A new study that could help retailers manage returns and improve profit by more than 8% uses illustrative machine-learning models that provide face-valid interpretations and reinforce the value of using prelaunch images to manage returns.
- The research team demonstrated that item images improve predictions of return rates, that policies based on predictions can improve profit, and that data-based insights are face valid, internally consistent, and suggest which items are returned at high and low rates.
- The team displayed that incorporating item images into models helps a firm decide, prior to launch, which products to include in its online store based on profitability net of returns.
- The approach is fully automated, scalable, implementable prior to product launch, and an improvement on current practice that does not incorporate product images.
- The approach has the advantage that it can be easily implemented by a retailer for each fashion collection.
- The focal retailer can thus improve profit by 8.3% and identify items with features less likely to be returned, says the study ‘Leveraging the Power of Images in Managing Product Return Rates’ authored by Prof Daria Dzyabura, New Economic School, Moscow; Prof Siham El Kihal, School of Finance & Management, Germany John R Hauser, MIT Sloan School of Management, USA, Asst Prof Marat Ibragimov, MIT Sloan School of Management, USA.
THE METHOD: The team developed a modelling framework to predict and interpret how product images relate to their return rates. And for this the team focused on a large European apparel retailer and observed that item return rates were averaging 53% ranging from 13% to as high as 96% for some items. This was in contrast with the 3% return rate in the same retailer’s offline channel, with the same set of items.
- Even with high margins, the items on the higher end of this spectrum generated a net loss for the firm’s online store.
- Machine learning models produce accurate predictions of an item’s return rate based on features of the product image and other characteristics available prelaunch. For example, including deep-learning image features in gradient-boosted regression trees (GBRT) predicts 13.5% better than a model based on traditional features alone.
- Using this model and the derived policy to decide on which items to display results in a profit improvement net of returns by 8.3% relative to displaying all items in the online channel.
- SHAP values (that relate the interpretable features to return rates) based on automatically-generated interpretable image-processing features suggest how the firm might design (or otherwise source) items less likely to be returned.
- The team tested a variety of alternative machine-learning models and features to suggest which do well and which do not on their data. Among those tested were deep-learning features, human-coded features, hand-crafted automated pattern and colour features, and the automatically-generated image-based interpretable features.
THE CONTEXT: In fashion, as in many industries, the product return rate is key input into any product management strategy. The problem in the fashion industry is that fashion seasons are short and return deadlines are generous.
- By the time an item’s return rate is observed, the fashion season is well under way or almost over. To effectively manage item assortment in light of returns, it is critical that the retailer is able to predict item return rates using only data available before the launch.
- 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.
