Hierarchical Demand Forecasting at Scale
Design a high-scale machine learning system for a global retailer to forecast daily demand for 1M+ SKUs across 10,000 stores. The system must handle hierarchical time-series data, integrate exogenous signals like promotional calendars and weather, and provide probabilistic forecasts for safety stock optimization. Address the technical challenges of 10B+ daily predictions, data sparsity, cold-start for new products, and the trade-offs between global and local modeling approaches in a production environment.
LightGBMSparkRayQuantile RegressionFeastTweedie LossAirflowS3Parquet
00