Serial returner scoring: how many returns make a pattern
Every returns team has a version of the same argument: at what point does a frequent returner become a serial returner? Set the bar too low and you harass good customers. Set it too high and the abusers operate with impunity. The answer is not a single number. It is a scoring model with thresholds calibrated to your own data.
Why return rate alone fails
A 60 percent return rate looks alarming until you learn the customer buys three sizes of everything and keeps one. Bracketing, the buy-multiple-sizes habit, produces sky-high return rates with zero fraud. Any threshold built on return rate alone will flag your most enthusiastic fit-uncertain shoppers alongside the actual abusers, and the false positives cost more than the fraud they were meant to catch.
The fix is combining signals. Return rate is one input. Wear evidence on inspection is another. Timing clusters, returns that spike around events or land suspiciously fast after delivery, are a third. Category skew, a return profile concentrated in occasionwear while basics are kept, is a fourth. No single signal convicts; the pattern across them does.
Building the score
Start with the signals you already collect, even if collection is informal. Warehouse inspection flags, return-reason distributions, days-between-delivery-and-return, and order frequency are usually available without new instrumentation. Weight them by how well each predicts confirmed abuse in your historical data, not by gut feel.
Keep the model explainable. A score that nobody can interpret produces decisions nobody trusts, and untrusted decisions do not get enforced consistently. A weighted checklist with four to six signals beats a black-box model for operational use, because the warehouse team and CX team need to act on it the same way.
Setting thresholds from your distribution
The most common mistake is importing someone else's threshold. A score of 70 means nothing until you know what fraction of your customers exceed it. Plot the score distribution across your customer base and look for the natural break: the point where the tail separates from the body. That break is your candidate threshold for review, and a higher break is your candidate for action.
Use tiers, not a cliff. A watch tier triggers closer inspection of future returns. A restrict tier triggers policy consequences like shorter windows or final-sale terms. A ban tier is reserved for confirmed, repeated abuse. Tiers give you somewhere to put the uncertain middle instead of forcing a binary innocent-or-banned decision that the data cannot support.
Calibrating against false positives
Every threshold has an error rate, and the errors have names. Pull a sample of customers above each candidate threshold and review them manually: how many are clearly abusing, how many are clearly innocent, and how many are ambiguous? If the innocent share is material, the threshold is too aggressive for automated action, though it may be fine for human review.
Track the overturn rate on appeals as a live calibration signal. A rising overturn rate means your thresholds drifted, your signals decayed, or shopper behavior shifted. Recalibrate quarterly, because the abusers adapt and the honest customers' behavior shifts with seasons and assortment.
When scores go stale: concept drift
Scoring models decay because both sides adapt. Abusers learn which behaviors trigger review and shift to just under the line; honest shoppers change habits with seasons, promotions, and assortment shifts. A model calibrated on last year's data slowly misreads this year's customers, and the decay shows up first as a rising review queue, not as a dramatic failure.
Fight drift with scheduled recalibration, not with alarms. Quarterly, re-plot the score distribution, re-check the natural breaks, and re-sample the above-threshold population for manual review. If the breaks moved, move the thresholds and document why. A scoring system with a maintenance calendar stays honest; one without it becomes folklore with a dashboard.
The takeaway
Score on multiple signals, set thresholds from your own distribution's natural breaks, act in tiers, and let the appeal overturn rate tell you when you drifted. A pattern is not a number; it is a number plus context, reviewed on a schedule.