I don't think it works just as good as in backtest.
But the past is reality.
If your algo captures every movement in complete asset's history - there is no better bet. Unless you are talking to time travelers.
Sure - some rules that worked in past might hinder in future. Which is why I'm pro-tinkering.
I find it similar to software development - if your code is scary to touch, contradictory to your feelings, you want to bash it more until it isn't.
Every bias, every fitting is an improvement at particular point in time. Trick is to align them so they don't interfere.
Speaking of ridiculous numbers - those are rookie numbers.
Once wrote a script that iterates backwards in time, captures everything in generated code. Theoretically speaking - you can get to gazillion in couple days starting from couple bucks. Slightly less as I didn't include fees that time but you get the idea.
