Being too easy to predict may break the blackbox nature of the PRNG (you might "predict" it by accident).
For example you can increase the quality of a RNG sequence r(n) by adding a multiplicative hash, r'(n) = A r(n) for some good constant A. A consumer treating r' as a blackbox expects B r'(n) = AB r(n) to be a good random sequence, but if B is close to the inverse of A, the constant AB may no longer be "good".
For example you can increase the quality of a RNG sequence r(n) by adding a multiplicative hash, r'(n) = A r(n) for some good constant A. A consumer treating r' as a blackbox expects B r'(n) = AB r(n) to be a good random sequence, but if B is close to the inverse of A, the constant AB may no longer be "good".