This article was really eye opening when I read it a few months ago.

I've always hated menu dropdowns where it would close if you try to skip to the next submenu you've opened, since you hovered on the other parts on your way to the submenu.

So you'd end up having to perfectly 'trace' the menu so it won't close down on you.

That article says nothing about preloading pages. The speed gains are purely a UI thing -- if you're clever about keeping submenus open even when the mouse leaves the submenu's entry in the main menu, you don't need to provide a delay for menu opening/closing.

I think the Amazon example was being used as an example of the "directional cone", not preloading.

A few years ago, I went to a tech demo/hackathon thing. Somebody had some neat experiments with preloading pages speculatively based on user behaviour

The coolest approach given was when a user types in the search box, kick off the request for the full results page if they move their mouse on a motion curve towards one of the results, with some machine learned data on how people move their mouse: Not quite in a straight line, and with some acceleration.