The Science Behind RaceAlpha
Discover how our advanced AI technology transforms racing data into powerful predictions.
AI Racing Intelligence Platform
Advanced machine learning algorithms process millions of data points to deliver precision-engineered racing predictions with unmatched accuracy
Data Ingestion
Multi-Source Collection
Continuous aggregation of race data, form guides, and historical performance from official racing databases
Neural Processing
Deep Learning Engine
XGBoost ensemble models analyse complex patterns across years of historical race outcomes
Pattern Analysis
Behavioral Recognition
Identifying winning combinations of jockey-horse connections and trainer performance patterns
Risk Modeling
Probability Engine
Race-level normalisation calculates calibrated win probabilities across varying field sizes
Value Detection
Opportunity Identification
Comparing calculated true probabilities against market odds to identify value opportunities
Prediction Output
Actionable Intelligence
Delivering precision-engineered predictions with confidence scores and value ratings
Live Data Pipeline
Data Ingestion
Neural Processing
Pattern Analysis
Risk Modeling
Value Detection
Prediction Output
Data Ingestion
Continuous aggregation of race data, form guides, and historical performance from official racing databases
Our AI Approach
RaceAlpha uses a proprietary multi-model ensemble architecture trained on over 160,000 race records. Our system processes 80+ carefully engineered features per runner, identifying patterns and correlations that would be impossible for human analysts to detect—while rigorously preventing data leakage for honest, production-ready predictions.
Ensemble Models
Three specialised XGBoost models work together—predicting win probability, top-3 chances, and finishing position—then combine their insights for robust predictions.
Dynamic Ratings
Our proprietary ELO rating system tracks over 27,000 horses, updating after every race to capture true form and ability in real-time.
Field-Size Optimization
Specialised sub-models trained on small, medium, and large field sizes deliver calibrated predictions regardless of runner count.
The Process
Why It Works
Most "AI" tipsters use basic statistics or leaky models that look good in backtests but fail in production. RaceAlpha is built differently—engineered for real-world deployment with rigorous temporal validation and honest performance metrics.
160,000+ Race Records
Our models train on a comprehensive dataset spanning years of racing history, with 80+ engineered features per runner that capture everything from career stats to jockey-trainer synergies.
27,000+ Horse ELO Ratings
Our proprietary rating system tracks every active horse with ratings that update after every race—giving us a real-time view of true ability that static databases can't match.
Production-Honest Models
We rigorously prevent data leakage—rejecting features that contain post-race information. Our temporal validation ensures we measure true predictive power, not pattern memorisation.
Value-First Philosophy
We don't just pick winners—we calculate true probabilities and compare to market odds. Our selections target positive expected value, where our edge compounds over time.
The result? A verified +3.7% ROI (2,549 Bets) — $1,688 profit, every result tracked live. All-time includes strong early results; 12-month rolling ROI ~4% with ongoing refinement.
Betting Unit Calculator
Enter your bankroll
Your RaceAlpha unit value
Recommended units: 250
This calculator helps you determine your unit size for Hong Kong racing. RaceAlpha recommends a standard 250 units for Hong Kong racing bankroll management, with each unit representing a small percentage of your total bankroll.
Frequently Asked Questions
Frequently Asked Questions
How accurate are RaceAlpha predictions?
Short Answer:Our ensemble model has delivered a verified +3.7% ROI (2,549 Bets) in Australian and Hong Kong racing. Our win model targets 8-15% accuracy (realistic for horse racing), while our top-3 model achieves 20-35% accuracy. We use time-based train/test splits to measure true predictive power—not pattern memorisation.
How often is the AI model updated?
Short Answer:Our horse ELO ratings update after every race meeting, ensuring real-time form assessment across 27,000+ tracked horses. The core prediction models undergo full retraining quarterly with hyperparameter optimisation, while our feature engineering pipeline continuously incorporates new data from our 160,000+ race record database.
What data sources does RaceAlpha use?
Short Answer:We engineer 80+ features from comprehensive racing databases including career win/place percentages, jockey-horse connection histories, trainer patterns, barrier statistics, weight differentials, and our proprietary ELO ratings (600-1934 scale). Every feature is validated against our automated leakage detection system.
How does RaceAlpha compare to traditional handicapping?
Short Answer:Most tipsters use basic statistics or 'AI' models with data leakage that look great in backtests but fail live. RaceAlpha uses a three-model ensemble architecture with temporal validation, automated leakage detection, and race-level probability normalisation. We measure true out-of-sample performance.
What is the ELO rating system?
Short Answer:Inspired by chess rankings, our dynamic ELO system rates every horse on a 600-1934 scale based on performance. Ratings update after each race, with gains/losses proportional to the quality of opposition beaten. This gives us a real-time view of true ability—not static historical averages.
How do you calculate win vs place probabilities?
Short Answer:We use a unified position-based probability system. Our Position Model predicts where each horse will finish, then we calculate win probability (P(position ≤ 1)) and top-3 probability (P(position ≤ 3)) using statistical distributions. This ensures P(top3) is always greater than P(win)—a constraint many prediction systems violate.
What prevents your models from overfitting?
Short Answer:Five layers of protection: (1) Automated leakage detection rejects features only available post-race, (2) Temporal train/test splits prevent future data leaking back, (3) Strong regularisation (L1/L2) prevents memorisation, (4) Shallow trees (max depth 4-6) limit complexity, and (5) We monitor train/validation gaps, rejecting models with >20% divergence.
What are connection features?
Short Answer:Connection features capture synergies between horses, jockeys, and trainers. We track how many races a jockey-horse pair have run together, their combined win rate, and whether there's a 'strong connection' (3+ races). Similarly for jockey-trainer partnerships (5+ races). These features often reveal hidden value the market misses.
How do field-size models work?
Short Answer:Race dynamics differ significantly between small fields (2-8 runners) and large fields (15+ runners). Our ensemble uses specialised sub-models trained on each field-size bracket, then selects the appropriate model at prediction time. This improves calibration by 0.3-0.5 positions compared to a single universal model.
What does "value betting" mean?
Short Answer:Value betting means finding horses where our calculated probability exceeds what the market odds imply. If we calculate a 15% win chance but the odds suggest only 10%, that's positive expected value (+EV). Over hundreds of bets, consistently finding +EV opportunities compounds into long-term profit—hence our verified +3.7% ROI (2,549 Bets).
Ready to experience the power of AI?
Join RaceAlpha today and start making data-driven decisions.