Improving sport horse selection using machine learning: Early prediction of elite-level show jumping performance.

Improving sport horse selection using machine learning: Early prediction of elite-level show jumping performance.

Overview of the Research

• This study investigates how machine learning, specifically the XGBoost algorithm, can be used to predict which show jumping horses will achieve elite-level performance based on early competition records.
• The goal is to identify horses with high potential before they fully mature, aiding in talent identification while considering the horse’s development and welfare.

Background and Challenges

• Show jumping performance is difficult to predict early because horses develop their abilities over time, and competition results can be highly variable.
• Elite-level success often only emerges after years of competition, making early selection of potential champions challenging.
• Traditional methods rely on simple thresholds such as the maximum height cleared by a horse but lack nuance in considering longitudinal performance data.

Data and Methodology

• Competition data were sourced from Fédération Équestre Internationale (FEI) and national-level competitions in Sweden, Belgium, and France.
• Horses with competition records before age ten were analyzed to predict if they would become “champions,” defined as those who achieved at least one clear round at 160 cm.
• “Non-champions” were horses that competed but never reached this elite-level benchmark during their career.
• The study used an Extreme Gradient Boosting model (XGBoost), a powerful machine learning algorithm well-suited for tabular data and classification tasks.
• Model training involved nested cross-validation to ensure robust performance evaluation and to avoid overfitting.
• Performance of the XGBoost model was benchmarked against simpler single-rule classifiers that classified horses solely on the basis of maximum obstacle height cleared.

Key Findings

• XGBoost significantly outperformed the simpler baseline models, with an average area under the ROC curve (AUC) of 83.5% and an average precision of 80.0%, indicating strong predictive accuracy.
• Prediction accuracy improved as more competition history was included, especially from around seven years of age onward.
• Probabilistic predictions showed increasing separation between future champions and non-champions as horses continued competing and accumulating performance data.

Interpretability and Important Predictors

• Permutation importance and SHapley Additive exPlanations (SHAP) analyses were used to interpret the model and understand which features contributed most to predictions.
• Key predictors included:
  • Obstacle difficulty—the height of jumps successfully cleared
  • Consistency in performance over competitions
  • Age-related progression, reflecting the horse’s development over time
• These insights provide transparency and can help trainers and breeders understand which early factors indicate future top-level potential.

Implications and Conclusions

• The study demonstrates that interpretable machine learning models can effectively extract early indicators of elite-level show jumping potential from longitudinal performance data.
• This approach supports evidence-based talent identification, allowing more informed decisions in sport horse selection.
• The method respects the development timeline of the horse and promotes welfare by avoiding premature judgments.
• Ultimately, this research advances the use of data-driven tools in equine sports for better long-term outcomes and improved resource allocation in training and breeding programs.