Predicting Returns Before They Happen

By the time an item ships, most of the information that predicts whether it will come back already exists. The customer's return history, the specific SKU, the size they picked, how many variants sit in the same cart, the depth of the discount that triggered the purchase, all of it is visible at checkout. A return-propensity model turns that into a single score: the probability that this order, or this line item, comes back. Building the score is the easy part. The part that separates useful from harmful is what you decide to do with it, and how carefully you avoid letting the prediction become the cause.

What goes into a return-propensity model

A return-propensity model does not need to be exotic. The strongest signals are the obvious ones, and a well-built model on a handful of them beats a deep-learning system fed noise. The inputs fall into a few families: who the customer is and how they have returned before, what the product is and how uncertain its fit is, and how the purchase was made. The table below lays out the signals that carry the most weight in practice, along with why each one predicts a return, so you can prioritize the data you actually need rather than instrumenting everything at once.

What you do with the score, band by band

A propensity score is worthless until it maps to an action, and the right action changes with the band. The mistake is treating the score as a single lever, usually a punitive one, instead of a graduated set of responses that mostly help the customer avoid the return in the first place. The most valuable interventions happen before the order ships, when a nudge can still change the outcome, not after, when all you can do is process the return you already failed to prevent.

1. 1Low propensity: do nothing. Most orders sit here, and any friction you add is pure cost against a customer who was never going to return.
2. 2Moderate propensity: help before purchase. Surface size guidance, fit recommendations, and clearer measurements at the point of decision, the same storefront size recommendation work that reduces fit returns before they start.
3. 3High propensity from fit signals: intervene on sizing. Prompt the customer toward the recommended size, show fit reviews from similar bodies, and make the exchange path obvious so a miss becomes a swap rather than a refund.
4. 4High propensity from abuse signals: flag for review, not for punishment. Hold instant credit until the item is received and graded, and route to manual review, but only when the signal is genuinely abuse-shaped, not merely a high-returning good customer.

Measuring lift, or you are guessing

The moment you act on a propensity score, you change the behavior you are trying to measure, which makes naive before-and-after comparisons meaningless. If you surface size guidance to high-propensity customers and their return rate falls, you cannot tell whether the guidance worked or whether those customers were regressing to the mean anyway. The only honest way to measure a propensity program is a holdout: withhold the intervention from a random slice of the flagged population and compare. The difference between the treated and untreated high-propensity groups is the real lift, and it is often smaller than the raw numbers suggest.

This is where propensity modeling connects to the broader returns data flywheel: every prediction the model makes, checked against what actually came back, is training data that sharpens the next prediction, but only if you log the prediction, the action, and the outcome together. Analytics groups such as McKinsey have documented how predictive models decay when the feedback loop between prediction and outcome is broken, and returns are a textbook case, because the intervention and the result are separated by days or weeks and usually live in different systems.

The moment you act on a return prediction, you contaminate the data that would prove it worked. Holdouts are not optional; they are the only honest measurement you have.

The bias and self-fulfilling traps

Return-propensity scoring has two failure modes that do real damage, and both come from acting on the score too bluntly. The first is the self-fulfilling loop: a customer returns once, the model tags them high-propensity, so you hold their credit and add friction, so their experience degrades, so they return more or churn, and the model records that as confirmation it was right. The score caused the outcome it predicted. The second is bias: history-based models bake in whoever returned in the past, which can mean penalizing categories, sizes, or customer segments for correlations that have nothing to do with the individual order in front of you.

The defense is discipline about which band gets which action. Fit-signal-driven interventions, better size guidance, clearer measurements, an obvious exchange path, help the customer and carry almost no downside if the model is wrong, because a good customer who did not need the help is only mildly inconvenienced. Abuse-signal interventions, holding credit and manual review, carry real cost when the model is wrong, so they must sit behind a genuinely abuse-shaped signal and, ideally, a check against lifetime value, the same principle we apply in return fraud prevention. ResReturn's role here is to close the loop honestly: structured return reasons and outcomes flow back so a score is trained on what actually happened, and interventions are graduated by band, so the default response to a high score is to help the customer avoid a return, not to quietly punish them for one they have not made yet.

• Start with the strong, obvious signals, return history and in-order bracketing, before reaching for exotic features.
• Map every score band to a graduated action; most orders are low-propensity and should see no friction at all.
• Bias interventions toward pre-ship help, size guidance, fit reviews, an obvious exchange path, over post-hoc punishment.
• Measure lift with a randomized holdout, because acting on the score contaminates any naive before-and-after comparison.
• Guard against the self-fulfilling loop: reserve credit holds and reviews for genuinely abuse-shaped signals, checked against lifetime value.