Model: Random Forest Regressor Trained on Merged NFL Stadium Attendance Dataset

This Random Forest Regressor model was trained on a custom-preprocessed dataset of NFL stadium attendance, constructed by merging three publicly available datasets from Kaggle, uploaded by Sujay Kapadnis. The original datasets include game-level statistics (games.csv), season standings (standings.csv), and attendance records (attendance.csv).

The datasets were merged and cleaned to form a single modeling-ready table. Categorical variables (such as teams, location, and outcome) were transformed using one-hot encoding, while numerical variables (such as scores and rankings) were standardized using StandardScaler. The target variable is weekly stadium attendance for NFL games.

The full dataset was split into:

• Training set: 70%

• Validation set: 15%

• Test set: 15%

A GridSearchCV was applied to a RandomForestRegressor from sklearn.ensemble, using 3-fold cross-validation on the training set. The model was optimized to minimize mean absolute error (MAE).

Best hyperparameters found:

• n_estimators: 250

• max_depth: None

• min_samples_split: 2

• min_samples_leaf: 3

• max_features: None

After tuning, the final model was retrained on the full training set, and its performance was evaluated on the validation set as a proxy for generalization performance.

Technical environment:

• Python version: 3.13.1

• Libraries used:

  • scikit-learn (sklearn)

  • pandas, numpy

  • matplotlib (for evaluation visuals)

The model is saved as a .pkl file and is intended for inference on similarly preprocessed data. Numerical features must be scaled using the same StandardScaler, and categorical encoding must follow the same one-hot schema used during training. The target scaler (target_scaler.pkl) is also provided to inverse-transform predictions back to the original scale.

Model Risks, Biases, and Limitations:

• Risks: Misuse for speculative dynamic ticket pricing strategies without ethical oversight.

• Biases: Dataset may under-represent regional differences in reporting or non-ticketed attendance.

• Limitations: Model does not account for special promotional events, weather conditions, or real-time ticket resale dynamics.

fair4ml:mlTask | Regression – NFL stadium attendance forecasting
fair4ml:modelCategory | Supervised → Ensemble → Random Forest
fair4ml:intendedUse | Stadium attendance forecasting for operational, marketing, and planning decisions in the NFL.
fair4ml:modelRisksBiasLimitations | - Risk: Model predictions could be misused for discriminatory pricing. - Bias: Original datasets may underreport regional attendance discrepancies. - Limitations: Model does not include external factors like weather, ticket resale, or last-minute promotions.
fair4ml:trainedOn | - https://doi.org/10.82556/4fst-m890
fair4ml:validatedOn | - https://doi.org/10.82556/djvg-rb67
fair4ml:testedOn | - https://doi.org/10.82556/zv9b-6h09
schema:license + fair4ml:legal / ethicalSocial | License: CC-BY-4.0
schema:codeRepository | {{ https://github.com/emilp-tuwien/nfl-attendance-prediction/tree/main