Optimization of delivery promise policy to control peak demand on a fulfillment network

Abstract

Maintaining high on-time delivery performance is a key objective for Amazon.com and an expectation from customers. This objective is particularly challenging to achieve during periods of peak demand, when extremely high order volumes must be delivered by the promised date. During these peak periods, managing the availability of two-day delivery promises is critical to preventing the overloading of either fulfillment or transportation capabilities. Delivery promise policy establishes daily deadlines by which orders must be placed to be eligible for a two-day promise. Historically, the policy has been set to align expected outbound shipment volume with available transportation capacity, without a similar level of consideration for fulfillment network order processing capacity.

To help prevent the fulfillment centers from receiving too much work due out too soon, particularly during peak periods, fulfillment network capacity should be more explicitly considered when making promise policy decisions. This paper presents a methodology for using the promise policy to control the rate of new workable demand (NWD) arrival to the fulfillment network during a peak period. Analysis of data from Black Friday 2017 reveals that the promise policy impacts the speed with which ordered items are dispatched to fulfillment centers. This effect is incorporated into an optimization model which seeks to maximize the availability of two-day promises each day while constraining the resulting NWD to a daily limit. The model output is the latest recommended promise policy cutoff time for each day as well as hourly and daily predictions for network-wide NWD. The results of a pilot study conducted during the week of Prime Day 2018 (July 16-20) are presented and discussed.

For this week, the model recommended that two-day delivery promises be cut off of at 18:00 Pacific Time on July 17 to maintain NWD below daily limits, with no cutoff times required for the rest of the week. The hourly and daily NWD predictions achieved a weighted mean absolute percentage error of 8.9% and 7.3% across the week, respectively.