That is overhyped marketing gibberish. Genetic algorithms in production are not all that complicated math wise. Nor is it rare to use this approach for parameter optimization. That website makes it sound esoteric to make their products appear more special. As the previous poster pointed out the risk of overfitting is the highest risk in this area and one should really understand what he is doing. Just in order to apply ML is not enough justification to implement it.
It's funny how many people here seem to miss the entire concept of optimization. Whether you use a simple parameter sweep, a proprietary system like this (why would I trust them?), a genetic algorithm, or some other esoteric optimization solution doesn't matter as long as you find the max/min accurately. I wish they would require some kind of book on linear and/or nonlinear programming for this forum. That's obviously a completely unrealistic idea, but you'd see a lot less people getting tricked by stuff like this.
Well...what to say...in the end everyone else makes their own choice. Whether discretionary trader relying on Technical Analysis or a quant or algorithmic trader relying on optimization techniques. Some stick to moving averages and others of granma's favorite tools on the TA side and the same blindly optimize the heck out of parameters. Both approaches are futile imho. The successful are the ones going a step further and actually trying to understand the goals and underlying math and avoid pitfalls. But back to the topic: I came across a product once that features spinning off uncountable backtests with the goal of generating trading strategies entirely on its own. No user input needed. That package made heavy use of optimization algorithms. Guess what. The system produced results with awesome statistics over the out of sample period. However when I more deeply investigated it turned out that not one out of the many hundreds of strategies, generated, was robust. In short, the system produced over fitted garbage.
A very good HFT I know gave me some advice. Always find a fundamental justification for why a trading strategy works. If you're making money, it's because you're participating in price discovery, not because you've found an optimal combination of parameters. He also uses optimization as an overlay on top of these fundamentally justified strategies to "get closer to the right answer" in his words. It's the approach I personally take as well. Start with a relatively concrete idea about a fair value for a security and then optimize the parameters that govern the strategy across the subspace of what's acceptable for the trader. What you say about out of sample performance is true. You need to cross validate results and test them out of sample. It's really easy to come up with historically profitable but overfit strategies. So easy, a set of nested for loops can do it.
Given the Processor time, just live back test the last 30mins or whatever, short as possible, what's working best, with different methods and play with the variables, has the best shot of working well next, obviously changes will happen and so will losses, but should still work out fine. I messed with this 9years back, MT3.51, too damn slow, MT4 is better, some C language would ofcourse be better and a new I7 type processor. Anything above these requires AI and personally I don't think we'll ever truely get AI, just simulations of, which are all preprogrammed outcomes only, sadly!
As I stated at the outset of this thread, most systems that I've looked at offered only static solution. Nearly all were merely walk-forward optimization of manual user input of parameters. However, some, like Trading System Lab, were more unique in that it does generate strategies "entirely on its own" based solely on profit potential. My ideal solution is a dynamic one that can "evolve" with the changing market. While that might sound too damn futuristic, it really ain't. Basically, I would like the system to cherry pick the best algorithm (from the basket of strategies) for the market condition and then create a if-then scenario based on the past N bars to come up with the most optimized parameters.
I bought biocomp dakota somewhere around the year 2006. I was not successful in generating anything truly outstanding or interesting. Maybe it was just me. Perhaps there have been improvements since then.
Sorry guys, none of our regulars are at fault, but no one is permitted to come here and violate our No Free Advertising policy by deliberately posting sales information, tech support, etc. His posts were removed along with any that quoted him.
At least snake oil salesmen have to cough up dough to Baron in order to spread their pest among us. But back to the topic: there are actually multiple libraries targeting different languages that implement various genetic algorithm techniques. If one is familiar with coding then those come very handy. I recall that Amibroker implements some simple genetic algorithms for those who want to tinker around with some basic stuff right out of the box.
For current but for future? This is the problem. If you have no validation segment the risk of over-fitting is high. http://eranraviv.com/sample-data-snooping/ http://www.priceactionlab.com/Blog/2012/06/fooled-by-randomness-through-selection-bias/ http://stats.stackexchange.com/ques...-in-sample-and-pseudo-out-of-sample-forecasts