Quote from braincell:
Good point.
My opinion, and experience is that people misunderstand what "machine designed" really is. It's actually human designed, with a search algo that does what a human wants it to do. It's never going to be as simple as "click and wait for money", no. Using machine learning is as complex and requires as much know-how as manually designing and testing strategies. The only benefit is that you can test more hypothesis and relationships more quickly. Often, the machine learning will give you ideas that you can modify slightly and improve upon.
PAL is just one method. Using indicators is another, but considering indicators is just a derivative of OHLC, you could in theory claim that there should be no difference in results if using patterns or indicators. However due to market dynamics, and the fact that it's a time series problem, there are. In addition to patterns and indicators, there is a lot of data that can be incorporated into machine learning, and tested with statistics for it's trading value. For example, a synthetic instrument as a derivative of a basket of correlated instruments, and then applying cubic spline moving averages onto them. No existing machine learning product really lets you test these ideas easily, so it's really hard for anyone to make a complete claim - does it work or not. There are so many different approaches and small changes often mean a big difference. Like jcl said, he did testing for patterns, but probably missed a crucial detail. In my view, it's best to transform manually designed strategies (hypothesis) into indicators, and let machines try to use them.
Nobody will ever have a straight answer, imho.
