Tech & Innovation: AI and Machine Learning, Transforming How Racing is Analysed and Predicted

Artificial intelligence in horse racing is not a single technology. It is an umbrella term for a range of computational approaches, machine learning, predictive modelling, natural language generation, neural networks, that share the capacity to process volumes of data and identify patterns at speeds and scales beyond human capability. The sport’s relationship with these tools has evolved rapidly over the past decade, from a niche area of interest to a mainstream consideration affecting how races are analysed, how horses are trained and how bets are placed.

From Intuition to Data

Traditional horse racing analysis was built on human expertise: the trainer who knew their horse better than any instrument could measure; the punter who studied form guides and watched replays; the journalist who accumulated a lifetime’s observation into editorial judgment. These skills have not been rendered obsolete by AI, the best form analysts and trainers still operate with tacit knowledge that data systems struggle to capture. But the analytical landscape has changed significantly.

The core advantage AI brings to racing analysis is the ability to process historical data at scale and identify non-obvious relationships. A human analyst reviewing the last 100 Cheltenham Gold Cup races can extract meaningful patterns, age profile, going preferences, trainer statistics, previous course form. An AI model can process the last 1,000 races, cross-reference going, weight carried, trainer, jockey, breeding, margin of victory, subsequent career, market movement and dozens of other variables simultaneously, identifying correlations that would take a human analyst months of work to uncover manually.

EquinEdge, a US-based AI handicapping platform, describes this capability as the ability to “dig into hidden patterns that people just can’t see.” The platform processes historical race outcomes, speed figures, consistency metrics, jockey-trainer combinations, track conditions and market odds to generate implied probability distributions for each runner. The self-learning aspect, whereby the model updates continuously as new race data arrives, means it adapts to changing racing patterns in ways that static analytical systems cannot.

What AI Analyses

Race outcome prediction: The most widely publicised application of AI in racing is outcome prediction, assigning probabilities to each runner in a given race based on historical patterns. The accuracy of these systems varies; racing remains a sport with a high degree of genuine randomness, and no model has consistently beaten the market with statistical significance over large sample sizes. However, the best systems have demonstrated an ability to identify mispriced horses, runners whose AI-assessed probability of winning differs meaningfully from the market odds, often enough to be commercially valuable.

Training optimisation: AI analysis of training data, combining GPS workout records, biometric readings, historical form and vet reports, allows training yards to optimise exercise intensity, identify horses who are underperforming in training (and might need rest) or overperforming relative to current fitness level. Paulick Report’s 2024 analysis of AI in training noted that “AI algorithms are already being used to analyse vast amounts of race data, identify patterns, and predict outcomes with remarkable accuracy.”

Market analysis: Sophisticated betting operations use AI to monitor odds movement across multiple bookmakers simultaneously, identifying patterns in how markets adjust to incoming information. Sharp money (large bets placed by professional bettors with strong analytical models) tends to move markets in ways that differ from public money; AI systems can identify these signatures and adjust positioning accordingly.

Content generation: Natural language AI is already used in racing journalism for automated race result reports, form summaries and routine content. This application is most mature in American racing, where the volume of daily race meetings exceeds any reasonable human content production capacity.

The Machine Learning Proof of Concept

One well-documented experiment illustrates AI’s capabilities in racing. A test conducted using Akkio’s no-code machine learning platform, documented by IEEE Spectrum, involved training a model on 700 historical race records from Saratoga and deploying it to predict the winner of 10 live races at Del Mar. The model correctly identified the winner in 6 of 9 races where the top-probability horse ran. Had bets been placed on the highest-probability selection in each race, the return on investment was 140%.

This single experiment does not establish that AI systematically beats racing markets, sample sizes are too small and conditions too specific. But it demonstrates that machine learning can identify patterns in racing data that translate to commercially useful predictions, at least in specific conditions, with modest training data. The question facing the industry is how this scales with larger data sets, more sophisticated models and longer track records.

Challenges and Limits

AI in racing faces several genuine constraints that prevent it from being the straightforward advantage it might initially appear.

Data quality: AI models are only as good as their training data. Racing data is notoriously variable in quality, inconsistent going descriptions across courses and eras, missing information about training conditions, and the inherent incompleteness of any available dataset.

Market efficiency: Betting markets aggregate the expectations of millions of participants, including some of the most sophisticated analytical operations in the world. Consistently finding information that the market has not already priced in is difficult; as more AI systems converge on similar methodologies, the edge they generate may diminish.

Unpredictability: Racing’s fundamental character, horses are not machines; races involve luck in running, interference, falls and unforeseeable events, limits the accuracy ceiling of any predictive system. A 65% strike rate for AI-identified winners would be extraordinary; it would also mean 35% of selections lose.

Ethical considerations: The use of AI for fraud detection (identifying suspicious market patterns) is broadly welcomed by regulators. The use of AI for market manipulation is not. The line between sophisticated analysis and market abuse is not always clear in practice.