Demand Forecasting and Inventory Intelligence Platform Using Prophet and XGBoost for Retail Analytics

The rapid growth of retail commerce has intensified the need for accurate demand forecasting
and intelligent inventory management systems. Manual and heuristic approaches are increasingly
inadequate given the scale and complexity of modern multi-store, multi-product retail environments. This
paper presents the Demand Forecasting and Inventory Intelligence Platform, a fully automated, data
driven system integrating two complementary machine learning paradigms: Meta’s Prophet library for
decomposable time-series demand forecasting and XGBoost, a gradient-boosted decision tree ensemble,
for multi-feature inventory quantity prediction. The platform processes a structured retail inventory
dataset (retail_store_inventory.csv) comprising 73,101 daily records spanning January 2022 to January
2024 across 5 retail stores, 20 products per store, and 5 product categories. A comprehensive feature
engineering pipeline is applied, incorporating lag variables (lag_7, lag_30), rolling window statistics
(roll_7_mean, roll_30_mean), calendar features, and an effective price signal. A rule-based three-tier
Stock Alert Engine classifies each inventory position as Critical, Warning, or OK based on the ratio of
inventory level to projected demand, translating model outputs into actionable restocking decisions. An
interactive Streamlit dashboard powered by Plotly provides real-time KPI metrics, configurable
forecasts, and downloadable alert reports accessible to non-technical business users. Experimental
evaluation on the held-out 20% chronological test set yields an XGBoost R² score exceeding 0.85 and
mean absolute error below 20 units per day, while Prophet achieves in-sample RMSE values of 15–25
units for typical high-volume SKUs. These results, coupled with the system’s modular architecture and
caching-optimised performance, demonstrate a scalable and practical solution for retail inventory
optimisation.