NFL Big Data Bowl 2026 — CAP4621 Group Project

This repo contains our team's code and dashboard for the NFL Big Data Bowl 2026 Prediction Competition on Kaggle.

---

Setup Instructions

1. Clone the Repository

git clone https://github.com/btabman/UF_CAP4261_F25_TEAM9.git
cd nfl-big-data-2026

2. Create and Activate Virtual Environment

python -m venv venv

• Windows:

venv\Scripts\activate

• Mac/Linux:

source venv/bin/activate

3. Install Required Packages

pip install -r requirements.txt

4. Set Up Kaggle API

Ensure Kaggle is installed:

kaggle --version

Log into Kaggle: https://www.kaggle.com/

Go to Profile → Account → Create New API Token

Move the downloaded kaggle.json file to:

• Windows

C:\Users\<YourName>\.kaggle\kaggle.json

• Mac/Linux

~/.kaggle/kaggle.json   

5. Download the Competition Data

kaggle competitions download -c nfl-big-data-bowl-2026-prediction
python -m zipfile -e nfl-big-data-bowl-2026-prediction.zip data/raw

6. Clean data set

Use the preprocess.py file to import the training and test data, perform some cleaning, and convert to parquet files. This will greatly inmprove speed compared to repeatedly reading from csv and/or holding the entire data set in memory.

7. Visualizing data in Power BI

Have a look at the NFL.pbix file to get familiar with the data. It will show the player movements in the field as well as the ball landing location (shown as brown diamond on the scatter plot). The primary_key is a concatenation of the game_id and play_id fields, which can be selected on the list slicer to visualize different plays. There are also range slicers for the frame_id and output_frame_id, to visualize movement through time for the training set and test set respectively. If you need to access to Power BI, you can do so with your Gatorlink credentials: https://it.ufl.edu/cloud/collaboration-tools/office-365/?

8. Running Models

MiniMax.py

This script includes a miniMax algorithm that functions by play. Each play object includes player objects. The Player() class controls the movement of each player as described by physics. The Play() class has all the functions needed to create the game tree, which currently implements alpha-beta prunning. Outside of both classes, is the predict() function, which takes in two polars dataframes: one with input data and another with output data. This function creates the needed Player() and Play() objects, runs the miniMax algorithm, and returns predictions. At the end of the script, there is a section that is in charge of loading datasets, running the predict() function for different plays, and reporting RMSE.

Transformer.py

How to run and train the model simply put into the ternminal `python3 src/models/transformer.py'. This will start training a model for yourself. In the config dictionary you can adjust any of the features to your liking/capability of your machine.

player_model.ipynb

First use feature_processing.ipynb to transform the original dataset into an enriched dataset with more physics attributes and clustered formation and play layout variables The player_model notebook uses the enhanced data to build an attention model by exploring various attention parameters to find an optimal configuration. If you wish to load a .pt model built on this data structure and run it directly, us ethe code in notebooks/player_from_pt

CNNRNNHybrid

To train the model run CNNRNNEvaluator.py. This uses the model configuration defined in CNNRNNHybrid.py and training utility defined in CNNRNNHybrid.py. In the evaluator you can specify certain configurations. Most notably you can change the MAX_SAMPLES to a lower value such as 500 for quick testing. To visual the results see the results directory when the program is finished running. Information about the effectivness of the model can be viewed there or printed to the terminal.

9. Running the app

To run the app you can put into the terminal `python3 src/app/main.py'. From the terminal there will be a link to a localhost where you can then run and use the models already stored and compare the different RMSE scores of all the models.