It is a very old thread, but I think some guys here seem to be able to use machine learning algos profitably but on the other hand do miss basic things about how the Z strategies work...

The Z strategies clearly state that there will be updates to the trained parameter sets in new zorro versions. This in fact means that the parametrization of the strategies are adapted to the current market behavior. Although WFO will basically help keeping the previous history record, it is not guaranteed to stay the same as WFOs are expanded at each training due to its' fixed segmentation logic. This is much more true for the latest WFO segment which provides parameter for actual trading.

As a logical consequence, the backtests will change naturally due to the parameters which are expanded in test time and optimized for the CURRENT markets and not the markets in 2015 or 2010. You may want to read about WFO, forward testing and market inefficiencies to understand the reasons.