NBA Stadium Attendance Prediction Using Game Intensity Modeling

This project predicts NBA stadium attendance by modeling game intensity and combining it with contextual data like weather. A two-stage PyTorch pipeline first estimates point differential as a proxy for excitement, then predicts attendance to support scheduling, promotion, and dynamic pricing decisions for the NBA.

🧠 Key Features

• Two-stage modeling pipeline: Predict game score margin and use it to forecast attendance.
• Data integration: Combines box score, weather, and game metadata from 2022–2025 seasons.
• Model serving: FastAPI-based inference with ONNX optimization.
• Online evaluation: Data drift detection and synthetic data testing for real-time robustness.

🏀 Motivation

NBA game schedules and TV allocations are often fixed and rarely optimized for fan interest. This project introduces an ML-based tool that forecasts attendance using game intensity signals and contextual factors, allowing the NBA to identify underperforming matchups and optimize resource allocation, promotions, and ticket pricing.

🧪 System Overview

1. Model 1: Predicts point differential using rolling averages of team statistics.
2. Model 2: Uses point differential and weather data to predict game attendance.
3. Serving: FastAPI endpoint for real-time inference; ONNX conversion for performance.
4. Evaluation: MLFlow for offline tracking, Alibi for online data drift detection.

📊 Evaluation Results

Component	Metric	Value
Model 1	RMSE / R²	Tracked via MLFlow
Model 2	RMSE / R²	Tracked via MLFlow
Online Eval	Response time	~1.2s avg
Drift Detection	Alibi + Gaussian	Detected expected changes
Final Output	Attendance MAE	Evaluated on 2024–25 games

📦 Data Pipeline

• Data Sources:
  • nba_api (box score, attendance)
  • WeatherAPI and Open-Meteo (temperature, wind, precipitation)
• Time Range: 2022–23 to 2024–25 regular seasons
• Preprocessing:
  • Rolling 5-game averages
  • Season-wise splits to prevent leakage
  • Weather joins by date and location

⚙️ Model Architecture

• Model 1 (Score Prediction):
  • PyTorch MLP
  • Input: Team stats (rolling averages)
  • Output: Point differential
• Model 2 (Attendance Prediction):
  • PyTorch MLP
  • Input: Output of Model 1 + weather data
  • Output: Predicted attendance
• Optimizations:
  • ONNX conversion
  • Graph and quantization optimizations for Model 2

🚀 Serving & Inference

• Interface: FastAPI web app
• Inputs: Game date, home team, away team
• Flow:
  1. Retrieve historical team stats and weather
  2. Run Model 1 → predict score diff
  3. Run Model 2 → predict attendance
• Backend Support:
  • .pth and ONNX model formats
  • Model staging, fallback logic
• Limitations:
  • Requires valid NBA team acronyms and dates
  • GPU-based performance tuning was partially constrained

🔍 Online Evaluation

• Synthetic Data:
  • Generated using Gaussian noise and future date shifts
• Drift Detection:
  • Implemented with Alibi Detect
  • Reports confidence scores and alerts

🧱 Infrastructure

Component	Description
2 × m1.medium VMs	Training and serving environments
A100 GPUs (x2)	Model training with DDP
Persistent Storage	Object store for data and artifacts
Floating IPs	Model endpoint + VM communication

🧰 Tech Stack

• Python, PyTorch, ONNX, FastAPI
• MLflow for experiment tracking
• Alibi for online evaluation
• Trovi + Chameleon Cloud (KVM@TACC) for compute

👥 Contributors

Name	Focus Area
Will Calandra	Model training
Lake Wang	Model serving, monitoring
SungJoon Moon	Data pipeline
All Members	Planning, integration, testing

📌 Future Improvements

• Full implementation of offline evaluation and load testing
• CI/CD integration for Ray-based job management
• Enhanced dashboarding for drift and performance monitoring
• Real-time feature enrichment (e.g., betting markets, injuries)