UFC-Predictor 🥊

The purpose of this project is to apply my passion for Machine Learning to the exciting world of MMA! In MMA, as well as any other combat sport, it is hard to consistently predict the victor due to the overwhelming amount of variables that can influence the outcome of a fight. However, I believe that it is worth wrangling the available data and leveraging machine learning to make statistically informed predictions. I am still working on the flow between Jupyter notebooks but it is currently designed as follows:

1. EDA, model creation, and training
2. Model evaluation on unseen data using classification error metrics (confusion matrix/classification report).
3. Running predictions for upcoming main card using the trained model as well as the associated W/L probabilities

Feature Selection

I used pandas built-in .corr() method to figure out which features appear most highly correlated with Winning. The original feature set included various gender and weightclass rankings so I made sure to filter them out with the following code:

pearson_corr = abs(master.corr()["Winner Dummies"]).sort_values(ascending=False)
ranks = []
for idx in pearson_corr.index:
    if "rank" in idx:
        ranks.append(idx)

pearson_corr = pearson_corr.drop(ranks, axis=0)

I then selected the ten features most highly correlated with winning from the filtered feature set:

features = master.loc[:, 
    [
        "B_odds", "R_odds", 
        "B_ev", "R_ev",
        "b_dec_odds", "r_dec_odds",
        "b_ko_odds", "r_ko_odds",
        "b_sub_odds", "r_sub_odds",
        "Winner"
    ]
]

Model Selection

Further analysis of the data must be conducted to acquire the features that appear most pertinent to winning. The current model used is a Support Vector Classifier (SVC) which I implemented with Scikit Learn's amazing library. A support vector classifier attempts to separate data geometrically via an (n-1)-dimensional hyperplane in an n-dimensional feature space while limiting the amount of misclassifications. The math behind this process is quite intensive, however, with the help of Scikit-Learn, this process is as simple as:

from sklearn.svm import SVC

model = SVC()
model.fit(X_train, y_train)

Pretty neat!

When I first began working on this project I was familiar with Logistic Regression, KNN, and SVM. I decided to utilize SVM when I first began this project, however I would like to transition to a Random Forest Classifier. There are a few reasons for this:

• Random Forests are an ensemble model (they are comprised of decision trees 🌲🌳), which means they may perform better.
• Decision trees are relatively robust to overfitting which allows me more room to control various hyperparameters (max depth, number of trees, etc)
• I can experiment with various meta-learning algorithms like Gradient Boosting and Adaptive Boosting to refine my model.

Edit

I used a GridSearchCV on a Random Forest Classifier and tried my best to increase its performance, however I found that a SVC model performed just as accurately with the same feature set. I compared these models against a Logistic Regressor and K Nearest Neighbors Classifier and a SVC appeared the most precise. I decided to integrate my trained SVC model into an API backend I created and will be monitoring its efficacy in the coming weeks.

TODO

• Perform more sensible EDA to find out which features are most highly correlated with winning.
• Add feature selection to README and other snippets showing relevant analysis of data.
• Configure a Random Forest Classifier model to train and evaluate.
• Automate download of new datasets from Kaggle repo using cron jobs.
• Develop backend API script that returns predictions as JSON.

Acknowledgements

This would not be possible without the Data consistently provided by mdabbert. I'm grateful that he takes the time to provide people with betting odds and relevant fighter data packed within a neatly formatted csv.