🚦 Event Wait Time Prediction at Barclays Center

This repository showcases my work on predictive modeling of ingress wait times for large-scale events (NBA, WNBA, concerts, family shows) at Barclays Center, home of the Brooklyn Nets and New York Liberty.
The project was developed during my Data Analytics Internship at BSE Global, where I built and evaluated machine learning pipelines to optimize event-day operations and enhance the fan experience.

---

📌 Project Overview

Problem

Managing large crowds at live events requires anticipating queue congestion at the main entry gate. Traditional staffing decisions relied heavily on historical averages and manual adjustments, often failing to capture event-specific dynamics (e.g., weather, genre, arrival patterns).

Objective

• Predict 95th percentile wait times per minute relative to event start.
• Enable operations teams to adjust staffing and messaging proactively.
• Reduce fan complaints and improve Net Promoter Score (NPS).

---

🧠 Methodology

Data Sources

• Ingress scans from turnstiles (line length, arrival time, throughput) from S3/Snowflake.
• Event metadata: genre, performer/team, property (Nets vs Liberty vs Barclays events) from Dbeaver.
• Weather API: temperature, precipitation, severe weather flags.
• External factors: MTA delays, day of week, weekend flags, holidays.

Feature Engineering

• Historical rolling queue statistics: queue_mean_3, queue_slope_5.
• Event patterns: genre congestion factors, early-arrival risk scores.
• Cyclical encoding for time (hour_sin, day_sin, minute_sin).
• Weather-driven modifiers (temperature bins, precipitation intensity).
• Attendance scaling and expected crowd intensity multipliers.

Models

• Quantile Regression for 95th percentile prediction.
• Tree-based methods (LightGBM, CatBoost, XGBoost, Gradient Boosting).
• Baselines: Historical averages and regression benchmarks.
• Grouped cross-validation to preserve event-level independence.

---

📊 Results & Insights

🔍 Key Modeling Insights

• Stable Accuracy Across Time: MAE stayed between 2–4 minutes across all 5-minute buckets relative to event start.
• Model Benchmarking: LightGBM and XGBoost multi-minute models outperformed single-step models, reducing MAE to ~1.7–1.8 minutes.
• Error Distribution: >80% of predictions fell within ±2 minutes, ensuring operational reliability.
• Congestion Tracking: Predicted wait time curves closely mirrored actual ingress patterns, including pre-event spikes.

---

🔧 Visualizations

1. Model Stability Over Time

Shows Mean Absolute Error (MAE) across 5-minute buckets relative to event start.
Model remains well below the 5-minute threshold, demonstrating robustness.

---

2. Model Benchmarking

Left: MAE comparison (single vs multi-minute models).
Right: Error distribution showing tight prediction intervals.

---

3. Actual vs Predicted Curves

Predicted curves successfully track real-world congestion patterns, allowing operations to anticipate bottlenecks.

---

⚙️ Tech Stack

• Languages: Python (pandas, numpy, scikit-learn, LightGBM, CatBoost, XGBoost)
• Visualization: Tableau, matplotlib
• Data Infrastructure: AWS S3 → Snowflake ETL pipelines
• Collaboration: Jupyter, Git, Slack, stakeholder presentations

---

🚀 Next Steps

• Model Refinement: Incorporate live ingress + real-time weather feeds.
• Hybrid Models: Combine quantile regression + boosting for sharper tail predictions.
• Operational Integration: Automate Tableau dashboards for day-of-event decisioning.
• Simulation: Test “what-if” staffing/gate allocation scenarios.