Prediction Model
The MMA Prediction Model is a system designed to forecast the outcomes of MMA bouts, with a focus on UFC fights. Leveraging historical data the model provides insights into potential winners, betting odds, and strategies for each match-up.
In this first version, the model only predicts the winner/loser of MMA fights, but this can be expanded to include if a fight is going to be a finish or even applied to other sports.
Data
The prediction model relies on data from UFCstats, fightmatrix, elevation data & historical betting odds from bestfightodds. Combining all these different data sources provides us with roughly 80 different features. From those 80 features, roughly 3000 engineered features are created. The performance of all these combined is measured, and then the most useful ones are selected and used in the actual model so as not to overcomplicate the model.
Core Features
Include, but are not limited to, the following features:
Fighter stats (reach, weight, height, stance, etc.)
Historical fight outcomes (win/loss record)
Method of victory (KO/TKO, submission, decision)
Round-by-round performance metrics
Engineered Features
Include, but are not limited to, the following features:
Fighter stats differences (reach, weight, height, stance, etc.)
averages,cumulatives, per min, over all stats
Peak and Valleys of all stats and measuring the declines.
ELO, days since last fight, octagon time.
measurement of stats from recent previous fights.
The model's performance is then fine-tuned using Optuna, and is redone every single time the model is used. On top of that, cross-validation will also be used to decrease the amount of overfitting.
Backtesting Infrastructure
A backtesting mechanism is implemented to evaluate the model's efficency. It simulates the model's performance over historical data, beginning from a specified start date (e.g., January 1, 2018), and iteratively applies the model to predict outcomes for subsequent events using only the data up until that particular point.
Betting Strategy
The current betting strategy is based on the Kelly Criterion and is integrated into the model. This criterion suggests the optimal amount to bet based on the predicted probability of a fight's outcome and the corresponding odds, maximizing the expected logarithm of wealth, which is equivalent to maximizing expected geometric growth. Kelly strategy gives a return of 20-25% ROI.
Besides Kelly, there will also be a more handpicked approach that still uses Kelly but only applies this to fighters where we are much more certain about how they will perform.
More intricate strategies are still pending further analysis of the backtesting data.
Last updated