How Do You Predict Demand and Set Prices For Products Never Sold Before?

 
 
How can a retailer use its own data to determine what to charge for its products on a day-to-day basis? Kris Ferreira explains the value of data-driven-pricing
 
 
by Carmen Nobel

How can a retailer use its own data to determine what to charge for products it has never sold before?

That’s a question Kris Ferreira considered during a presentation at Future Assembly, an event at Harvard Business School where business leaders and academics discussed the challenges of operating in a digitally-transformed economy.

“All of these decisions would be easy to make if I knew what consumer demand will be. The problem is that I have a lot of uncertainty in demand”

All retailers face tricky tactical decisions related to assortment, inventory, product placement, and pricing. “All of these decisions would be easy to make if I knew what consumer demand will be,” Ferreira said. “The problem is that I have a lot of uncertainty in demand.”

Ferreira believes that the trick to tactical decision-making lies in quantitative analysis.

Summary of Data Used for Demand Prediction.

She explained that the world of business analytics includes descriptive analytics (analyzing what has happened), predictive analytics (analyzing data to figure out what will probably happen), and prescriptive analytics (using data to decide what to do next). “So how can we combine predictive analytics to predict demand with prescriptive analytics to make tactical decisions?” she said to a packed audience of executives, data scientists, and scholars. “I believe the answer lies in data.”

Ferreira presented field work she and colleagues conducted with the Boston-based online retailer Rue La La, while Ferreira was a doctoral student at the Massachusetts Institute of Technology. Rue La La is a so-called flash sales business, offering deeply discounted, extremely-limited-time offers on designer clothes and accessories.

Often, these limited-time offers include products the retailer has never sold before. Historically, some of these items would sell out almost immediately, suggesting the retailer could have charged more. But some items didn’t sell well at all during the sales period, suggesting they may have been priced too high.

Thus, the challenge the researchers aimed to solve: How do you predict demand and set prices to maximize revenue of products that have no historical sales data? To that end, they set out to develop a pricing decision support tool that would use existing data to maximize revenue on new products.

The researchers used machine learning techniques to estimate historical lost sales (demand of products that had sold out) and predict demand for new products that Rue La La was planning to sell in the future. As a part of the analysis, they realized that the demand for any given item depended on the price of the other items in that product’s category. Because of this, the researchers then developed an efficient multi-product price optimization algorithm to concurrently recommend prices for all products on Rue La La’s site in a given day.

Determining optimal pricing

Then, in January 2014, they teamed up with Rue La La on a field experiment to help the retailer set optimal pricing for new products.

Schematic of Pricing Software Developed for Rue La La to Predict Demand and Optimize Prices.

The researchers showed that Rue La La could increase revenue of products in the experiment by approximately 9.7 percent through price changes recommended by the algorithm, with minimal impact on aggregate sales quantity.

While this research focused specifically on a flash sales setting, they believe it could be useful for any retailer that needs to make pricing decisions for new products before the selling season begins. More broadly, this work illustrates how predictive and prescriptive analytics can be combined to develop a tactical decision-making tool that makes a big impact on the bottom line.

For details about the research and the successful machine learning tool, see the paper “Analytics for an Online Retailer: Demand Forecasting and Price Optimization” by Kris Johnson Ferreira, Bin Hong Alex Lee, and David Simchi-Levi. The paper won the 2014 INFORMS Revenue Management and Pricing Section Practice Award.

Related Reading from Future Assembly:
When Hosts Attack: The Competitive Threat of Online Platforms

About the Author

Carmen Nobel is the senior editor of Harvard Business School Working Knowledge.

Post A Comment

In order to be published, comments must be on-topic and civil in tone, with no name calling or personal attacks. Your comment may be edited for clarity and length.