A Drift-Aware MLOps Pipeline for Retail Demand Forecasting and Inventory Replenishment: A Semi-Synthetic Proof of Concept

Retail demand forecasting can lose accuracy as demand patterns evolve, which motivates operational pipelines that monitor model errors and govern replacement decisions. This study describes a drift-aware machine learning operations (MLOps) framework integrating stock-keeping unit (SKU)-level Prophet forecasts, rolling mean absolute error (MAE) monitoring, a mean absolute percentage error (MAPE)-based candidate-model safety gate, and inventory replenishment calculations. It is a controlled semi-synthetic proof of concept based on 160,000 Kaggle retail transactions: unit quantities are synthesized from category heuristics, daily series are extended beyond the observed period with a historical-mean procedure, and four controlled drift windows are injected into selected SKUs. Preprocessing yields 52,632 daily records for 72 SKUs. In a leakage-free rolling-origin evaluation over July 3 to 31, 2025 (2,088 one-day-ahead SKU-day forecasts), Prophet achieves an aggregate MAE of 1.49 units/day and a MAPE of 11.65%, against 1.83 and 1.85 units/day for the naive and weekly seasonal-naive baselines and 1.47 units/day for a historical-mean baseline. The illustrative SKU APPL-001 has an MAE of 1.43 units/day and a root mean square error (RMSE) of 1.72 units/day. These measurements do not identify an effect of retraining. The rolling-MAE detector detected 21 of 48 SKU-window injections (median delay 6 days) and raised 436 alert-days outside the injection windows; the held-out MAPE safety gate accepted 139 of 231 candidate models (60.2%). The pipeline design is coherent, but the study does not establish causal retraining benefit, real-drift detection accuracy, operational scalability, or production effectiveness, which require unit-sales data with naturally occurring drift.