Say you, a SF resident, look at a book today on Amazon.com. Based on your past actions, Amazon knows very well how likely you are to purchase right away, or tomorrow, etc. [...] They'll ship the book to the warehouse in case you come back tomorrow to look at that book.
It's unlikely that it would be worth doing this if one person looks at a book; the return-and-buy-later rates wouldn't be high enough. But if ten people look at a book, it makes sense to move one or two copies to a closer warehouse -- and that's where Amazon's scale really wins. With O(N) customers and good historical data they can predict future purchases to within O(sqrt(N)) thanks to the central limit theorem.
> It's unlikely that it would be worth doing this if one person looks at a book
Yeah... But then again, if a truck is leaving soon from warehouse A to warehouse B, why not put that book on the truck, if no one near A has looked at it recently?
There IS a marginal cost to touching a book and to moving it, even if the truck is going ANYWAY, it might be pennies of fuel + pennies of labor to move it. This could eat into the margins.
Even then if you really have data for everything, you can take into account, "are we overstaffed in warehouse A right now? what are the staffing costs in warehouse B when we arrive?", and do more or less physical work depending on that information
It's unlikely that it would be worth doing this if one person looks at a book; the return-and-buy-later rates wouldn't be high enough. But if ten people look at a book, it makes sense to move one or two copies to a closer warehouse -- and that's where Amazon's scale really wins. With O(N) customers and good historical data they can predict future purchases to within O(sqrt(N)) thanks to the central limit theorem.