UFC Fight Prediction System: How to Surpass Academic Papers in Combat Sports Prediction Using Machine Learning

This article introduces an open-source UFC fight prediction system. Through strict temporal feature engineering and a five-model ensemble strategy, it achieves a prediction accuracy of 68.45% on unseen data from 2023 to 2026, surpassing the best academic result (66.71%) published in ACM ICIIP 2024. This system addresses the data leakage issue in sports prediction and provides a prediction solution applicable in real-world scenarios.

UFC prediction faces two core challenges:

1. Data Leakage Issue: Existing models often use future information, leading to excellent backtesting results but actual failure. This project uses strict temporal segmentation (training up to before 2023, testing covering 2023-2026) to eliminate leakage.
2. Dynamic Changes in Fighters' States: Career average statistics cannot reflect temporal changes in skills and physical fitness; rolling feature calculation methods are needed to capture dynamics.

Five-Model Ensemble Architecture

The system uses the average of prediction probabilities from five models:

• XGBoost (500 trees, depth 3, learning rate 0.01)
• LightGBM (500 trees, depth 3, learning rate 0.01)
• Random Forest (500 trees, depth 6)
• Logistic Regression (standardized scaling)
• CatBoost (500 rounds, depth 3, learning rate 0.01)

Rolling Feature Engineering

Features are calculated using only historical data before the match date, covering:

• Career performance (streak difference, total win difference, etc.)
• Physical attributes (height difference, reach difference, etc.)
• Offensive efficiency (strikes per minute difference, etc.)
• Strike distribution (head/body strike ratio difference, etc.)
• State decay (strike accuracy change rate, etc.)
• Finishing/defensive ability (finishing rate difference, takedown defense success rate difference, etc.)

Other Innovative Methods

• Style Collision Quantification: Position/target style distance, wrestling advantage, etc., to depict counter relationships
• Market Information Integration: Rankings, implied probability difference from odds, etc., provide incremental value (ELO rating has poor performance)

Confidence Stratification Performance

Tier	Confidence Threshold	Historical Accuracy	Backtest Return Rate
High Confidence	80%+	89.9%	+3.3%
Medium-High Confidence	75%+	86.6%	+4.2%

Temporal Leakage Verification

Extreme time span test (training before 2020, testing after 2024) still maintains 65.91% accuracy, proving no leakage; the version with leakage (v9) has 63.04% accuracy but fails in practice.

Feature Ablation Experiment

• Ineffective attempts: Sliding window (-0.61pp), exponential decay (-0.73pp), weight class split training (+0.06pp)
• Effective supplements: Defensive features (+0.12pp), reversal move features (+0.12pp)

Academic Comparison

Study	Accuracy	Method	Limitation
Yan et al. (ICIIP2024)	66.71%	GBDT	No temporal segmentation
This Project	68.45%	Five-model Ensemble	Strictly no leakage

The project provides a Streamlit web application:

1. Upcoming Event Prediction: Automatically loads the next UFC event, obtains real-time odds, and sorts prediction results by confidence
2. Custom Match Prediction: Select two fighters from a library of 2241 fighters for a match, supports inputting odds for comparison, and displays a statistical data comparison table

Requires The Odds API key (free version allows 500 requests per month); prediction is still possible without a key.

1. Temporal leakage is a core trap: Strict temporal segmentation and rolling features are the foundation of real performance
2. Feature engineering takes priority over model complexity: The improvement from a single XGBoost (66.02%) to ensemble (68.45%) comes from feature refinement
3. Domain knowledge creates differences: Features like style collision and state decay rely on understanding combat sports
4. Confidence stratification is key: Ignoring reliability stratification leads to failure in high-risk scenarios

This system demonstrates the potential of machine learning in sports prediction. The core strategies (temporal features, multi-dimensional features, model ensemble, confidence stratification) form a usable framework. It provides practitioners with verified technical paths and lessons learned, emphasizing reproducibility and practicality in real-world scenarios.