This project utilizes a proprietary machine learning pipeline to predict UFC fight outcomes and identify positive-expected-value (+EV) betting opportunities. Moving beyond basic physical attributes, the model leverages advanced feature engineering to quantify fighter durability, recovery curves, and psychological momentum.
By integrating historical betting odds with these domain-specific metrics, the system detects when public perception (and market pricing) drifts away from statistical reality. The algorithm specializes in identifying high-value upset scenarios where a fighter's durability or stylistic matchup is undervalued by the general market. In out-of-sample testing spanning from 2023 through late 2025, the model's "Underdog Hunter" strategy identified 262 value bets. From a total hypothetical wager of $26,200, the strategy yielded a net profit of over $18,600, representing an return of investment of approximately 71%.
The technical challenge of this project resided in quantifying relevant domain-specific concepts like "chin health" and "ring rust" into vectorizable features. Due to the proprietary nature of this feature engineering and the algorithm's potential commercial application, the source code remains private. However, the repository can still be found below.
View Code on GitHub