Hierarchical Forecasting for Amazon EC2 demand

4:18 PM - 4:36 PM
Yiming Wang, Amazon; Ebrahin Nasrabadi, Amazon.
Amazon’s Stores, Digital, and Other (SDO) is the largest user of AWS EC2 capacity. This poster introduces our in-house Hierarchical Bayesian and Hierarchical Reconciliation forecasting methods to forecast SDO EC2 demand. The demand forecasts are published at bi-monthly cadence, covering a rolling 2-year horizon, and at granularity of business group (Stores, Digital, Other), availability zone (e.g., us-east-1a), and instance type (e.g., c5.12xlarge) level. The forecasts can be used for AWS capacity planning to drive operational and financial decisions on global AWS data center infrastructure and space plan. To address the cold-start challenges in forecasting new instance generation and availability zones’s demand, we developed a novel Hierarchical Bayesian time series forecasting model. In back-testing, our model was proven to reduce the forecast error measured in Weighted Absolute Percentage Error (WAPE) to 18.6% from 30.0% using Meta’s open-source Prophet package for new AWS EC2 instances launched in 2023.