Vendor Performance Tracking and Cancellation Prediction in Amazon’s Direct Fulfillment Operations

Efficient monitoring and forecasting of vendor performance are critical to the scalability and reliability of Amazon’s Direct Fulfillment (DF) model, where third-party vendors ship products directly to customers without involving Amazon’s own warehouses. This model requires reliable integration systems and timely order fulfillment across a broad network of vendors spread out over different regions. This thesis aims to develop scalable methods to monitor vendor performance and integration health, and forecast vendor cancellations across Amazon’s DF supply network. To achieve this, we designed and implemented the Vendor Health Dashboard, a centralized tool that automates monitoring of thousands of vendors worldwide, significantly reducing manual oversight and extending coverage to long-tail vendors who were previously difficult to track. For forecasting, we conducted a comparative evaluation of classical time series models, including ARIMA and ARMA-GARCH, alongside neural network models based on Gated Recurrent Units (GRUs). We assessed these models on their ability to predict cancellations, as well as their scalability and generalizability across different vendors and regions. Results show that ARMA-GARCH models perform well for short-term forecasts involving vendors with high variance, while GRUs deliver higher accuracy and better generalization for longer-term predictions across various vendor types and regions. Notably, both models maintained strong performance when applied to unseen vendors without requiring custom modelling, supporting their scalability for large operational deployments. These findings support a hybrid forecasting strategy that adapts to each vendor’s behaviour, enabling teams to step in proactively and use data to better manage risks such as order cancellations. Bringing together monitoring and forecasting in a flexible dashboard gives a solid way to improve supply chain efficiency, minimize disruptions for customers, and support future improvements such as real-time alerts, additional metrics, and automated management of predictive models.