Book

Description

This paper examines what happens when retail replenishment systems replace traditional forecasting (ARIMA, exponential smoothing) with machine learning, looking beyond forecast accuracy to system-level supply chain performance. It is framed for submission to Production and Operations Management. The literature shows ML models, especially gradient boosting, consistently beat traditional methods on MAE and RMSE, but prior work has focused on accuracy rather than operational consequences. In volatile omni-channel demand, highly responsive models can make replenishment unstable, adding a new structural driver to the bullwhip effect: algorithmic responsiveness. The study uses Gradient Boosting and LSTM networks on three years of point-of-sale data (about 10 million transactions). A controlled A/B deployment across live distribution centers compared ARIMA and exponential smoothing (control) with the ML models (treatment). The results show a clear trade-off. ML cut stockouts by 18%, improving service levels and revenue capture, but raised upstream order variability by 7%, as the models react to high-frequency POS fluctuations and pass volatile signals to suppliers. The paper calls this the "algorithmic bullwhip effect." It concludes that accuracy alone is not enough. ML forecasting should be paired with stabilising policies, such as order smoothing and multi-step demand planning, so downstream gains don't come at the expense of upstream stability. That's 1,469 characters including spaces.