baselines, statistical and predictive models, walk-forward validation, calibration, simulation, interpretation, and honest reporting. The skills work without this repository’s Python package or pipelines.

The repository also contains an optional sports_ds toolkit for acquiring and normalizing public NFL, NBA, and MLB data. The sports-ds-bridge skill connects that toolkit to the portable artifacts consumed by the standalone skills.

Install the skills

npx skills add WalrusQuant/sports-analytic-skills

The default command lets you select skills interactively. Install one skill non-interactively:

npx skills add WalrusQuant/sports-analytic-skills --skill eda-sports -y

Install every skill for every detected agent:

npx skills add WalrusQuant/sports-analytic-skills --all

That is enough to use the skills. You do not need to clone this repository, install sports_ds, or run an end-to-end pipeline.

Ask your agent to:

• explore a team-game or player-game dataset before modeling;
• audit candidate features for decision-time leakage;
• establish constant and domain baselines;
• build a time-ordered validation design;
• assess probability calibration;
• interpret the largest misses and error slices;
• simulate outcomes from a schedule and win probabilities;
• write a results report or model card.

Each skill documents the input columns or artifact shape it needs. Bundled helpers operate on user-owned CSV, Parquet, or JSON files and public Python dependencies; they do not import sports_ds.

A composable analysis path

The skills are independently invocable. Combine only the ones the task needs:

question and decision time
  -> data acquisition
  -> EDA
  -> time-safe features / ratings
  -> baselines and candidate models
  -> time-ordered validation
  -> leakage and calibration checks
  -> interpretation, simulation, and reporting

Examples:

Use eda-sports to inspect this CSV at team-game grain. Report coverage,
missingness, duplicated game rows, target balance, and modeling red flags.

Use feature-rules and leakage-audit on this feature table. Decision time is
kickoff. Identify every column that would not have been knowable then.

Use baseline-models, predictive-modeling, and validation-design to compare a
constant baseline with logistic regression under season walk-forward folds.

Finding data

The data skills work directly with public ecosystems:

• nflreadpy for nflverse data;
• sportsdataverse-py for supported multi-sport sources;
• pybaseball for Statcast and MLB tables;
• data-sources for choosing a source and grain.

If the user wants the repository’s prebuilt loaders, normalized panels, or CLI, use sports-ds-bridge. It owns all optional sports_ds integration and hands a portable artifact back to the standalone skill.

public source -> sports_ds loader (optional) -> CSV/Parquet/JSON
                                            -> standalone skill

Optional sports_ds toolkit

The toolkit is useful when you want ready-made public-data adapters, reusable feature/model components, or reference pipelines. It is a separate opt-in runtime, not a dependency of the skills.

git clone https://github.com/WalrusQuant/sports-analytic-skills.git
cd sports-analytic-skills
python3 -m venv .venv
source .venv/bin/activate
pip install -U pip
pip install -e .

Optional multi-sport loaders:

# macOS only: XGBoost, pulled by sportsdataverse, needs OpenMP
brew install libomp

pip install -e ".[multi]"

The Homebrew command is unnecessary for skill-only use and normally unnecessary on Linux. It is only a native runtime prerequisite for the optional multi-sport toolkit path on macOS.

Discover the current toolkit surface with:

sports-ds --help
sports-ds feature-registry

Pipeline commands remain useful reference benchmarks. They are documented in sports-ds-bridge, not embedded as prerequisites throughout the generic skills.

Skills

Foundation and data

Exploration and features

Modeling and validation

Interpretation and reporting

Skill structure

skills//
  SKILL.md
  references/     # conditional detail
  scripts/        # optional standalone helpers
  agents/         # optional UI metadata

The path to a bundled script is resolved relative to its SKILL.md; instructions must not assume the user’s current directory is this repository.

Repository architecture

skills/                         primary product; standalone
  sports-ds-bridge/             optional integration boundary

src/sports_ds/                  optional Python toolkit
  data, eda, features, ...      reusable components
  pipelines/                    reference orchestration
  cli.py                        toolkit command surface

tests/                          toolkit + skill-independence tests

See ARCHITECTURE.md for dependency rules.

Design rules

1. A skill-only install works without sports_ds.
2. Generic skills consume documented user artifacts and public dependencies.
3. Only sports-ds-bridge may direct users to the optional toolkit.
4. Predictive features must be knowable at the declared decision time.
5. Time-ordered sports data uses walk-forward validation, not random shuffles by default.
6. Baselines come before model complexity.
7. Reports retain sample sizes, uncertainty, failures, and limitations.

Development

Repository contributors can install the optional toolkit and test dependencies:

python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"
pytest -q -m "not live"

Live data integrations are a separate gate because they call third-party services:

SPORTS_DS_LIVE_TESTS=1 pytest -q -rs -m live

GitHub Actions runs the offline suite on Python 3.10 through 3.13. A separate scheduled and manually triggered workflow exercises the live integrations.

Read docs/skill-authoring.md before changing a skill.

License

MIT