Data-Driven Horse Racing Analysis
HorsePlay:
Where Probability
Meets the Races
Quantitative betting intelligence for South African flat racing — not just tips, but a full portfolio with every bet sized and justified.
What Is HorsePlay?
HorsePlay is a sophisticated horse racing prediction and betting analysis system built to bring institutional-grade quantitative methods to serious recreational and semi-professional punters. It covers South African (SAF) flat racing, with a model tuned to the market’s distinct conditions, data patterns, and racing calendar.
The system grew out of a personal project called horseplay-php and has since evolved into horseplay-cogent: a full Python/Django backend that ingests runner data, engineers over 80 predictive features per runner, scores every horse through a machine learning ensemble, and produces a rated betting portfolio — complete with recommended stake sizes — before the gates open.
Think of it as Bloomberg for horse racing. The output isn’t a hunch or a hot tip — it’s a structured, mathematically defensible view of each race, expressed as win probabilities, expected value scores, and risk-adjusted bet recommendations.
How It Works
Feature Engineering
Each runner is profiled across more than 80 engineered features: speed figures, recent form, distance suitability, going preference, jockey and trainer strike rates, draw bias, Markov state sequences, HMM-derived momentum, and more — all normalised and imputed for the model.
ML Ensemble Scoring
A blended LightGBM + Logistic Regression ensemble — trained and validated on SAF racing — scores every runner and outputs calibrated win probabilities, using SHAP-selected features for maximum signal density.
14-Factor Confidence
Raw probabilities are adjusted by 14 independent confidence factors: venue, race slot, jockey, trainer, season, trainer×jockey combo, trainer×venue, jockey×venue, distance, draw, trainer streak, jockey switch, class change (ΔMR), and a Bayesian venue–slot Markov momentum score.
Portfolio Construction
Kelly criterion-based stake sizing converts confidence-adjusted probabilities into a structured betting portfolio. Win bets, exactas, trifectas, quartets, swingers, and any-to-come multi-race chains are all priced and ranked by expected value and Sharpe ratio.
What You Get
- ● Win probability scores for every runner in a race, calibrated by a blended LightGBM + Logistic Regression ensemble
- ● Rated betting portfolios with mathematically sized stakes — not arbitrary unit recommendations
- ● Exotic bet combinations — swingers, exactas, trifectas, quartets, and any-to-come chains priced using Harville-Lo-Bacon-Shone mathematics
- ● Indicative average return and Sharpe ratio per bet — know not just what to bet, but whether the risk is worth taking
- ● 14-factor confidence-adjusted recommendations accounting for venue, race slot, jockey, trainer, seasonal patterns, connection combos, distance, draw, and Markov momentum
- ● Affordable daily access — a Day Pass (R15) for full single-day coverage or a Weekend Pass (R25) for Friday–Sunday racing, via secure PayPal checkout
Why HorsePlay?
Portfolio Thinking, Not Gut Feeling
Most tipster services hand you a selection and walk away. HorsePlay treats your bankroll as a portfolio — every bet is sized relative to your edge and the inherent uncertainty of the outcome, following the Kelly criterion used by professional investors. R1 equals one TAB bet unit.
Confidence That Adapts to Reality
A 14-factor confidence multiplier — venue, race slot, jockey, trainer, seasonal calibration, trainer×jockey combo, trainer×venue, jockey×venue, distance suitability, draw bias, trainer streak, jockey switch, class change, and Bayesian Markov momentum — is applied to every recommendation. All factors are Bayesian-shrunk toward the mean, so small samples don’t distort signals.
Built for South African Racing
HorsePlay is purpose-built for SAF racing — its track profiles, data availability, and seasonal patterns — rather than a generic model stretched across incompatible conditions.
The Numbers
Model performance metrics at a glance. AUC (Area Under the ROC Curve) measures how well the model discriminates winners from non-winners — 0.5 is random chance, 1.0 is perfect.
| Metric | SAF Region |
|---|---|
| Model AUC Score | 0.8189 |
| Classification Quality | Excellent |
| Features Engineered | 80+ |
| Model Input Features | 12 (SHAP-selected) |
| Confidence Factors | 14 — Venue · Slot · Jockey · Trainer · Season · T×J · T×V · J×V · Distance · Draw · Streak · Switch · ΔMR · Markov |
| Exotic Bet Pricing | Lo-Bacon-Shone / Harville · R1 = 1 TAB unit |
AUC benchmarks: >0.85 = Excellent | 0.75–0.85 = Good | 0.65–0.75 = Fair | 0.5 = Random chance
Ready to Back Your Bets
with Genuine Data?
HorsePlay isn’t about finding every winner — it’s about finding every edge and making sure your bankroll reflects it. If you’re serious about racing and tired of guesswork dressed up as analysis, this is the system built for you.
Day Pass R15 · Weekend Pass R25 · Secure PayPal checkout
Covering SAF flat racing · Powered by LightGBM + Logistic Regression ensemble · 14-factor confidence system