How does AI predict returns before checkout? Flagging risky baskets explained

KEY TAKEAWAYS Summary by the editors

1. Return prediction at checkout uses machine learning to estimate, before an order is placed, how likely each item or basket is to be returned, so a retailer can intervene while the shopper can still change the order.
2. Strong predictors include basket composition, such as the same item in several sizes or colours, and the customer's own return history; research shows accuracy drops sharply without historical customer data.
3. A 2024 study on a German online retailer of occasion wear reported balanced accuracy of up to 0.86 on future orders, but only 0.61 without historical customer data.
4. The most defensible interventions are helpful ones, such as size advice or a fit warning on the product, rather than blocking orders or penalising customers.
5. In the EU, online shoppers generally have a 14-day right of withdrawal, and the GDPR gives individuals the right not to be subject to solely automated decisions with legal or similarly significant effects, which limits how predictions can be used.

Predicting returns before checkout means scoring each item or basket for return risk while the shopper is still on the site. Machine learning models look at signals such as several sizes of the same item in one basket, the product's historical return rate and the customer's past behaviour, then trigger an intervention, typically a size suggestion or fit note, before the order is placed. The aim is to prevent avoidable returns, not to stop customers exercising their legal rights.

Why predict returns before checkout rather than after?

Once an order ships, a return is mostly a cost to be managed: transport, inspection, repackaging and often markdown. Before checkout, the retailer can still change the outcome. A shopper who adds a dress in two sizes may be persuaded by good size advice to keep only one; a shopper about to buy an item that is known to run small can be told so.

The scale of returns explains the interest. The US National Retail Federation estimated in October 2025 that total US retail returns would reach $849.9 billion in 2025, that online returns would equal 19.3% of online sales, and that 9% of returns were fraudulent. Fashion is among the most return-intensive online categories, so even a small reduction in avoidable returns is material.

How does a return prediction model work?

A return prediction model is a classifier: for each item in an order, it estimates the probability that it will be returned. It is trained on historical orders, labelled with whether each item was kept or sent back, and then applied in real time to open baskets.

Research has explored this for several years. A 2019 paper titled “Early Bird Catches the Worm: Predicting Returns Even Before Purchase in Fashion E-commerce” described a model, tested with a major fashion e-commerce platform, that estimated return probability on the cart page in real time so the retailer could take preemptive measures before the order was placed. A 2024 study presented at the ICAART conference by researchers at the August-Wilhelm Scheer Institut tested several algorithms, including XGBoost, on data from a German online retailer of festive and occasion wear.

Read also

Why returns are fashion e-commerce's costliest problem

How accurate is return prediction?

Accuracy depends heavily on the data. In the 2024 study, models reached a balanced accuracy of up to 0.86 on future orders and 0.86 on new products, provided historical data on customer behaviour was available. Without that customer history, balanced accuracy could not exceed 0.61. The dataset also showed how unusual some segments are: the overall return probability was 0.73, rising to 0.94 for multi-item orders, which reflects the occasion-wear category rather than fashion in general.

Two lessons follow. First, the customer's own history is often the most predictive signal, which makes the model strongest for repeat customers and weakest for first-time visitors. Second, results from one retailer or category do not transfer automatically to another; each business needs to validate on its own data.

What can a retailer do with a risky basket?

The choice of intervention matters more than the model. Helpful interventions reduce returns while improving the shopping experience; punitive ones risk alienating customers and attracting regulatory scrutiny.

• Size guidance: when a basket contains several sizes of one item, show a size recommendation and ask the shopper which fits best.
• Fit warnings: flag items that historical returns show run small or large.
• Better product information: surface measurements, model sizes and close-up images for items often returned as “not as described”.
• Delivery choices: suggest in-store try-on or collection where available.
• Review, not block: route extreme, fraud-like patterns to human review rather than automatically refusing orders.

What are the legal and ethical limits?

In the EU, consumers buying online generally have a 14-day right of withdrawal without giving a reason, running from delivery for goods, according to the European Commission's Your Europe guidance. Return prediction cannot remove that right. It can only help shoppers make better choices before they buy.

Data protection law adds a second limit. Article 22 of the GDPR gives individuals the right not to be subject to a decision based solely on automated processing, including profiling, which produces legal effects or similarly significantly affects them. Showing a size tip is unlikely to reach that threshold; automatically refusing to sell to a customer, or changing their terms because of a risk score, may well do so and needs careful legal assessment, transparency and human involvement.

Read also

How is AI used in fashion customer service, and where does it fail?

How should a retailer get started?

1. Clean the training data: link every return to the order line, SKU, size and a structured reason.
2. Start with a simple, explainable model and a single intervention, such as size advice for multi-size baskets.
3. Run a controlled test comparing return rates, conversion and customer satisfaction against a control group.
4. Document the model, its data and its effects for privacy and legal review.
5. Retrain regularly, because assortments, promotions and customer behaviour change every season.

Frequently asked questions

Can AI predict which online orders will be returned?

Yes, with useful but imperfect accuracy. Research on fashion e-commerce data shows models can estimate return probability before purchase, with accuracy depending heavily on whether the retailer has historical behaviour data for the customer.

What is bracketing in online fashion?

Bracketing is ordering the same item in several sizes or colours with the intention of returning those that do not fit or suit. It is one of the clearest signals return prediction models use, and size advice is the most common response.

Is it legal to refuse orders from customers who return a lot?

Rules vary by country, but in the EU the GDPR restricts decisions based solely on automated processing that significantly affect individuals. Retailers should take legal advice, keep humans involved in such decisions and be transparent, and they cannot remove the statutory 14-day withdrawal right for online purchases.

What data is needed to predict returns?

At minimum, historical order lines linked to return outcomes, product attributes and structured return reasons. Customer-level history greatly improves accuracy, which is why prediction works better for repeat customers than first-time shoppers.

GuideThe complete guide to AI in fashion e-commerce, marketing and retailRead the complete guide