Data-driven optimal stopping: A pure exploration analysis