Well, not to split hairs or anything, but if you can do a smaller integer it will alway save the disk space over a larger one, and each additional byte in a billion-row table is a gigabyte of storage.

The necessary precision should not be driven by Netezza's constraints but by your business rules. So if you need the additional widht or precision, just do it. The compression will cover you either way.

Another aspect is that float does not compress. So stay in the numeric if at all possible (these are just scaled integers anyhow).

Another aspect of your data-size testing - you need to use a very large sample to test with because Netezza has a default-minimum database size that will be allocated no matter what the size of the table, so you could get a false positive (that they all appear the same) because your sample is too small.

The only thing you would need to test, past all this - is to make sure that your BI math doesn't spontaneously combust because of overflow. Don't jump to convert them to float until you read my blog entry A Tastier Float here on the Netezza community (Gather 'Round the Grill)