Time-Series Forecasting in Sports: Using LSTM and GRU for Stadium Attendance Prediction

Abstract This study investigates the use of Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) models for predicting stadium attendance in National Football League (NFL) games. Using a comprehensive dataset spanning 26 years and incorporating various game-specific, economic, and temporal features, the performance of LSTM and GRU architectures in forecasting attendance rates is compared. The analysis reveals that both models effectively capture underlying patterns in attendance data, with the LSTM model demonstrating slightly superior predictive accuracy. Optimal configurations are identified through a comparative evaluation of different hidden sizes and layer counts. The best-performing models achieve a Root Mean Squared Error (RMSE) of 7.81% and a Mean Absolute Error (MAE) of 5.62%, representing a significant improvement over previous approaches. While the LSTM model exhibits better adaptability to sudden variations in attendance, the GRU model offers faster convergence and more consistent predictions. To enhance transparency, SHapley Additive exPlanations (SHAP) were used to interpret model outputs. The results revealed seasonality, team identity, and economic indicators as the most influential factors, aligning with domain knowledge and supporting practical applications. These findings contribute to the growing field of AI-driven demand forecasting in sports, offering valuable insights for financial decision-making and operational planning in the sports industry.