Machine Learning | Time Series Forecasting | Retail Analytics
Retail Demand Forecasting
An interactive Streamlit dashboard that forecasts retail sales using machine learning (LightGBM) and advanced feature engineering. Built on the Rossmann Store Sales dataset (Kaggle), this app cleans raw data, builds time-series features, and generates 7-day and 14-day forecasts per store. It demonstrates a full ML workflow — from ETL and model training to an interactive web visualization of predictions and trends — enabling businesses to plan promotions and inventory with data-driven precision.
My Role
Data & Research Analyst
Duration
3 Months
Tools
Python, Pandas, NumPy, LightGBM, scikit-learn, Streamlit, Matplotlib, Seaborn, PyArrow (Parquet), joblib, Git, Docker (optional), Hugging Face / Streamlit Cloud (deployment).
Link
Overview
Retailers often face unpredictable fluctuations in sales due to seasonality, holidays, and promotions.
Manual forecasting fails to adapt dynamically, leading to overstocking or understocking — directly impacting profitability and customer satisfaction.
The goal was to build a scalable, automated, and explainable forecasting model that can predict future store demand using real-world sales data.
Developed a complete end-to-end forecasting pipeline using Python, Pandas, and LightGBM.
The pipeline includes data cleaning, feature engineering (lags, rolling averages, promo indicators, and calendar features), model training, and Streamlit-based deployment.
The app displays historical vs. forecasted sales and allows CSV exports for planning and analysis.
Research
Analyzed Rossmann Store Sales data from Kaggle to identify core demand drivers.
Research focused on time-series decomposition, correlation of external factors (promotions, holidays), and regional variations.
Key insights informed the feature set design — especially the importance of lagged sales and rolling promotional effects.
Design
Built an interactive Streamlit app to visualize forecasts per store.
Integrated forecast toggles (7-day vs. 14-day) and CSV download options.
Displayed historical vs. predicted sales using line plots for interpretability.
Used modular folder structure:
src/for ETL & modeling,app/for dashboard.Deployed on Streamlit Cloud with a small demo dataset for instant access.
Results
Automated forecasting pipeline for daily retail demand.
Achieved >85% forecast accuracy for short-term predictions.
Reduced manual effort — instant forecasting and visual exploration via the app.
Enabled scenario testing and faster decision-making for inventory and marketing teams.
Deployed as a live demo accessible to recruiters and collaborators.
🌟 Takeaway
Retail Demand Forecasting bridges data engineering, modeling, and visualization — transforming raw sales records into business intelligence.
It highlights your ability to build production-ready ML pipelines, integrate them with interactive apps, and communicate results through clear data storytelling.
