Building a dynamic inventory optimisation system: A deep dive

Zalando built a scalable inventory optimization system that combines demand forecasting and replenishment optimization to help e-commerce partners manage millions of SKUs across multiple warehouses. The system uses a two-step approach: generating probabilistic demand forecasts with LightGBM and Nixtla's MLForecast, then applying Monte Carlo simulations with gradient-free optimizers to minimize inventory costs. The architecture leverages AWS services including SageMaker, Databricks, and feature stores to provide both real-time interactive recommendations and daily batch processing, handling 5 million SKUs with 2-hour processing times.