Airline Passenger Demand Forecasting System

An end-to-end Machine Learning solution for long-term passenger traffic prediction. Developed with a focus on airline industry standards, specifically targeting high-precision forecasting for carriers like Lufthansa.

Key Achievements

• High Accuracy: Achieved a 3.46% MAPE on a 24-month recursive backtest (blind forecast).
• Production-Ready: Containerized architecture using Docker & Docker Compose.
• Scalable Backend: REST API built with FastAPI, ready for integration with frontend dashboards.

---

System Architecture

The project follows a modern MLOps approach, separating data storage, model logic, and the serving layer.

1. Database: PostgreSQL (running in Docker) stores historical SFO passenger traffic.
2. ML Engine: XGBoost regressor with custom recursive forecasting logic.
3. API Layer: FastAPI service providing high-performance inference endpoints.

---

Tech Stack

• Languages: Python 3.10
• ML Libraries: XGBoost, Scikit-learn, Pandas, Statsmodels (SARIMA for baseline)
• Database: PostgreSQL + SQLAlchemy
• API: FastAPI + Pydantic + Uvicorn
• DevOps: Docker, Docker Compose

---

Feature Engineering

To overcome the limitations of decision trees in time-series forecasting, the system generates:

• Temporal Features: Month, Quarter, Year (to capture seasonality).
• Lags: 1-month, 3-month, and 12-month (critical for airline summer/winter cycles).
• Moving Averages: 3-month and 6-month rolling windows to capture trends.

---

Model Evaluation (Backtesting)

The model was evaluated using a Recursive Backtest starting from Jan 2023 to prove long-term stability:

• MAE: ~151,976 passengers
• RMSE: ~193,394 passengers
• MAPE: 3.46% (Average error over 2 years)

---

Installation & Running

Ensure you have Docker and Docker Compose installed.

1. Clone the repository:

   git clone https://github.com/mapi-developer/Passenger-Demand-Forecasting.git
   cd passenger-demand-forecast
   

2. Launch the entire stack:

   docker-compose up --build -d
   

3. Test the API:

   The API will be available at http://localhost:8000.

   You can access the interactive Swagger documentation at http://localhost:8000/docs.

API Example (Inference)

Send a POST request to /predict with the following JSON:

{
    "month": 7,
    "quarter": 3,
    "year": 2026,
    "lag_1M": 4500000,
    "lag_12M": 4450000,
    "rolling_mean_3M": 4300000,
    "pct_change_1M": 0.05
}

Output:

{
    "predicted_passengers": 4679611
}