E-commerce platforms such as Amazon and JD.com have grown tremendously over the past decade—but so have order-fulfillment costs, which can hurt the bottom line. While Amazon’s revenue grew twentyfold from 2009 to 2021, its fulfillment costs surged twice as much, observe Chicago Booth PhD student Mohammad Reza Aminian, Columbia’s Will Ma, and Booth’s Linwei Xin. The researchers created a model that could help online retailers contain costs.
At issue is multiordering, where customers place several separate orders in quick succession. For example, a customer who ordered a pair of sneakers on Amazon might put in an order for yoga pants 25 minutes later. Such customer behavior is fairly common; 30 percent of JD.com’s repeat customers place another order within half an hour.
To address this, companies have introduced a pause to potentially consolidate orders before dispatching them. Assuming the items are at the same warehouse, a shipping delay would enable Amazon to send the sneakers and yoga pants together, cutting delivery expenses. (For more, read “Why Making Customers Wait Pays Off for Online Retailers.”)
Platforms could implement, say, a 30-minute delay for all orders—as some have done. But that can result in universal delays, the researchers observe. As the researchers note, universal delays are impractical because they mean “all orders must be temporarily held . . . even if only a handful are adjusted.”
Companies can also run into capacity issues and have to decide which orders to hold. In that case, platforms could delay only certain orders on the basis of past customer behavior, the researchers write. If a particular customer on the platform often orders two items within a short period, it would make sense to hold the first order. And if another person’s history doesn’t show such a pattern, that order could go right to the warehouse for picking. Some companies are, in this vein, personalizing delays.
However, this may sound simpler than it is in practice, as retailers still face trade-offs related to specific situations and goals, such as how many orders they can hold and for how long. As Ma further explains, “Good criteria for holding are orders with a high chance of consolidation and orders that have not already been in the system for too long. However, sometimes these criteria are at odds with one another, which creates a trade-off.”
The researchers’ model captures the trade-offs and works out solutions that take into account a range of circumstances.
A dynamic algorithm could help with inventory management.
Why Retailers Should Reject Some Online OrdersAminian, Ma, and Xin proposed three algorithms and analyzed them in the model. The first one targets the immediate “reward rate” of the order, or savings on shipping costs acquired proportional to the multiorder rate of the customer.
Assume that Customer A orders socks on Amazon. Let’s say the maximum time to hold an order is three hours, there is one fixed cost for shipping each parcel, and Amazon can hold only one order. There’s a 50 percent chance Customer A will order something else in any of the next few hours. Now suppose that Customer B orders a coffee maker two hours later. If past behavior suggests Customer B has only a 20 percent chance of placing another order in any of the next few hours, Amazon might pay to ship the coffee maker and keep holding the socks for another hour, particularly if the retailer is looking chiefly toward maximizing the immediate reward rate.
This is the model’s “most conservative” algorithm for deciding whether to keep holding an order, Aminian says, and it works best when the platform is extremely busy. When demand is high, aiming for immediate rewards and taking less risk is optimal, the researchers find.
However, a downside of the first algorithm is that it ignores how long each order has already been in the system. A second algorithm aims to maximize the total remaining reward for each customer. Consider the same scenario. Because the maximum hold time is three hours, the socks order for Customer A must go out within an hour by the time Customer B buys the coffee maker. Yet Customer B’s order could be held for a full three hours longer.
On the basis of the potential return and the time remaining, the second algorithm would dispatch the socks first and hold the coffee maker. This is because the three-hour window is seen as an advantage over the remaining one-hour window, even though Customer B has a lower probability of placing a second order.
This is the “least conservative” algorithm, Aminian says. “It’s the most volatile—it tries to switch as much as possible as long as it sees the opportunity.” This would be the best option when there are very few orders, he says.
A third algorithm aims to land somewhere in the middle, managing a wide variety of order situations. For example, when order demand is moderate, it seeks to find a balance between immediate and total remaining rewards and hold or ship orders accordingly. This strategy would also be effective—though not optimal—when volume is high or low, the researchers find. A retailer trying to cover all situations without switching strategies might prefer this one.
The researchers note that their model could further cut costs by capturing order modifications or cancellations before delayed shipments go out. Beyond e-commerce, the researchers suggest that it could be useful for cloud computing platforms managing capacity, hospitals deciding discharges based on patient urgency, and businesses determining whether to outsource incoming new projects or ongoing ones when they run out of their own capacity.
Along with this research shedding light on key trade-offs, these algorithms could “affect the efficiency of how such companies make their decisions under future uncertainty through increasing welfare in healthcare, profit in cloud computing, and savings in fulfillment cost for e-commerce retailers,” Aminian says.
Mohammad Reza Aminian, Will Ma, and Linwei Xin, “Real-Time Personalized Order Holding,” Working paper, November 2023.
During the pandemic, algorithms overrode the human tendency to panic buy.
When Retail Buyers Panic, A.I. Keeps Calm
Regional transportation networks can blunt the impact of supply-chain disruptions.
Why Some Regions See Gas Price Surges After a Storm—And Why Others Don’t
Researchers create a computationally efficient model for assemble-to-order operations.
Ever Closer to an Optimally Cost-Efficient Assembly-Line OperationYour Privacy
We want to demonstrate our commitment to your privacy. Please review Chicago Booth's privacy notice, which provides information explaining how and why we collect particular information when you visit our website.