Problem
OLX is a classifieds marketplace. With Pay&Ship the seller sends the item by a carrier, and OLX holds the payment until the package arrives. Buyers weigh delivery time when choosing Pay&Ship, yet the ad page showed only carriers and prices. The goal was to raise the share of buyers who choose Pay&Ship and cut churn among those with a bad delivery experience. OLX does not ship goods, so the total time is the seller's handover time plus the carrier's transit time.
Example from the article: the ad page listed each provider with its cost and no delivery time; the mockup added a range such as "Arrives in 3-5 days".
Solution in brief
First show a cheap per-carrier range to test whether buyers use an estimate at all. Then train two CatBoost quantile models (alpha 0.05 and 0.95) that give the lower and upper bounds, and keep them only if they beat that baseline.
Questions this case answers
- How to predict a duration as a range that is correct for most cases when the duration is the sum of two independent handover steps by different parties
- How to justify the cost of a prediction model by first testing the value of showing any estimate with a cheap statistical baseline
- How to show buyers a delivery time estimate when the marketplace does not ship the goods itself?
- How to prove that a delivery estimate is worth building before training a model?
Tools and methods
- Per-carrier mean and standard deviation — baseline range without a model.
- Quantile regression — two models predict the 5th and 95th percentile of delivery days, giving the range bounds.
- CatBoost — gradient boosting that handles categorical features such as carrier and region.
- Batch prediction — scores ads when they are created, with no buyer data.
- MLflow, Docker, GitLab, Airflow, AWS — tracking, packaging, deployment and automation.
Google or a one-time email code. No password.