Store Demand Forecasting

An end-to-end machine learning pipeline for forecasting retail store demand across multiple products. This project ingests transactional sales data, engineers time-series features, handles extreme data sparsity, and utilizes advanced gradient boosting architectures to generate reliable multi-step predictions.

Core Features

• Robust Time-Series Pipeline: Automatically processes raw transactional logs into continuous, chronological daily records. Efficiently handles missing dates by dynamic reindexing and zero-filling to maintain statistical integrity.
• Advanced Forecasting Models: Implements standard LightGBM for dense, continuous demand, alongside specialized formulations (LightGBM Tweedie Regression and dual-stage Hurdle classifiers) to effectively model zero-inflated, highly sparse retail environments without introducing false-positive volume spikes.
• Automated Feature Engineering: Generates temporal indicators, shifted lags, and rolling historical moments (mean, standard deviation, min, max) while enforcing strict temporal masking to guarantee zero data leakage.
• Interactive Dashboard: A Streamlit-based web application providing a user-friendly interface to upload datasets, select store-item combinations, and visualize historical sales trends against future forecasts.

Technology Stack

• Data Processing: Pandas, NumPy
• Machine Learning: LightGBM, Scikit-Learn
• Web Interface: Streamlit

Setup and Installation

1. Clone the repository:

   git clone <repository_url>
   cd Store-Demand-Forecasting

2. Create a virtual environment (Recommended):

   python -m venv venv
   source venv/bin/activate  # On Windows use: venv\Scripts\activate

3. Install dependencies:

   pip install -r requirements.txt

Usage

To launch the interactive forecasting dashboard:

streamlit run streamlit_app.py

1. Access the application in your browser (typically http://localhost:8501).
2. Upload your historical sales dataset (sales_data_sample.csv or similar).
3. Select the desired store and item combination from the sidebar.
4. Set your forecasting horizon and view the generated predictions.

Architecture Overview

The system architecture focuses on strict validation methodologies and computationally efficient inference:

• Validation Strategy: Time-aware chronological validation utilizing the final 30 days of data strictly segregated from the historical training window.
• Sparsity Management: Uses a conditional approach—employing Standard LightGBM for high-volume items and Tweedie regression for intermittent demand where >99% of dates represent zero sales.
• Latency: Capable of generating thousands of multi-step forecasts in under 50ms (measured prediction time).