This repository contains my solution to a demand forecasting challenge involving the prediction of daily probabilistic demand distributions for item-store combinations. The solution is structured across two Jupyter notebooks and is designed to be easy to run end-to-end.

Place the following files in the same directory:

├── Basic_Data_Exploration_and_Merging.ipynb       # EDA, cleaning, feature engineering
├── Probabilistic_Forecast_Generation.ipynb        # Modeling, forecasting, PMF construction
├── sales.csv                                      # Historical sales data / not included
├── promo_price.csv                                # Promotional pricing data / not included
├── regular_price.csv                              # Regular pricing data / not included
├── requirements.txt                               # Python dependencies

pip install -r requirements.txt

Run the notebooks in order, without skipping cells:

1. Basic_Data_Exploration_and_Merging.ipynb

   • Loads and merges raw data
   • Handles missingness, outliers, and promo adjustment
   • Performs exploratory data analysis and feature engineering

2. Probabilistic_Forecast_Generation.ipynb

   • Trains LightGBM quantile regression models
   • Generates forecasts for the target week (Sept 12–18, 2022)
   • Constructs a discrete probability mass function (PMF) per forecasted demand

✅ Both notebooks assume all files are in the same folder. ✅ No additional configuration is needed.

• Python, Jupyter Notebooks
• LightGBM (quantile regression)
• Pandas, NumPy, Scikit-learn
• Plotly, Matplotlib (interactive and static visualizations)

The final output includes:

• Forecasted quantiles (q10, q50, q90) for each item-store-date
• A discrete probability mass function (PMF) over demand values
• Visualizations of historical trends, forecast intervals, and features' importance

Gregory Barbas
📧 Email: gregorybarbas@gmail.com
💼 LinkedIn
🖥️ GitHub

For questions or contributions, feel free to reach out!

• For the sales and price data, feel free to reach out.

This project is licensed under the MIT License.