SuperKart Revenue Prediction

Business problem

Retail planners need product-store revenue estimates with uncertainty ranges so inventory and commercial decisions are not based on a single point prediction.

Dataset and context

Retail dataset with 8,763 rows and 12 raw columns. The target is Product_Store_Sales_Total. The README notes that this is not a true time-series forecasting setup and that random split/store overlap should be considered when interpreting results.

Methodology

• -Cleaned and prepared product-store features for supervised revenue regression.
• -Compared naive baseline, linear models, Random Forest, and XGBoost-style approaches.
• -Selected a serving path that supports point predictions plus q10/q90 prediction intervals.
• -Used SHAP to explain major revenue drivers to business users.
• -Deployed a Streamlit interface and documented a FastAPI serving option.

Technical approach

The project balances predictive performance with deployability. Random Forest performed strongly on RMSE, while the chosen serving approach emphasizes point prediction, interval output, explainability, and a practical app surface.

Key insights

• -Final RMSE is $281.04 and MAPE is 4.23%.
• -RMSE is reduced by 73.7% versus the naive baseline.
• -Prediction-interval coverage is 85.7%, giving stakeholders a useful uncertainty range.
• -The README transparently flags that the setup is supervised regression, not full time-series forecasting.

Business recommendations

• -Use predictions as planning support for product-store combinations with enough comparable history.
• -Show uncertainty intervals wherever estimates inform inventory, pricing, or staffing decisions.
• -Avoid over-interpreting random-split results as future-period performance until time-based validation is added.
• -Use SHAP summaries to identify product and store attributes that explain revenue variance.

Results and impact

Metrics are from the public project README. Production business impact would require real planning cadence, inventory constraints, margin, and out-of-time validation.

What I would improve in production

• -Add time-based validation if date fields or regular periods become available.
• -Monitor interval calibration and segment-level error by store and product category.
• -Connect predictions to planning outcomes such as stockouts, waste, or margin.
• -Clarify dataset redistribution rights before packaging the project for broader reuse.

Recruiter takeaway

This project shows practical machine learning for business planning: baseline comparison, regression quality, uncertainty intervals, SHAP explanations, and a deployable interface.