Inside an RTO Score: The Signals That Predict a Return Before You Ship

Ask an experienced ops person which order will RTO and they will guess right more often than chance — COD, new customer, rural PIN, no phone. What they're doing is mental arithmetic on signals. An RTO score is that arithmetic made explicit, run on every order, and honest about what it doesn't know.

The signals, weighted

Not all warnings weigh the same. Cash on delivery is the heaviest single feature — refusal costs nothing to a buyer who hasn't paid, so COD orders return at roughly two to three times the prepaid rate. Within COD, value matters: a ₹12,000 cash parcel carries more downside than a ₹500 one.

The address contributes its own score: a structural check that fails (see our piece on address risk) adds meaningfully — and COD stacked on a weak address compounds, so the combination carries its own penalty beyond the sum.

Customer history is the most predictive per-person signal: a first-time buyer sits in the population where RTO concentrates, while a customer with three delivered orders and no returns earns a credit. Prior returns count against, and a pattern — half or more of their orders returned — counts hard.

The phone is cheap to check and disproportionately useful: missing, placeholder or invalid numbers mean the courier cannot rescue a failed attempt. Geography rounds it out — rural PINs return more often, and a measured delivery-success rate for the destination prefix is the only signal here that isn't a prior: it's your own track record.

Why additive beats opaque

A black-box model might edge out a weighted sum on raw accuracy. It loses on the thing that matters operationally: trust. When the tool says "hold this order," someone must be able to read why — COD, no phone, PIN with a 60% delivery rate — and disagree if they choose. "The model thinks so" is not a reason a merchant accepts twice.

Additive scoring also degrades honestly. No customer history? The signal is absent, not assumed. PIN never seen before? Its zone contributes nothing rather than a flattering default. Every adjustment is a named code with a weight, which means the score can be audited, argued with and recalibrated as your own data grows.

A risk score you can't explain is a risk score you'll start ignoring by the second week.

What the bands mean

The score lands in one of three bands tied to your own hold threshold: Healthy (ship normally), Suspicious (worth a confirmation or a second look — the band starts at half the threshold), and Risky (at or above it — hold for review or convert to prepaid first). The point was never to stop orders; it's to spend the five minutes of attention where it prevents a ₹200 mistake.

The bottom line

RTO is mostly predictable, and predictability is cheap when the signals already live in the order. Add the points, show the reasons, keep your own thresholds — and the returns you do take start looking like the price of business rather than a tax on guessing.