tl;dr
What it is: The Returns Prediction Model gives you item-level insights into which orders and items are likely to be returned — before the return actually happens.
How it works: We built a data science model that runs predictions for every individual store, day, order, and item.
Why we've built it: If you have a high return rate (like a fashion brand), you want to make decisions based on your real net revenue and derived ROAS before returns start coming in.
1. How the Return Prediction Model Works
Most return reporting works at the store or category level — averages, totals, last quarter's number. Useful, but not actionable when you're trying to make a decision on a single order.
We've built something different: a model that predicts the probability of return for every line item in every order, the moment it's placed. You no longer have to wait for the return window to close — you see what's coming back on day one, item by item.
Under the hood, the model is a gradient-boosted classifier (CatBoost) trained separately on each merchant's order history. It learns from customer behavior, basket composition, SKU patterns, and shipping data, among other signals.
The model also accounts for order age. A two-day-old order and a three-week-old order are scored fairly, even though one has had no time to be returned yet and the other has nearly reached the end of its return window. This is the part most return models get wrong — and it's what makes ours reliable on fresh orders, not just historical ones.
When tested across 5 shops with return rates between 17% and 35%, the model predicted net revenue within 1.1% of the actual figure — across €131.9M in gross revenue and 3.8 million line items, scored on a strict out-of-time holdout with no per-merchant calibration applied.
2. Using the dashboard
2.1 Macro Level
The dashboard goes beyond predicting returns. Got additional logistics costs that kick in when something comes back? Those are factored in too. Every metric you rely on in Klar is available with predicted returns — and all connected costs — already priced in from the moment an order arrives.
2.2 Micro Level
Alongside your macro-level shop metrics, a Flex Table lets you slice by dimensions like channel, order name, or product SKU. This gives you a detailed view of where returns are likely to concentrate, so you can make budget and inventory decisions before the real returns come in.
2.3 Limitations
To use the model, your store needs at least €20,000 in net revenue, 100 returns, and a 5% return rate on average per month. This threshold exists because the model doesn't perform reliably with limited input data — and we'd rather set a clear bar than surface predictions you can't trust.
2.4 Detailed Results
Want to dig deeper into how the model is built and scored? Click here to get the detailed insights.