Detecting Unmarked Sold Listings on Facebook Marketplace from Existing Signals

You are the product manager for Facebook Marketplace. Many sellers never use the option to mark a listing as sold, so a listing can stay up long after the item is gone. Using functionality Marketplace already has, and metrics you can build from it, how would you judge whether a given listing has probably sold? ```hint Compare before and after a sale Think about what a seller and an interested buyer each do on Marketplace while an item is still for sale, and what changes once a deal has happened. Every difference that leaves a trace in an existing feature is a candidate signal. ``` ### Clarifying Questions - What will the estimate be used for: cleaning up search and browse results, prompting sellers, or measuring how many items actually sell? The answer sets how costly a wrong "sold" call is. - Should the answer stick to signals that exist today, or may it add small touches such as asking sellers to confirm availability? - Can the content of buyer–seller messages be analyzed, or only metadata such as reply times and thread counts? - Are business inventory, vehicles and rentals in scope, or only individuals selling used items? ### What a Strong Answer Covers - Why unmarked sold listings matter to buyers, sellers and Marketplace, and how the intended use sets the balance between precision and recall - A structured inventory of existing signals from seller actions, buyer–seller conversations, buyer behavior and listing age, with the direction each one points - Metrics derived from those signals and a way to combine them into a likelihood, including how to get trustworthy labels when "sold" is under-reported - How the estimate would be validated, plus privacy limits and failure modes such as listings with several identical units ### Follow-up Questions - Once a listing scores as probably sold, what would you do with it in search, on the listing page and with the seller, and how would you limit the harm when the score is wrong? - Sellers who mark items as sold may behave differently from those who never do. How does that bias a model trained on marked listings, and how would you correct it? - How would you tell a listing that sold apart from one the seller simply abandoned, and does the difference change what you do?