Building a return prediction model from basket contents
The short answer
A return prediction model scores each order on basket signals that past returns have shown to matter: the same item in two or more sizes, several high-return products and a customer with a history of heavy returns. High-scoring orders can be flagged before they ship, giving service a chance to step in.
Why it matters
Every return costs freight, handling and often a markdown on resale. Flagging likely returns before shipping gives you a chance to confirm sizes, suggest the right fit or adjust free-return terms for the riskiest orders.
The question to ask
Copy this into Lexi as written.
What the answer looks like
Illustrative output on sample data
Illustrative example: how Lexi answers this for an apparel retailer scoring online orders before they leave the warehouse.
- Adds points for the same style in two or more sizes.
- Adds points for items with a return rate above 30%.
- Adds points if the customer returned more than 40% of purchases in the past year.
- Flags orders scoring 70 or above.
What the data shows
640 of this week's 9,800 orders are flagged. In back-testing, about 60% of flagged orders were returned.
What it means
Most flagged orders contain size duplicates, so a fit problem can be fixed before the parcel ships.
What to do
Send flagged customers with size duplicates a fit-advice email before dispatch.
The score predicts likelihood, so some flagged orders will be kept. Want me to push the flagged customers to Klaviyo?
Frequently asked questions
What data does return rate prediction in ecommerce need?
A year or more of online orders with line-level detail, a flag for each returned item and the return reason where you have it. Add product return rates, size and color, the discount used and each customer's past return history. Store returns of online orders need to be matched back to the original order.
Should a high return score stop an order shipping?
Rarely. Most flagged orders are legitimate, and blocking them costs sales and goodwill. Use the score to change what happens around the order, such as fit advice, a size check or different return terms, and keep any restrictions for a small group with a clear, documented pattern.
How accurate does a return prediction model need to be?
Accurate enough that acting on a flag costs less than the returns it prevents. Test the model on recent orders it has not seen. Check what share of flagged orders were returned and what share of all returns it caught. Set the threshold where the action you take pays for itself.