Why it is a question
THEORY-375 reports one simulator (LIT-1113, read in NOTE-904): distance from the current query beat LRU where queries clustered or wandered with loops, and MRU was best on regular cyclic paths. The paper gives no analysis of the crossover, and NOTE-904 notes that its numbers point to cycle length relative to cache size without testing it. Which property of a workload decides between a forecast by content and eviction by recency is open.
What would answer it
An analysis, competitive or under a stated distribution of query paths, that says when a value based on distance from the current query beats recency-based eviction, or experiments in another system that vary cycle length relative to cache size and recover the crossover. More runs of the same simulator on the same synthetic paths do not settle it.