In this project I build a Mixed Martial Arts Fight Predictor from the ground up, i.e. from scraping data to evaluating a model.
In order to run the experiments in your local machine you should do the following steps.
- Clone the repo by running
git clone https://github.com/ManosL/UFC-MMA-Fight-Predictor.git - Afterwards, install virtualenv in pip3(if you did not do that already) by running
pip3 install virtualenv - Then move to this repository directory.
- Then create and activate the virtual environment by running the following commands
virtualenv <venv_name>
source bin/activate
- Finally install the requirements by running
pip3 install -r requirements.txt - You are ready to move to
src/directory and run the experiments and demo programs!
In order to run the experiments done in order to write the report, go into src/ directory and run the following
command:
python3 experiments.py
WARNING: This will take time in order to complete.
While running this program you will see logs in terminal and the graphs, except the correlation matrix, will be opened in browser.
In order to run the demo you should move again into the src/ directory and run the following command:
python3 demo.py -t <training_dataset_path> -f <fighters_dataset_path> -p <prediction_dataset_path>
where the names of those variables are pretty descriptive, but we should also mention the following:
- Because the default datasets are in
src/datafolder the required command to run the demo is the following:
python3 demo.py -t ./data/Fights.csv -f ./data/Fighters.csv -p ./data/Matchups.csv
I defined those parameters in case someone wants to run the demo with different datasets, but with the same form.
<prediction_dataset_path>is a csv file where each row has the following form:
<weight_class>|<title_fight(true or false)>|<rounds(3 or 5)>|<fighter_1_id>|<fighter_2_id>
where fighters' ID should be taken from dataset given in the -f parameter.
- Demo trains a
RandomForestClassifierwith 150 estimators and the dataset is converted to be a Double Difference dataset, because with these configurations I saw that out classifier gave the best results. For more details, checkMachine_Learning_Report.pdf. - Note that because I could not take the gender of the fighters from their pages, I find initially using a library, which is prone to error. Thus, there is a case that a correct matchup will raise an error related to making matchup between fighters of different gender.
- Data are in
./src/datadirectory. - Crawlers and further data configuration are in
./datasetsfolder. - In order to get the latest datasets, run the
get_latest_datasets.shscript. This will run the fights and fighters webpages crawlers, then configure those datasets and finally they will get moved intosrc/data/directory. - Match-ups should be created by hand in order to predict them.
At 05/02/2022, with the data that I could acquire until now, I tested my model in two upcoming events
"UFC Fight Night: Hermansson vs. Strickland" and "UFC 271" in order to see how it will do on real-time.
In order to do that I run the demo and wrote the appropriate Match-up dataset which is in ./src/data/Matchups.csv
file. The results were the following:
The predictions done along with the actual results are the following:
1 . Jack Hermansson vs Sean Strickland
Jack Hermansson: 40.0
Sean Strickland: 60.0
draw: 0.0
Actual Winner: Sean Strickland
2 . Punahele Soriano vs Nick Maximov
Punahele Soriano: 36.67
Nick Maximov: 63.33
draw: 0.0
Actual Winner: Nick Maximov
3 . Shavkat Rakhmonov vs Carlston Harris
Shavkat Rakhmonov: 63.33
Carlston Harris: 36.67
draw: 0.0
Actual Winner: Shavkat Rakhmonov
4 . Sam Alvey vs Brendan Allen
Sam Alvey: 30.67
Brendan Allen: 68.67
draw: 0.67
Actual Winner: Brendan Allen
5 . Tresean Gore vs Bryan Battle
Tresean Gore: 48.67
Bryan Battle: 50.67
draw: 0.67
Actual Winner: Bryan Battle
6 . Julian Erosa vs Steven Peterson
Julian Erosa: 52.0
Steven Peterson: 46.67
draw: 1.33
Actual Winner: Julian Erosa
7 . Miles Johns vs John Castaneda
Miles Johns: 68.67
John Castaneda: 29.33
draw: 2.0
Actual Winner: John Castaneda
8 . Hakeem Dawodu vs Michael Trizano
Hakeem Dawodu: 56.67
Michael Trizano: 40.67
draw: 2.67
Actual Winner: Hakeem Dawodu
9 . Chidi Njokuani vs Marc-Andre Barriault
Chidi Njokuani: 55.33
Marc-Andre Barriault: 44.67
draw: 0.0
Actual Winner: Chidi Njokuani
10 . Alexis Davis vs Julija Stoliarenko
Alexis Davis: 36.67
Julija Stoliarenko: 61.33
draw: 2.0
Actual Winner: Alexis Davis
11 . Jailton Almeida vs Danilo Marques
Jailton Almeida: 52.0
Danilo Marques: 46.67
draw: 1.33
Actual Winner: Jailton Almeida
12 . Jason Witt vs Phil Rowe
Jason Witt: 45.33
Phil Rowe: 54.0
draw: 0.67
Actual Winner: Phil Rowe
13 . Malcolm Gordon vs Denys Bondar
Malcolm Gordon: 47.33
Denys Bondar: 52.0
draw: 0.67
Actual Winner: Malcolm Gordon
This means that my predictor achieved approximately 76% accuracy on this event, by predicting correctly 10 out of 13 fights which is pretty good and probably its better than teh performance of a casual fan like me, who cannot even predict correctly fights between two relatively unknown fighters.
The predictions done are the following(results are pending because the event is on 12/02):
1. Israel Adesanya vs Robert Whittaker
Israel Adesanya: 64.0
Robert Whittaker: 36.0
draw: 0.0
Actual Winner: Israel Adesanya
2. Derrick Lewis vs Tai Tuivasa
Derrick Lewis: 47.33
Tai Tuivasa: 50.0
draw: 2.67
Actual Winner: Tai Tuivasa
3. Jared Cannonier vs Derek Brunson
Jared Cannonier: 38.67
Derek Brunson: 61.33
draw: 0.0
Actual Winner: Jared Cannonier
4. Kyler Phillips vs Marcelo Rojo
Kyler Phillips: 72.0
Marcelo Rojo: 28.0
draw: 0.0
Actual Winner: Kyler Philips
5. Bobby Green vs Nasrat Haqparast
Bobby Green: 40.0
Nasrat Haqparast: 58.0
draw: 2.0
Actual Winner: Bobby Green
6. Andrei Arlovski vs Jared Vanderaa
Andrei Arlovski: 50.67
Jared Vanderaa: 48.0
draw: 1.33
Actual Winner: Andrei Arlovski
7. Roxanne Modafferi vs Casey O'Neill
Roxanne Modafferi: 21.33
Casey O'Neill: 78.67
draw: 0.0
Actual Winner: Casey O'Neill
8. Alex Perez vs Matt Schnell
Alex Perez: 64.0
Matt Schnell: 36.0
draw: 0.0
Actual Winner: N/A
9. William Knight vs Maxim Grishin
William Knight: 59.33
Maxim Grishin: 39.33
draw: 1.33
Actual Winner: Maxim Grishin
10. Mana Martinez vs Ronnie Lawrence
Mana Martinez: 49.33
Ronnie Lawrence: 49.33
draw: 1.33
Actual Winner: Ronnie Lawrence
11. Alexander Hernandez vs Renato Moicano
Alexander Hernandez: 38.0
Renato Moicano: 62.0
draw: 0.0
Actual Winner: Renato Moicano
12. Carlos Ulberg vs Fabio Cherant
Carlos Ulberg: 52.67
Fabio Cherant: 45.33
draw: 2.0
Actual Winner: Carlos Ulberg
13. AJ Dobson vs Jacob Malkoun
AJ Dobson: 52.0
Jacob Malkoun: 47.33
draw: 0.67
Actual Winner: Jacob Malkoun
14. Douglas Silva de Andrade vs Sergey Morozov
Douglas Silva de Andrade: 32.0
Sergey Morozov: 68.0
draw: 0.0
Actual Winner: Douglas Silva de Andrade
15. Jeremiah Wells vs Blood Diamond
Jeremiah Wells: 42.67
Blood Diamond: 56.0
draw: 1.33
Actual Winner: Jeremiah Wells
This means that my predictor achieved approximately 57% accuracy on this event, by predicting correctly 8 out of 14 fights which is not so good, especially considering the performance on last event.
This means that our model is generally unstable considering its real-time performance in these two events. However, it is consoling that in most of the wrong prediction our model selects as winner the favorite fighter, given the odds of a bookmaker.
DISCLAIMER: This project was done for educational purposes ONLY and does not support any betting activity that will result in addiction and property losing. However, if you want to apply this model in the real world, I suggest use that in order to level up in Verdict App.