GreedAndFear, do you have a way to quantify when it will be "good enough" to go live?
The ideal would be to be profitable in every simulated session (=recorded market data) or at least be 'significantly' profitable overall.
Also, there should be enough different kinds of market sessions in the DB to assume/proof that it will be successful in all those situations. I have several wild whipsaw-sessions from the 2011-2012 period, but the algo should also perform well during a flat session and not overtrade.
For now I'm mostly working on my neural network for generating daily trading signals, which is working just fine for the last 2 year (actual trading). I may not even need the algo in the end, but it's just great fun building it ;-)