Huggy Turf — Baseline Bernoulli (Winner Prediction)
Baseline shipped as a taste of what the full Huggy Database unlocks: 1,128,888 races across 44 countries, including 426,963 French races. Serious backtests need the full archive — see pricing below.
🇬🇧 English
Baseline winner-prediction model for French horse races, built on the Huggy Database — French Horse Races dataset. Meant as a reference point for evaluating more advanced models (XGBoost, LightGBM, neural networks).
Approach
Bernoulli logistic regression on final PMU odds plus 8 form features (recent finishes, jockey_win_rate_90d, trainer_win_rate_90d, draw, normalized distance, weight, age, sex).
Performance (Q4 2025 validation)
| Metric | Baseline | Public odds |
|---|---|---|
| Log loss | 0.298 | 0.312 |
| Top-1 accuracy | 22.4% | 21.1% |
| Flat ROI (1€ stake) | -8.2% | -14.5% |
Usage
import joblib, pandas as pd
model = joblib.load("baseline_bernoulli.joblib")
X = pd.read_parquet("runners_upcoming.parquet")
probs = model.predict_proba(X)[:, 1]
Limitations
- Not calibrated for real betting.
- No performance guarantee — research use only.
- The full dataset requires a commercial license — see https://huggydatabase.com/.
Get the full dataset
| Tier | Price | Contents |
|---|---|---|
| R&D Pack | 2 950€ | 2 years sample · full schema |
| Full Archive | 9 500€ | 1,128,888 races · 44 countries · 1996-2026 · PMU/PMH spread · 40+ vars |
| Reasoning Engine | 45 000€+ | Source code · Monte-Carlo · Query Builder |
Order: https://huggydatabase.com/ · contact@huggydatabase.com
🇫🇷 Français
Modèle baseline de prédiction du gagnant d'une course hippique française à partir du dataset Huggy Database — French Horse Races. L'archive complète réunit 1 128 888 courses dans 44 pays, dont 426 963 courses françaises. Ce modèle sert de point de comparaison pour évaluer des modèles plus avancés (XGBoost, LightGBM, réseaux de neurones).
Approche
Régression logistique Bernoulli sur les cotes PMU finales et 8 features de forme (musique, jockey_win_rate_90j, entraineur_win_rate_90j, corde, distance normalisée, poids, âge, sexe).
Performance (validation Q4 2025)
| Métrique | Baseline | Cote publique |
|---|---|---|
| Log loss | 0.298 | 0.312 |
| Top-1 accuracy | 22.4% | 21.1% |
| ROI flat (mise 1€) | -8.2% | -14.5% |
Limites
- Modèle non calibré pour paris réels.
- Aucune garantie de performance : à des fins de recherche uniquement.
- Le dataset complet nécessite une licence commerciale — voir https://huggydatabase.com/.
Obtenir le dataset complet
| Offre | Prix | Contenu |
|---|---|---|
| R&D Pack | 2 950€ | 2 ans d'échantillon · schéma complet |
| Full Archive | 9 500€ | 1 128 888 courses · 44 pays · 1996-2026 · spread PMU/PMH · 40+ variables |
| Reasoning Engine | 45 000€+ | Code source · Monte-Carlo · Query Builder |
Commander : https://huggydatabase.com/ · contact@huggydatabase.com
Links / Liens
- Dataset : https://huggingface.co/datasets/huggy-engine/horse-racing-database-fr
- Space demo : https://huggingface.co/spaces/huggy-engine/huggy-database-query-builder
- Site : https://huggydatabase.com/
🏆 Huggy Quinté Challenge — 23 August 2026
An open prediction benchmark on Kaggle: forecast the finishing order of the top 5 horses in French Quinté+ races. Train on 3,030 historical quintés, climb the public leaderboard on 854 held-out races, and settle it on one real race nobody has the answer to — the Quinté+ of Sunday 23 August 2026. Winner gets a Huggy R&D Pack licence (2,950 € value).
Un benchmark de prédiction ouvert sur Kaggle : trouver l'ordre d'arrivée des 5 premiers chevaux. Finale sur le Quinté+ du dimanche 23 août 2026. Vainqueur : une licence Pack R&D (valeur 2 950 €). Benchmark de prédiction, pas un concours de paris.
- Competition — https://www.kaggle.com/competitions/huggy-quinte-challenge
- Starter data — https://www.kaggle.com/datasets/huggydatabase/huggy-quinte-challenge
- Starter notebook — https://www.kaggle.com/code/huggydatabase/huggy-quinte-challenge-starter
🔗 Huggy ecosystem / Écosystème Huggy
The same project is mirrored on several platforms. Everything points back to the canonical source, huggydatabase.com.
| Where | Link |
|---|---|
| 🌐 Official site — full archive, pricing, Query Builder | https://huggydatabase.com/ |
| 📄 Methodology article (Bernoulli, Poisson, calibration, ROI) | https://huggydatabase.com/methode |
| 🤗 Hugging Face — dataset | https://huggingface.co/datasets/huggy-engine/horse-racing-database-fr |
| 🤗 Hugging Face — baseline model (you are here) | https://huggingface.co/huggy-engine/turf-baseline-bernoulli |
| 🤗 Hugging Face — Query Builder Space | https://huggingface.co/spaces/huggy-engine/huggy-database-query-builder |
| 📊 Kaggle — dataset (free sample) | https://www.kaggle.com/datasets/huggydatabase/french-horse-races-1996-2026 |
| 📓 Kaggle — starter EDA notebook | https://www.kaggle.com/code/huggydatabase/french-horse-racing-pmu-eda-starter |
Kaggle and Hugging Face host the same free sample; the full 1,128,888-race, 44-country archive lives at huggydatabase.com — contact@huggydatabase.com.
Kaggle et Hugging Face hébergent le même échantillon gratuit ; l'archive complète de 1 128 888 courses dans 44 pays est sur huggydatabase.com — contact@huggydatabase.com.
- Downloads last month
- -