Developed as part of my Machine Learning Internship at Future Interns (Task 1), this project implements a high-precision forecasting engine to predict daily sales for 1,115 Rossmann stores.
In the 2026 retail landscape, static forecasting is insufficient. This project utilizes a Multivariate Facebook Prophet approach to account for complex external drivers, ensuring that predictions remain robust against market volatility.
- Language: Python 3.12+
- Modeling: Facebook Prophet
- Data Wrangling: Pandas & NumPy
- Validation: Prophet Diagnostics (Cross-Validation & Performance Metrics)
Unlike standard univariate models, this implementation dynamically incorporates store-specific features to improve accuracy. Key regressors include:
- Promotional cycles (Promo, Promo2)
- Competition distance and entry dates
- State and School holidays
- Store-specific assortments and types
I developed a programmatic workflow to iterate through store attributes and inject them as regressors into the Prophet model, allowing for a scalable and maintainable codebase.
To ensure the model is "business-ready," I performed time-series cross-validation across multiple horizons (3 to 12 days).
The model was evaluated using RMSE, MAE, and MDAPE to ensure stability across different prediction windows.
Rossmann Sales Model Summary:
Best Accuracy: 7-Day Horizon (MDAPE: 12.3%) Average Error: MAE 1336.00 Confidence: 78% Average Coverage across 10-day forecast period.
This project uses the Rossmann Store Sales Dataset. The dataset is known for its high dimensionality and the challenge of managing store-specific trends.
git clone https://github.com/AnoMi-1/FUTURE_ML_01.git
pip install -r requirements.txt
jupyter notebook FUTURE_ML_01.ipynb