I wrote a little Python program that lets you try it with an arbitrary integer. It assumes that for instances where you have to lay out a zero that you need one bit (for 0).

This times out quickly if you use too big of a number, but you can play around with integers here: http://codepad.org/4mnHI9E3

Based on what I see, it compresses poorly at the beginning and then reaches parity. At 1,000,000 (tested on my machine, not Codepad since it times out), you get one bit of profit, but 1,000,005 is at parity, suggesting that it's not very stable.

Also, this is assuming that the decompressor can make sense of the absolute smallest representation you could lay out these numbers with, which is basically not going to happen. That additional padding would destroy any potential savings you got.