Prediction of UFC Lightweight Winners Using Ensemble Machine Learning

Abstract

The Ultimate Fighting Championship (UFC) lightweight division presents significant prediction challenges due to factors including knockout variability, injuries, and fluctuating fighter momentum. This study develops an intelligent prediction system for UFC lightweight fight outcomes using ensemble machine learning, deployed as a web-based platform. Historical data from UFCStats.com comprising 6,000 fights and 675 fighters were collected and preprocessed. Feature engineering generated 63 differential attributes, including stance compatibility, recent performance metrics (last five fights), win streak differential, age difference, reach difference, and striking/takedown statistics. Multiple models, including XGBoost, LightGBM, and Logistic Regression, were optimized using Bayesian hyperparameter tuning, with Synthetic Minority Over-sampling Technique (SMOTE) applied to address class imbalance. The soft voting ensemble classifier achieved 79.25% accuracy and 88.67% ROC-AUC on time-based test data, representing a 13.7% to 14.2% improvement over previous state-of-the-art approaches. The primary contributions of this study include: (1) development of 63 domain-specific engineered features with quality adjustments and temporal weighting, (2) achievement of state-of-the-art prediction accuracy through optimized ensemble architecture, and (3) deployment as an accessible web application providing real-time predictions with confidence scores and market odds comparison—transforming academic findings into a practical decision-support tool. Validation against betting market odds demonstrated 76% agreement with market favorites and 82.1% accuracy in consensus cases, confirming alignment with domain expertise while identifying value betting opportunities.