How better predictive models could lead to fewer clothing returns

In a perfect world, retailers would benefit greatly from online sales. The fixed cost of displaying an item on a website is marginal, and online stores don’t need sales staff.

Reality is another story, of course. In sectors such as fashion, average return rates for items bought online can be almost 20 times higher than those purchased in a store. Along with refunding customers and restocking items, retailers often have to pay shipping costs for items being returned.

“It’s not very profitable, and particularly in fashion it’s not very sustainable,” said a professor of marketing at MIT Sloan. If a winter item gets returned at the beginning of the spring, there’s no value in restocking it. “You just throw it away, and it ends up in a landfill.” (The U.S. Environmental Protection Agency estimates more than 85% of clothing and other textiles don’t get recycled.)

In a new paper, Hauser and co-authors Daria Dzyabura, ’07, PhD ’12, a professor at New Economic School in Moscow, Vienna University of Economics and Business professor Siham El-Kihal, and Emory University assistant professor Marat Ibragimov, SM ’22, PhD ’23, look at how predictive models can help retailers estimate return rates as they decide what to feature on their websites.

By incorporating product images into existing predictive models, the model predicted return rates more than 13% better than the baseline, and using it to dictate online offerings improved profits more than 8%, the researchers found.

Online shopping and high return rates

For the study, researchers looked at data from a women’s clothing retailer based in Germany that operated 39 brick-and-mortar stores and an online store that accounted for more than 30% of sales. From September 2014 to August 2016, the store sold roughly 1.23 million items; the analysis focused on 4,585 items that were sold at least 20 times.

Under European Union policy, customers can return items within 14 days without providing a reason, but they must be returned in the same channel of purchase. Return rates for products sold online ranged from 13% to 96%, with an average rate of 56% — compared to just 3% for items purchased in person.


Return rates for clothing sold online at a European store ranged from 13% to 96%, with an average rate of 56%.

Why the difference? Physical stores allows customers to look at and feel items, Hauser said. Customers can also see how items fit and pair with other items.

“You can imagine on the web, but every item is displayed on its own,” he said. “Until it arrives, you never know how it looks, feels, or fits. And even if it looks good on you, it has to fit within your wardrobe.”

More accurate return predictions using machine learning

Retailers already aim to predict return rates based on an item’s seasonality, price, and type of garment. “They’d be crazy not to,” Hauser said.

Researchers sought to augment this analysis by incorporating additional factors into a predictive model that can find common traits with items that are often returned. They began with color, as machine learning can detect subtle differences in shades better than the naked eye. Incorporating color alone improved the model’s predictions by less than 3%.

From there, the researchers added images to the model. Images are vital to any online clothes buying experience, and recommendation engines within shopping sites already use them to suggest similar styles and substitutes. Plus, images include colors as well as patterns, embellishments, and shapes. The latter is especially important for dresses. When purchased online from this retailer, they were returned 72% of the time, more than any other type of garment.

As the sample size included fewer than 5,000 images, which aren’t enough to train a predictive network, researchers added additional images from ResNet, a neural network commonly used to recognize and classify objects. RestNet was trained on a data set that included 1.3 million images.

Fewer returns, greater profits, less waste

The model enhanced by the neural network thus predicted return rates based on image and color in addition to the baseline of seasonality, price, and type of garment. This model improved the prediction rate by 13.5% compared to the baseline.

Related Articles

With that information, the retailer could choose not to display about 7% of the items in its inventory on its website. In doing so, the company could improve profits by 8.3%.

“Admittedly, it’s not a huge effect, but it’s still a lot of money for a profit-maximizing business,” Hauser said. It’s also an effect realized by exploring just one factor in returns. Other gains could be realized by focusing on customer behavior, search patterns, pricing models, and marketing strategies. In other markets, the impact of return policies would also warrant study.

Hauser added that the model offered the retailer two benefits beyond predicting returns. One was a form of diagnostics that let the company see which features of a garment were more likely to result in a return. Horizontal stripes and dark colors were returned at lower rates than vertical stripes and solid, bright colors. Form-fitting garments were more likely to be returned than casual items.

The second benefit was the potential for a more collaborative design process. Similar work from Hauser has explored this possibility in automotive design. Here, predictive models help narrow down the most aesthetically pleasing designs without the need for many expensive and time-consuming theme clinics, where customers are brought in to look at physical car designs.

“If a designer presents a retailer with a garment, and the predictive model says it won’t sell, then you can take that information back to the designer,” Hauser said. All the garment may need before it hits the online shelves is a change of color or embellishment. “You can gain a lot by running the model, and you’re contributing less waste.”

Read the paper: Leveraging the power of images in managing product return rates

Read next: Try this data framework for analytics advantage


Leave a Reply

Your email address will not be published. Required fields are marked *