There is a great explanation in "Stereochemistry of Organic Compounds" by Eliel, Wilen, and Mander. I'll try to summarize:
First, draw a three dimensional representation of your molecule. Next, superimpose three dimensional coordinate axes onto your molecule with the origin at the carbonyl carbon and one of axes coinciding with the C-O bond in the carbonyl group. These are your octants. If you were to look down the axis that contains the carbonyl group (with the O in front and the C in back), you would see what appear to be quadrants. In general, the 'front' octants in this projection aren't needed, so they are ignored. The four rear octants are labeled with the top left and bottom right octants + and the top right and bottom left octants -. If most of the steric bulk (or more accurately, the most polarizable group) resides in a + octant, then a + CD is expected.
This seems to work pretty well if you are sure of the most favorable conformation of the molecule and the distribution of the bulk is unambiguous. In conflicting cases where there are relatively similar polarizable groups in both a + and a - octant, then the predictions are not as good, although you generally still predict the correct sign of CD, the magnitude is often small and therefore kind of unreliable.
Here is a slide that I made for a recent presentation to our group. I hope this helps. The reference to the 1986 Tetrahedron review is also very useful and has a lot of examples.