The trick would be between finding the difference between actually needing the book later, or just hoarding something that will never be needed again...or at least, that introduces more of a cost to keep than it would cost to dispose of it and buy it again later if it does come back into fashion.
The only exception is if you knew a school were to assign the book as required reading to an entire grade - but that's really a rare and special case.
FWIW I think there is room for curated "should read" sections in a library. Optimizing purely for use is more the purview of a bookstore. But to support that the bulk of the shelf space should be serving the broadest use possible.
It seems like one of those "big data things", looking for odd relationships between things, like vampires being more popular during Democratic presidencies and zombies being more popular during Republican ones (as a tongue-in-cheek example: http://www.mrscienceshow.com/2009/05/correlation-of-week-zom...)
The librarian in the article cited a future need to re-buy a book that was previously culled. I guess what I'm trying to work toward would be an idea of finding non-obvious signs of when certain categories of books might become more and less popular, what patterns (if any) govern that schedule, and from that, make predictions of expected cost in keeping certain books around, as opposed to selling them off immediately. I'm playing around with the idea of whether there's a better way than "current readership + age of book" to predict, to a useful degree, the probability of it being more or less popular in the future.