Hi ak47,
The prediction should take no more that ~15% of the total time of the movie. However, if the final CQ will fall in the unlinear range (varies if CQ or CQ_VBR), then it can take WAY longer, because there's just no way to calculate
So the best thing that CQMatic be do, is start bracketing down (or up) on CQ values, until the target is reached. In another way, say I want a target sample of 10MB. I encode with a CQ of 50 and I get a 20MB sample. If TMPEG's CQ was linear, then simply dividing by 2 (50/2) would give me a 10MB sample, but this is not the case, and I can wind up with 14MB. So you see, it's a real pain to calculate the correct target, and every CQ value creates a different file size completely out of proportion.
So the best you can do is ~2 passes, to a max of "I don't know", depending on your material.
-kwag