9,577 of 45,000 orders came back before anyone opened them. Your courier files it as remorse. Your own exchange records say a twelfth of it is a size chart that runs small.
11 reason codes on the manifest, and not one of them can say ran small.
A fit failure reaches you as REFUSED_COD — owner listed as checkout, marked rarely savable.
So the fix gets budgeted as fraud, freight and address hygiene — and the pattern never changes.
Nobody in this list owns the number. That is the whole problem.
Same ₹70.0L. Same orders. The left column is the axis your courier hands you; the right is the one your returns desk already keeps.
Every one of these is a delivery outcome. None is a product defect. The biggest bar, REFUSED_COD, is the bucket a customer lands in when the size was wrong and they never told you.
Nothing here asks the shopper for anything, and nothing waits on a courier to add a reason code they have no way to observe.
Click any style and the number decomposes into the exchanges that produced it — order id, size ordered, size kept, reason string.
A 35% exchange propensity is the conservative end of the range. Raise it and the attributed figure goes up, not down. We show the floor.
Styles injected with a bad chart are marked in the table only to score the estimator. The attribution never sees the flag.
239 customer identities across 1,838 households, 633 orders, ₹3.2L written off — returning at 3.3× the 21% baseline. These are not fit failures. They are the same graph, asked a different question.
Ranked by size-attributed cost. The offset is what the widget applies at the product page — a negative number means the garment runs small and the shopper should be sized up.
| Style | RTO | Size exch. | Attributed | Offset |
|---|---|---|---|---|
| Zoya Silk-Blend Midi Dresstruly small | 58 | 19.9% | ₹27k | −4.0cm |
| Kiara Silk-Blend Angrakhatruly small | 52 | 15.6% | ₹22k | −2.7cm |
| Noor Linen Kurtatruly small | 70 | 17.5% | ₹22k | −3.5cm |
| Saanvi Silk-Blend Midi Dresstruly small | 64 | 12.9% | ₹21k | −2.0cm |
| Kiara Cotton Anarkalitruly small | 60 | 12.5% | ₹20k | −2.3cm |
| Rehana Cotton A-Line Kurtatruly small | 73 | 11.0% | ₹19k | −1.9cm |
| Meher Georgette A-Line Kurtatruly small | 48 | 15.3% | ₹18k | −3.6cm |
| Kiara Khadi Angrakhatruly small | 67 | 10.0% | ₹15k | −1.9cm |
| Anaya Chanderi Straight Kurtatruly small | 49 | 12.4% | ₹15k | −2.6cm |
| Noor Georgette Anarkalitruly small | 41 | 18.2% | ₹14k | −3.7cm |
| Saanvi Linen Straight Kurtatruly small | 73 | 13.3% | ₹13k | −2.2cm |
| Meher Muslin Angrakhatruly small | 54 | 17.4% | ₹12k | −2.5cm |
Truly small is ground truth from the generator, shown only to validate the estimate. The model never sees it.
No. Everything on this page comes from a synthetic corpus of 45,000 orders generated for this project, where a known set of styles was deliberately given a bad size chart. That is the point: it is the only way to score the attribution against an answer key. On your catalogue the same pipeline runs, minus the answer key.
Because most shoppers who receive the wrong size never raise an exchange; they refuse the parcel, or keep it and never buy again. Counting only the exchanges you can see gives a ₹3.0L floor. Dividing by an assumed 35% propensity gives ₹8.5L. Both numbers are on the page, and the floor is defensible on its own.
It is a deliberately conservative assumption, not a measurement. Published apparel returns research puts the share of dissatisfied buyers who actually initiate a return well above that in most categories. If your real propensity is higher, the attributed figure shrinks; if it is lower, it grows. We picked the end of the range that makes our own case weaker.
Nothing. That is the design constraint. The attribution runs entirely on order history, exchange records and size charts — data the brand already owns. No new reason codes, no manifest changes, no integration.
It shifts the recommendation at the size picker on the product page, and feeds the try-on pose estimate. A −4.0cm offset on a style means the size chart overstates the garment by four centimetres at the binding measurement, so the shopper is guided up.
Against this corpus, ranking styles by attributed cost recovers 100% of the styles that were injected with a bad chart, and the total attributed sits under the true injected cost. Recall is the number worth watching; a ranking that surfaces the wrong styles wastes a pattern-making cycle.
The join takes an afternoon. The pattern correction takes a season. The freight you stop paying for compounds.