title	UFC Predictor
emoji	🥋
colorFrom	blue
colorTo	red
sdk	docker
python_version	3.12
pinned	false

UFC Predictor

Documentation available here

Model deployed in this Hugging Face space

The Model

The model uses the statistics from previous matches as features to predict the outcomes of upcoming fights. It consists of two stacked neural networks. Key statistics for each fighter are input into the first network, producing two vector outputs.

These outputs are then fed into the SymmetricFightNet, yielding a value between 0 and 1, where 0 indicates that the first fighter is predicted to win and 1 indicates the opposite. The loss function applied to the model optimises profit based on the odds input into the network, making the odds a crucial part of the model.

The output value of the network should be considered as a "normalised" predicted bet, which is scaled to the range (-100, 100) in the Gradio implementation.

Predictions

In examples the notebooks results_* show the fight forecast for the periods 2023/01-2024/10 and 2024/01-2024/10. Two different models are shown:

• With odds: This is the default model, where the odds are used in the loss function, but they are also used as input features in the model.
• Without odds: This model is trained using the same loss function, but the odds do not enter into the model.

The accuracy of both models on these datasets is summarized in the table below, where Favorite Accuracy is the accuracy if we just predict the favorite to win.

Period	Odds in model	Accuracy (%)	Favourite Accuracy (%)
2024	No	60.3	64.0
2024	Yes	64.0	64.0
2023-2024	No	58.6	62.1
2023-2024	Yes	61.8	62.1

The accuracy of the model is comparable to that implied by the opening odds, although slightly lower. When odds are incorporated into the model, the accuracy aligns with that of predicting favourites. However, the difference in predictions can still lead to profit, depending on the strength of the prediction and the odds. In the following plot, we compare the return for the different datasets:

while the model shows underperformance in certain periods, leading to loses, its overall performance shows potential for profit. When evaluating return, we see an improvement with the inclusion of odds in our model; but the behaviour of both curves is similar.

Deployment

The app is deployed in the following Hugging Face space, allowing for the prediction of fights by including the event date, fighter names and the odds assigned for each fighter.

If you want to deploy you own, you can follow the following guidelines:

Input data

In order to successfully deploy the prediction model, data needs to be downloaded and stored in a local folder. This data is expected to follow the structure defined in balaustrada/ufcscraper, with the following tables:

• event_data.csv
• fighter_data.csv
• fight_data.csv
• round_data.csv
• fighter_names.csv
• BestFightOdds_odds.csv

Alternatively, data can be downloaded on the fly if a Hugging Face token to the repo datasets/UFCfightdata is provided (currently a private repo).

Local installation

Install ufcscraper and ufcpredictor (tested for python==3.12):

RUN pip install git+https://github.com/balaustrada/ufcscraper.git
RUN pip install git+https://github.com/balaustrada/ufcpredictor.git

Run (without UFCfightdata token)

In this case you need to specify the data folder where the csv files are stored, and the model:

ufcpredictor_app --data-folder data --model-path models/model.pth

The model can be accessed in http://localhost:7860.

Run (with UFCfightdata token)

In this case you need specify the data folder where the data will be downloaded into, but also the token and the model:

export DATASET_TOKEN=DATASETTOKEN; ufcpredictor_app --data-folder data --model-path models/model.pth --download-dataset

The model can be accessed in http://localhost:7860.

Deployment with Docker

Include data (without UFCfightdata token)

We need to modify the Dockerfile uncommenting and editing the following line to include the folder where the csv files are stored. (Remember Dockerfile will only work with folders inside the relative path).

#ADD --chown=user data_folder $HOME/app/data

We should also modify the last line to remove the --download-dataset option:

CMD ["python", "app.py", "--data-folder", "data", "--model-path", "models/model.pth", "--port", "7860", "--server-name", "0.0.0.0",]

Include data (with token)

To include the data we only to set the environmental variable DATASET_TOKEN in our Dockerfile or at runtime.

Build image

The first step to deploy with Docker is to build the image. We need first to modify the Dockerfile uncommenting and editing the line:

docker build -t ufcpredictor:latest .

Run container (without token)

docker run -p 7860:7860 ufcpredictor

The model can be accessed in http://localhost:7860.

Run container (with token)

docker run -e DATASET_TOKEN=DATASET_TOKEN -p 7860:7860 ufcpredictor

The model can be accessed in http://localhost:7860.