sports-trends-dataset

TL;DR — A continuously-updated, medallion-architecture data lake for football, basketball, tennis and cricket: immutable raw ingests, cleaned/standardized layers, an engineered feature store, and ready-to-train chronological splits in partitioned Parquet. Built to be leakage-safe and reproducible.

Pairs with thesports-trends-models repo.

What's inside (medallion layout)

Large data never lives in GitHub — it lives here, partitioned by sport / date / layer.

raw/<sport>/provider=<p>/date=YYYY-MM-DD/fixtures.json   # immutable ingest (source of truth)
bronze/  silver/                                          # cleaned + standardized records
gold/
  features/<feature>/sport=<sport>/date=.../part.parquet  # engineered feature store
  training/<sport>/{train,validation,test}.parquet        # chronological, leakage-safe splits
  inference/tomorrow/date=YYYY-MM-DD/inference_input.parquet
quality/{leakage_report,schema_validation,...}.json       # automated data-quality reports
registry/{dataset_manifest,latest_versions}.json          # versioning + production pointers

• Raw is append-only and immutable — full provenance back to the source API.
• Bronze/Silver normalize to one canonical schema with stable IDs and de-duplication.
• Gold is analytics/ML-ready: a feature store plus train/validation/test splits.
• ML data is partitioned Parquet (by sport/date/layer) — never one giant CSV.

Engineered features (gold)

Every feature is computed leakage-safely — for a given fixture, only matches with date < fixture.match_date are used:

The prediction label is the realized match outcome (home/draw/away or 2-way per sport).

Data quality & leakage prevention

• Chronological splits. Training data is ordered by time — train precedes validation precedes test. No random shuffling across the time boundary.
• Leakage assertions. Automated checks (quality/leakage_report.json) verify feature/label disjointness and chronological ordering; a regression test plants a future blowout and confirms pre-match features are unchanged.
• Schema validation. Records are validated against the canonical schema (quality/schema_validation.json).
• Versioned manifests. registry/dataset_manifest.json records version, timestamp, sports covered, layers present, and quality counts.

Load the data

import pandas as pd
from huggingface_hub import hf_hub_download

REPO = "ruslanmv/sports-trends-dataset"

# Ready-to-train football split (chronological, leakage-safe)
train = pd.read_parquet(hf_hub_download(REPO, "gold/training/football/train.parquet", repo_type="dataset"))
print(train.shape)
print(train.head())

Or stream everything for a sport with thedatasets library:

from datasets import load_dataset
ds = load_dataset("ruslanmv/sports-trends-dataset", data_files="gold/training/football/*.parquet")

Update cadence

All updates run automatically via GitHub Actions.

Sources, license & intended use

• Sources. Aggregated from free sports APIs and public-domain feeds (e.g. OpenFootball for World Cup data), normalized into a common schema. Raw partitions retain provider provenance.
• License. MIT for this dataset's structure, schema, and engineered features. Please respect the terms of the underlying upstream providers for the raw factual data.
• Intended use. Sports analytics, ML research, education, and powering the ruslanmv.com/sports-trends dashboard.
• Not betting advice. Provided for information and entertainment only.

Citation

@dataset{magana_sports_trends_dataset_2026,
  author  = {Ruslan Magana Vsevolodovna},
  title   = {Sports-Trends: A leakage-safe multi-sport match data lake},
  year    = {2026},
  url      = {https://huggingface.co/datasets/ruslanmv/sports-trends-dataset},
  note     = {Live dashboard: https://ruslanmv.com/sports-trends/}
}