Predicting Product Returns in eCommerce

Returns are a margin problem wearing a customer service badge

An online fashion retailer selling 30 pound dresses with free returns can find that a returned item costs it two shipping legs, a warehouse inspection, repackaging and sometimes a markdown because the season moved on. Across a range where a large share of orders come back, returns quietly consume the profit on the orders that stay.

Most retailers treat returns as an operations cost to process efficiently. Return prediction treats them as something partly foreseeable, which means partly preventable. Our post on returns and exchanges on Shopify covers the process side; this one is about seeing them coming.

Two levels of return prediction

Product-level analysis is less glamorous and usually worth more. A single dress whose size runs small, a pair of boots whose colour looks different in person, a sofa cover described as fitting models it does not fit: each can generate a steady stream of returns that no order-level model will stop, because the cause is on the product page.

Signals that predict an order will come back

• Bracketing. The same item in two or three sizes or colours in one basket is the clearest signal there is.
• Product return rate. Items and categories with high historical returns.
• Customer's return history. Useful, but handle with care, as below.
• Discount depth. Heavily discounted impulse purchases come back more often in many categories.
• Size chosen vs customer's usual size. A shift from their normal size suggests uncertainty.
• Time of order and device. Late-night mobile orders in some categories behave differently.
• Delivery promise. Late deliveries turn into returns, especially for occasion-wear.

Return reasons are valuable too, when customers give honest ones. Most retailers' reason codes are too coarse ('did not suit') to help much. Free-text return comments, classified by a language model into specific causes such as 'runs small in the shoulders', often reveal more than the dropdown ever did.

What to do with predictions

1. Rank products by excess return rate against their category, and review the top twenty every month with buying and content teams
2. Show size guidance at checkout when the basket contains bracketed sizes or a size unlike the customer's history
3. Add fit notes to product pages where returns show a consistent cause
4. Feed predicted return volumes into warehouse staffing for the weeks after peak
5. For marketplace sellers, adjust stock allocation for items likely to return in resaleable condition

Order-level prediction also improves financial reporting. Revenue recognised before returns are known is optimistic, and predicted return rates make weekly trading figures more honest.

Why punishing serial returners usually backfires

The obvious idea is to identify customers who return a lot and restrict them: no free returns, fewer promotions, or an account closure. Some large retailers do this for genuine abuse such as wardrobing or fraud, and that is reasonable.

For most businesses, though, heavy returners are also heavy buyers. A customer who orders 2,000 pounds a year and returns half may still be among your most profitable once kept orders are counted. Restricting them based on a return score risks losing exactly the customers you want. Look at net value per customer, not return rate, before deciding who is a problem. Predicted customer lifetime value net of returns is the right lens here.

Fix the products people return before you fix the people who return them.

When return prediction is not worth building

Categories with low return rates, such as consumables, books or most groceries, rarely justify a model; a monthly product report is enough. Small stores with a few hundred orders a month can spot problem products by reading return comments. And if return reasons are not recorded at all, collecting them properly is the first project.

Some eCommerce platforms and returns apps now offer basic analytics on return rates by product. Start there. Custom prediction becomes useful at larger volumes, with many sizes and variants, or when return predictions need to feed checkout, warehouse and finance systems together.

How SpiderHunts would approach it

At SpiderHunts we would start with a product-level excess-returns report and a classification of free-text return reasons, because that usually finds money within weeks. Order-level prediction comes second, first as a quiet forecast, then as checkout nudges tested against a control group. Our machine learning work on returns is mostly careful data joining, since orders, returns, refunds and exchanges are frequently stored in different systems that disagree.

Frequently asked questions