✈️ Flight Delay Analysis & Prediction (Nov 2024 to Oct 2025 – US DOT BTS)

📌 Project Overview

This project is an end-to-end Data Analytics and Data Science case study based on flight data from the U.S. Department of Transportation (DOT) Bureau of Transportation Statistics (BTS), covering from November 2024 to October 2025.

The goal is twofold:

1. Data Analytics: Perform exploratory data analysis (EDA) and build meaningful KPIs and visualizations to understand flight delays and cancellations.
2. Machine Learning: Build predictive models to classify whether a flight will be delayed (≥15 minutes), comparing multiple algorithms.

The project follows a complete pipeline:

• Data cleaning & preprocessing
• Feature engineering
• Exploratory Data Analysis (EDA)
• Visualization
• Machine Learning modeling & evaluation

---

🧰 Tech Stack

• Python
  • pandas, numpy
  • matplotlib, seaborn
  • scikit-learn
  • imbalanced-learn
  • xgboost
• Visualization
  • matplotlib / seaborn (EDA)
  • Tableau (dashboarding)
• Machine Learning Models
  • Logistic Regression
  • Decision Tree
  • Random Forest
  • XGBoost

---

📂 Dataset

Source: U.S. DOT – Bureau of Transportation Statistics (BTS)

Scope:

• Domestic US flights
• From November 2024 to October 2025

Key variables include:

• Flight date components (year, month, day, day of week)
• Origin & destination airports
• Departure and arrival times
• Departure and arrival delays
• Delay indicators (15+ minutes)
• Cancellation flags and cancellation codes
• Delay causes (carrier, weather, NAS, security)

---

🧹 Data Cleaning & Feature Engineering

Main preprocessing steps:

• Handled missing values while preserving cancellation information
• Converted time columns into datetime format
• Created categorical bins for departure time (dep_bin)
• Added engineered features:
  • season (Winter, Spring, Summer, Autumn)
  • delay_category (On Time, Minor Delay, Moderate Delay, Severe Delay, Cancelled)
• Encoded categorical variables using OneHotEncoder
• Scaled numerical variables using StandardScaler
• Built a unified preprocessing pipeline using ColumnTransformer

Special care was taken to avoid data leakage by embedding preprocessing directly inside ML pipelines.

---

📊 Exploratory Data Analysis (EDA)

The EDA focuses on operational and business-relevant KPIs, including:

• Distribution of delay categories
• Average delays by:
  • Day of week
  • Time of day (bins)
  • Season
• Cancellation rate by season and cancellation code
• Delay categories by season
• Calendar-style heatmap of average daily delays
• Identification of the most problematic routes

Visualizations include:

• Bar charts
• Pie charts
• Stacked bars
• Heatmaps
• Route maps
• Calendar heatmaps

Conclusions:

• The majority of flights are On-Time, so this will cause an Imbalance in our Machine Learnig models.
• The likelihood of flight delays increases as departure times move later into the day.
• Summer presents the highest volume of delayed flights, with delays occurring most frequently on Mondays, Thursdays, and Sundays, highlighting clear seasonal and weekly patterns.
• North Central West Virginia Airport records the highest average delays, with PSA Airlines Inc. identified as the airline most associated with delayed flights.
• The route with the highest probability to get delayed is from Roanoke-Blacksburg Airport to Orlando Sanford International Airport. The top four routes with the highest probability of delay are all scheduled to arrive in Orlando, highlighting Orlando as a key destination associated with increased delay risk.
• These analyses highlight temporal patterns, seasonal effects, and operational bottlenecks.

---

🤖 Machine Learning

🎯 Objective

Binary classification:

• 0 → On-time flight
• 1 → Delayed flight (≥15 minutes)

Primary business goal:

Maximize recall for delayed flights in order to proactively detect operational disruptions.

---

⚙️ Preprocessing Pipeline

A unified pipeline was used for all models:

• Numerical features → StandardScaler
• Categorical features → OneHotEncoder
• Models trained on top of the same preprocessing logic

---

🧠 Models Trained

Four supervised models were implemented and compared:

1. Logistic Regression (baseline linear model)
2. Decision Tree (non-linear baseline)
3. Random Forest (ensemble model)
4. XGBoost (gradient boosting)

Each model was evaluated using:

• Precision (delay class)
• Recall (delay class)
• F1-score
• Confusion Matrix

Later on, class imbalance was addressed using:

• Oversampling
• Undersampling
• SMOTE

Also used Hyperparameter Tuning, optimize model hypernaeters to maximize recall on imbalanced dataset. The methods I used:

• Grid Search: recall of 67%
• Cross-Validation: recall of 69,49%
• Bayesian Optimization: recall of 74,8%

Conclusion: Hyperparameter tuning did not improve recall over the simple undersampling approach.

---

📈 Model Comparison

Models were compared on the same test set using consistent metrics.

Key findings:

• Logistic Regression provides a strong baseline but struggles with non-linear relationships.
• Decision Trees improve recall by capturing feature interactions.
• Random Forest further stabilizes performance through ensembling.
• XGBoost delivers the best overall performance, achieving the strongest balance between recall and precision for delayed flights.

XGBoost was selected as the final model due to superior performance in detecting delayed flights.

---

✈️ Flight Delay Prediction WEB

All of this work led to the creation of a Flight Delay Prediction Web tool. The interface allows users to check if their specific flight is likely to be on time or delayed based on the real-world patterns we've identified. Link: http://192.168.68.103:8501

✅ Key Conclusions

• Flight delays show strong dependency on seasonality, time of day, and specific routes.
• XGBoost provides the best operational value for proactive delay detection.

---

Real-World Application

In a real airline environment, this model could be integrated into scheduling and operations systems to:

• Identify flights with high delay risk before departure
• Optimize crew and aircraft allocation
• Improve gate management and turnaround planning
• Provide proactive passenger notifications

🚀 Future Improvements

Potential next steps:

• Incorporate real-time weather data
• Perform deeper hyperparameter optimization