Anomaly Detection in Automated Buying Systems

Automated buying systems at Amazon order hundreds of billions of dollars worth of inventory each year into thousands of fulfillment centers across dozens of countries and millions of products. These systems provide hundreds of billions of vendor-product inbound recommendations each year. These decisions are made at the last responsible moment to maximize long term free cash flow and provide an optimal service level to the customers. These models use inputs from 100+ systems and consider factors such as demand, product price, vendor cost, vendor lead time uncertainty, vendor constraints, supply risk, variability of economic terms, customer demand conversion, salvage value, and the value of being in stock at various delivery speeds while adhering to network constraints. This necessitates development of automated systems to identify defective inputs which would have resulted in sub-optimal decisions. In this talk, we present machine learning based automated anomaly detection approaches to mitigate impact of defective inputs by building awareness of historical context into real-time buying decisions. The goal of these approaches is to stop egregious automated buys and reduce reaction time to large scale changes in inputs/outputs. This has a direct impact on Amazon’s profitability. We will also share real-world examples as part of this talk.