MMA PREDICTOR WORKFLOW¶

By Patrick Kong¶

Website | Github | LinkedIn

Contents:¶

This notebook shows the workflow my MMA Predictor project.

Section 1. Introduction¶

1. Background/questions
2. Objectives

Section 2. Data¶

1. Collection
2. Preview of raw fighter data
3. Preview of raw fight data
4. Cleaning/processing
5. Elo ranking formula
6. Preview of aggregated dataset

Section 3. Analysis¶

1. Summary statistics
2. Correlation matrix
3. Correlation heatmap

Section 4. Prediction Testing¶

1. Training models
2. Logistic regression
3. Decision tree
4. Random forest classification

Section 5. Conclusion and Next Steps¶

Section 1. Introduction

¶

1. Background/questions

Mixed Martial Arts is a unique sport. There exists a large number of variables from stemming from the multitude of styles and attacks allowed in the sport which can make the fights quite unpredictable. Therefore, I sought to create a predictor that could make sense of the sport's chaos and answer the following questions:

• What variables are the most important in determining the winner? (ie. striking? takedowns?)
• Is there a metric that can be used to determine who is the better fighter?

2. Objectives

The objective of this project is to create a model that:

• Predicts better than 50% accuracy
• Create a ranking system for fighters

Section 2. Data

¶

1. Collection

The data was gathered from UFC Stats using webscrapers built with Python's Beautiful Soup library. Two separate types of data were collected: fighter data and fight data. Both types of data were stored in separate Pandas data frames. The webscraper scripts can be found at the webscraper section of my github repository.

2. Preview of raw fighter data

Description of columns:

• id - fighter id
• name - fighter's name
• height - fighter's height (ft, in)
• weight - fighter's weight (lbs)
• reach - fighter's reach (in)
• dob - fighter's date of birth
• ss_min - significant strikes per minute
• str_acc - striking accuracy
• str_a_min - significant strikes absorbed per minute
• str_def - striking defense
• td_avg - takedowns landed average
• td_acc - takedown accuracy (%)
• td_def - takedown defense (%)
• sub_avg - submissions attempted avergae
• wins
• losses
• wl_diff
• momentum - current win/loss streak

3. Preview of raw fight data

Description of columns:

• event_id
• event_name
• f1_id - id of fighter 1
• f1_name - name of fighter 1
• f2_id - id of fighter 2
• f2_name - name of fighter 2
• winner_id - id of winner

4. Cleaning/processing

The follow steps were taken to clean the datasets:

• Fighter data
  • All rows with missing height and weight were dropped
  • Missing reach values were replaced with the mean reach of the weight class
  • Heights were changed into inches
  • Outliers were removed from the columns 'ss_min', 'str_acc', 'str_a_min', 'str_def'
• Fight data
  • This data was aggregated with the cleaned fighter data to create a composite database that contains the every fighter, their fights, and their attributes
• Aggregated dataset
  • Each row consists of a fight instance that shows a fighter, their opponent, the differences between their attributes, and the outcome of the fight
  • The columns 'elo' and 'elo_opp' were added to create an elo rating for each fighter, similar to that of chess

The data cleaning scripts can be found at the data cleaning section of my github respository.

5. Elo ranking formula

Formula for expected score of player A:

${\displaystyle E_{\mathsf {A}}={\frac {1}{1+10^{(R_{\mathsf {B}}-R_{\mathsf {A}})/400}}}~}$

A fighter's expected score is their probability of winning. Therefore, the higher elo a figher has, the higher their probability of winning.

Formula for updating elo:

${\displaystyle R_{\mathsf {A}}'=R_{\mathsf {A}}+60(E_{\mathsf {A}}={\frac {1}{1+10^{(R_{\mathsf {B}}-R_{\mathsf {A}})/400}}})~}$

6. Preview of aggregated dataset

The features of the aggregated data are as follows:

• id - id of fighter
• name - name of fighter
• id_opp - id of opponent
• name_opp - name of opponent
• event_id - event id
• event_name - event name
• elo - elo of fighter
• elo_opp - elo of opponent
• height_diff - difference between fighter and opponent's height
• reach_diff - difference between fighter and opponent's reach
• ss_min_diff - significant strikes per minute differential
• str_acc_diff - striking accuracy differential
• str_a_min_diff - significant strikes absorbed per minute differential
• str_def_diff - significant strike defense differential
• td_avg_diff - takedown average differential
• td_acc_diff - takedown accuracy differential
• tdd_def_diff - takedown defense differential
• sub_avg_diff - submission average differential
• wins_diff - difference in wins between fighter and opponent
• losses_diff - difference in losses between fighter and opponent
• momentum_diff - difference in win streak between fighter and opponent
• wl_diff_diff - difference in win-loss differentials
• outcome - outcome of the fight

Section 3: Analysis

¶

1. Summary statistics

2. Correlation Matrix

• We can see that the predictors are not very correlated with each other, meaning a low possibility of multicollinearity
• str_min_diff, str_acc_diff seem to have higher positive correlations with momentum

3. Correlation Heatmap

Section 4: Prediction Testing

¶

1. Training models

Using scikitlearn, we train and test a logistic regression, decision tree, and random forest classifier model.

Here we load a test dataset with the outcome column removed.

2. Logistic regression

Running our test dataset through a logisitic regression models gives a prediction rate of 89.93%.

3. Decision tree

Running our test dataset through a decision tree model gives a prediction rate of 73.83%.

4. Random forest classification

Running our test dataset through a decision tree model gives a prediction rate of 85.91%.

Section 5: Conclusion and Next Steps

¶

Summary of results:

Logistic regression

• 75.71% prediction accuracy on training set
• 75.21% cross-validation score
• 89.93% prediction accuracy

Decision tree

• 100.0% prediction accuracy on training set
• 67.82% cross-validation score
• 73.83% prediction accuracy

Random forest classification

• 99.90% prediction accuracy on training set
• 76.43% cross-validation score
• 85.91% prediction accuracy

Based on the results, the logistic regression model has the overall greatest prediction accuracy at 89.93%, with the random forest classification model close behind at 85.91%. These prediction rates meet our initial objective of predicting better than 50% accuracy.

As for next steps, we can improve feature engineering through greater analysis of decision variables, and create an interactive dashboard to display real-time predictions.