README.md

---
license: cc-by-4.0
language:
- it
- en
pretty_name: Sport Intelligence Benchmark
tags:
- benchmark
- reproducible-research
- zenodo
- sports-analytics
- brier-score
- lightgbm
- synthetic-dataset
---

# Sport Intelligence Benchmark

Ensemble V2 LightGBM meta-learner methodology and a synthetic reproduction dataset for football match outcome prediction, reported as two distinct, deliberately never-reconciled results (D-07). Reproducible: `python run_benchmark.py --seed 42` against the shipped synthetic sample (`data/derived_sample.csv`, itself generated via `python export_dataset.py --rows 4000 --seed 42 --output data/derived_sample.csv`) prints a Brier score (3-class summed, range 0.0-2.0) that any third party can verify end to end. Production: Brier 0.5783 on the real 97,000-match production corpus — this figure is measured on non-redistributable data and cannot be verified by third parties; it is reported for transparency, not for independent reproduction. See DATA_PROVENANCE.md for the full reconciliation. Code: MIT. Contents of `data/`: CC-BY-4.0.

> Metodologia del meta-learner ensemble V2 LightGBM e un dataset sintetico di riproduzione per la previsione degli esiti di partite di calcio. Il risultato di produzione (Brier 0.5783 su circa 97.000 partite reali) e il risultato sintetico riproducibile sono dichiarati come due numeri distinti, mai riconciliati.

This is a **self-deposited** research artifact — an author-submitted Zenodo record, not a publication reviewed by an independent third party before release.

**Synthetic sample disclosure:** the file(s) under `data/` in this mirror are a fully **synthetic** reproduction sample (see `DATA_PROVENANCE.md` in the Files below) — not a copy, subset, or export of the production dataset. The headline production figure (Brier 0.5783 on ~97,000 real matches) is reported here for transparency but is measured on non-redistributable data and **cannot be verified by third parties** — only the synthetic in-sample number is independently reproducible from what this repository ships (`python run_benchmark.py --seed 42`).

## Source & Attribution

- **Canonical page (IT):** https://federicocalo.dev/blog/sport-intelligence-benchmark-97k-partite
- **Canonical page (EN):** https://federicocalo.dev/en/blog/sport-intelligence-benchmark-97k-matches
- **Author:** Federico Calò — https://federicocalo.dev (Wikidata Q139562320, ORCID 0009-0004-4102-281X)
- **Licence:** This Hugging Face front-matter declares `cc-by-4.0` — the licence of the **data** you download here (this repository's dual licence: code MIT, data CC-BY-4.0). The corresponding Zenodo deposit's structured `license` field instead reports `mit-license`, because Zenodo's structured field records the **code** licence; the data licence lives in prose there (`LICENSE`, `README.md`, and the data-directory `LICENSE`). This divergence between the two records is deliberate, not an inconsistency — on Hugging Face, what a user actually downloads is the data, so the front-matter reflects the data licence.
- **Methodology:** see `DATA_PROVENANCE.md` in the Files below

## Files

- `README.md` — this file (the Hugging Face dataset card)
- `CITATION.cff`
- `DATA_PROVENANCE.md`
- `LICENSE`
- `REPO_README.md`
- `data/LICENSE`
- `data/derived_sample.csv`
- `export_dataset.py`
- `requirements.txt`
- `run_benchmark.py`
- `tests/test_reproduce.py`

## Citation

Zenodo is the DOI **of record** for this artifact — cite the concept DOI below, not this Hugging Face page (Hugging Face mints one DOI per revision, with the previous one becoming outdated, so it has no stable identifier semantics):

```
Federico Calò, "Sport Intelligence Benchmark", Zenodo, 2026. https://doi.org/10.5281/zenodo.21602378
```

- **Concept DOI (stable, cite this):** https://doi.org/10.5281/zenodo.21602378
- **Source code / machine-readable `CITATION.cff`:** https://github.com/fedcal/sport-intelligence-benchmark