With a lot of data, the aggregate probabilities will become reasonably accurate. If they predict that there's a 60% chance you'll buy, that they'll make $2 from you if you do, and they'll lose $2 on wasted shipping if you don't, then on average they'll make money by prepositioning your prospective purchase.
For Prime orders I can see it making a huge difference. Two-day shipping gets really expensive compared to slower methods when shipping over large distances, but for short distances it can be done entirely on normal ground delivery trucks and is cheap. If prepositioning wastes a few bucks sometimes and saves a bunch of money on expensive two-day shipping other times, it could be a big win.
For Prime orders I can see it making a huge difference. Two-day shipping gets really expensive compared to slower methods when shipping over large distances, but for short distances it can be done entirely on normal ground delivery trucks and is cheap. If prepositioning wastes a few bucks sometimes and saves a bunch of money on expensive two-day shipping other times, it could be a big win.