Flight Delay Prediction

Project Overview

This project implements a machine learning pipeline to predict flight departure delays of 15 minutes or more (DepDel15) using integrated operational and weather data. By analyzing 105,783 flight records across 10 major U.S. airports, the system identifies high-risk delay windows to support proactive airline operations planning.

Key Performance Metrics (XGBoost)

• ROC-AUC: 0.735
• Recall: 0.543
• PR-AUC: 0.448

High-Level ML System Architecture

Data Engineering Pipeline

The system utilizes a modular processing architecture to merge disparate datasets:

1. Flight Data Processing:
   • Ingested 3 months of Bureau of Transportation Statistics (BTS) data.
   • Filtered to 10 primary hubs (ATL, EWR, JFK, LAS, LAX, MCO, MIA, ORD, SEA, SFO).
   • Feature selection reduced 100+ raw columns to 15 high-impact predictors.
2. Weather Data Integration:
   • Flattened hourly JSON observations into structured features.
   • Extracted: tempC, windspeedKmph, precipMM, visibility, and weatherCode.
3. Feature Engineering:
   • Temporal alignment: Merged datasets on Airport, Date, and Departure Hour.
   • Engineered dep_hour and DayOfWeek to capture cyclic delay patterns.

Key Insights

• Propagation Effect: Delay probability scales from ~10% in the morning to ~30% in the evening.
• Non-Linearity: Weather impacts (wind/visibility) exhibit threshold-based behavior rather than linear correlations.
• Interaction Driven: The strongest predictors are interactions between Departure Hour and Day of Week.

Modeling & Evaluation

We compared multiple architectures, prioritizing Recall to maximize early-warning capabilities for operational risk mitigation.

Model Comparison

Model	ROC-AUC	PR-AUC	Precision	Recall	F1
XGBoost (Final)	0.735	0.448	0.403	0.543	0.463
Random Forest	0.730	0.436	0.443	0.461	0.452
Gradient Boosting	0.730	0.440	0.461	0.404	0.431
Logistic Regression	0.679	0.321	0.423	0.221	0.291

Final Model Configuration

• Algorithm: XGBoost
• Hyperparameters: learning_rate: 0.1, max_depth: 4, n_estimators: 300, subsample: 0.8.
• Interpretability: SHAP analysis identifies dep_hour, Origin, and windgustKmph as the primary drivers of model decisions.

Tech Stack

• Language: Python 3.13.5
• Data Science: pandas, numpy, scikit-learn
• ML Framework: xgboost
• Visualization: matplotlib, seaborn, shap

Future Roadmap

• Data Expansion: Incorporate full-year seasonality and aircraft tail-number rotation tracking.
• Advanced Features: Integrate real-time METAR/TAF weather alerts and airport congestion metrics.
• Deployment: Implement temporal cross-validation and an automated early-warning alert system.

Usage

Installation

pip install -r requirements.txt

Execution Order

Run the following notebooks sequentially to reproduce the results:

1. 00_data_directory.ipynb
2. 01_flight_data_preprocessing.ipynb
3. 02_weather_data_preprocessing.ipynb
4. 03_merge_flight_weather_data.ipynb
5. 04_exploratory_data_analysis.ipynb
6. 05_baseline_modeling_data.ipynb
7. 06_final_modeling_data.ipynb