Pricing Multiple Features in Microseconds

Professor Saša Pekeč designs more efficient online ad allocation

May 29, 2018
Big Data, Operations
Saša Pekeč designed a more efficient way to allocate online ads

In the microseconds after you click on any kind of online content, automated transactions happen at lightning speed as companies bid on what ads you’ll see when the page loads.

An associate professor at Duke University’s Fuqua School of Business has designed an auction mechanism to more quickly and efficiently match advertisers with the audiences they want.

The model created by Saša Pekeč leverages the fact that advertisers typically value target audiences who possess some features of interest, such as age, income level and location. But the research provides insights beyond digital advertising and can also help businesses price any portfolio of products with multiple features or options.

Pekeč’s work is focused on maximizing value by optimizing prices when consumers demand not just a single product, but bundles of distinct products.

“As long as I know you value something – even if I don’t know how much you value it – I can figure out the optimal prices,” Pekeč said. “If I have a demand forecast, if I have some data-driven estimate from previous sales or from industry on what consumer values might be, I can price them. And if I don’t have this information, I can use an auction to elicit it.”

The model sets what are known as market-clearing prices which ensure every buyer gets what they value the most at those prices.

“Market-clearing prices don’t always exist. But we show they can be set for buyers – advertisers, in this example – who want to buy bundles of ads targeting consumers with multiple demographic features,” Pekeč said. “We show how this can be done, and done quickly.”

The research, Efficient Allocation and Pricing of Multifeatured Items, is newly published in the journal Management Science. Pekeč worked with Ozan Candogan of the University of Chicago.

“The publisher gets the top dollar for their ad slot, and the advertiser doesn’t waste money.”

Millions of micro-targeted ads, processing massive amounts of data, are sold in automated penny auctions every second. Ad platforms like those on Google or Facebook have vast amounts of information about consumers. This information is valuable for targeted advertising only if it is possible to compute the value-maximizing allocation and pricing fast enough.

Amazon, for example, might be able to snap up the opportunity to show an ad to a user identified by their IP address, who has an item in his or her Amazon cart, as a reminder to complete the purchase.

“Even if all advertisers would tell you how much they value each customer – which they don’t, they keep that private – it would take all the time in the universe to efficiently clear one-second worth of the digital ad market,” Pekeč said.

His new model resolves the issue by expressing the private value advertisers put on a certain volume of each feature, such as a million ads shown to residents of the New York City metro area, or 2 million ads shown to men over 30, or a million ads to those who make more than $100,000 per year.

 “If there is a relationship among features, one example being a hierarchical structure, then we can find the right allocation without any loss of efficiency, and do it fast enough that it can actually be implemented,” Pekeč said. “The publisher gets the top dollar for their ad slot, and the advertiser doesn’t waste money.”

The research is useful beyond digital advertising. In labor markets, firms value a variety of skills, of which individual workers could have more than one. Pekeč’s approach can help efficiently clear labor markets by ensuring workers are employed by firms who value and pay them the most.

The research also provides guidance to companies making products with multiple optional features, such as automobiles or software packages, Pekeč said. His findings suggest the most efficient approach is not to offer features a-la-carte, but to bundle them carefully, sometimes even into just a simple hierarchy of basic, regular and premium packages.

“If you have only 10 optional features, there are 1,024 possible combinations and you cannot set prices for all of them. It would get out of control,” Pekeč said. “If you carefully design bundles of the features, even just sticking with good, better and best product portfolio, and if you have some information about consumer demand, you can quickly set correct prices,” he said. “This allows you to find the optimal price at which goods can be transacted so everyone gets what they want the most and at the most attractive price to them.”

This story may not be republished without permission from Duke University's Fuqua School of Business. Please contact media-relations@fuqua.duke.edu for additional information.

Contact Info

For more information contact our media relations team at media-relations@fuqua.duke.edu.