Hours of testing, with about 100 individual tests later, and it seems the issue has at least two different causes:
1. Deleting all training related data files *.par before changing DataSplit and running a new training session sometimes has an influence on the outcome.
2. If DataSplit is not set, it seems to change its value from 0 to 85% at some point. The problem surfaced because, I used its value in calculations at a point in the program when it was still 0, it later changed to 85%. This explains at least some of the behaviour.
There is a clue about this inconsistency in the manual - which statement is the correct one? Walk Forward Optimization
If DataSplit is not set otherwise, WFO uses a default training period of 85% and a default test period of 15%.
DataSplit
Splits the simulation in a training period (given in percent) and a following test period. F.i. when set at 60, the training period has a length of 60% and the test period has a length of 40%. This works with or without WFO and ensures that the test always uses out-of-sample data.
Typical range:
50..90 (default = 0 = no separate training / test period).
Conclusion: Something needs to be sorted out by the Zorro Developers and/or manual author. Or maybe I am just not understanding things right...