📺 Live Demo & Dashboard

The system concludes with a functional Streamlit Dashboard that bridges the gap between raw data science and business decision-making.

Key Dashboard Features:

Real-time KPIs: Displays average daily demand and dataset health.

Interactive History: Adjustable look-back periods (7 to 180 days) to analyze seasonal trends.

Automated Forecasting: Utilizes the saved XGBoost model to generate a 7-day rolling demand forecast.

How to Run the Dashboard:

1. Ensure you have run the Ml.ipynb notebook to generate daily_demand.csv and demand_model.pkl.
2. Run the following command in your terminal:

streamlit run app.py

Experiment Tracking with MLflow

I used MLflow to manage the model lifecycle. This ensured that every hyperparameter tweak was logged and measurable.

Experiment Name: Retail_Demand_Forecasting

Logged Metrics: Mean Absolute Error (MAE), Root Mean Squared Error (RMSE).

Artifacts: Automated versioning of the XGBoost model for deployment.

🛠️ Data Pipeline Architecture

1. Ingestion: Cleaning 1,067,371 rows of UCI Online Retail II transactional data.
2. Transformation: Aggregating hourly transactions into daily demand totals.
3. Feature Engineering: Creating lag variables ($t-1, t-7$) and holiday flags.
4. Modeling: Training an XGBoost Regressor with time-series cross-validation.
5. Inference: Serving predictions via the Streamlit web interface.