This project demonstrates a complete end-to-end Machine Learning pipeline built in the cloud using AWS SageMaker. The goal of this project was to ingest raw, transactional retail data, clean and aggregate it, and train an XGBoost model to forecast total daily revenue for the crucial holiday shopping season.
- Cloud Platform: Amazon Web Services (AWS)
- Compute: SageMaker Notebooks (
ml.t2.medium), SageMaker Training Jobs (ml.m5.large), SageMaker Real-Time Endpoints - Storage: Amazon S3
- Languages & Libraries: Python, Pandas, NumPy, Scikit-Learn, Matplotlib, Boto3
- Data Engineering & Cleaning (Pandas): * Ingested over 500,000 rows of raw transactional data.
- Handled missing values and removed negative quantities (canceled orders/returns) using Boolean masking.
- Aggregated transactional data into a daily time-series format (
groupby). - Reindexed the timeline to handle missing days (zero-sales days) to ensure a continuous time-series.
- Feature Engineering:
- Engineered time-based features (Month, Day, DayOfWeek, Year) from the
Datetimeindex to prepare the data for the XGBoost algorithm.
- Engineered time-based features (Month, Day, DayOfWeek, Year) from the
- Cloud Model Training (XGBoost):
- Staged the cleaned, split data (Train/Test) into Amazon S3.
- Configured and launched a SageMaker Training Job using the built-in XGBoost container.
- Deployment & Inference:
- Deployed the trained model artifact to a SageMaker Real-Time Endpoint.
- Formatted the holdout test data (Nov-Dec holiday season) and invoked the endpoint for predictions.
- Evaluation & Visualization:
- Visualized Actual vs. Predicted revenue using Matplotlib.
- Evaluated model accuracy using Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE).
- Safely deleted the endpoint post-evaluation to optimize cloud costs.
The model successfully captured the weekly seasonality and the general upward trend of the holiday shopping season.
