Hi many thanks for the replies.

My immediate thoughts are that if even a simple model like this is so sensitive to small changes in parameters, it will be hard to gain confidence in the robustness of *any* (algo) strategy, especially more complex ones.

Presumably different classes of strategies are more/less sensitive to initial conditions (and over-training)?

(Also, @simonkrebs please could you post your code that ran the test? I would help me get my head around a few things!)