Lotto ball machines work well to randomize ball picks because of their mixing properties.
A closely related problem is how many shuffles are sufficient to randomize a deck of cards.
The following curve is pulled from an AMS (American Mathematic Society) post on this topic. A link to the post is given below.
https://www.ams.org/publicoutreach/feature-column/fcarc-shuffle
d is a measure of how close the pdf (probability distribution function) of the card ordering is from uniform random. k is the number of shuffles. The curve shows that after 17 shuffles, the cards are very uniformly randomized.
Is this game over for lotto predictors? That depends on how well the mathematical mixing model models a real world lotto picking machine.
During my investigations into PB and MM prediction, I've discovered that you can about double (over uniform random expected) the probability of a correct ball pick for the powerball/megaball and the first white ball (sigma>5 confidence). More interesting to me, the prediction formula is the same (up to ring size) for the first white ball and powerball. For the PB, I'm getting 7.25% accuracy (NumTests = 455) (expected is 3.85%). For the first white ball pick, I'm getting 2.86% accuracy (NumTests=455) (expected is 1.45%). Since this initial discovery, I've discovered another powerball predictor which gets 9.77% accuracy (NumTests=430).
Note that even with these improvements over uniform random picking, PB and MM are still big time losing games for all but a relatively few lucky players.
For white ball picks 2 through 5, I have not discovered anything to help prediction (at least, anything with > 4 sigma confidence).
In summary, my investigations to date show that PB and MM ball picking machines are sufficiently random. I have not found any statistically significant advantage for 4, 5, or 6 ball wins.