Alex, that is an excellent start; I hope you continue the series. One problem in genetic optimization I have run into is where a parameter with little or no efficacy is associated with a set of values that are the top performing group in the generation. This can cause a number of what amounts to the same value set to be promoted to the next generation, squeezing out alternative sets.

i suggest exploring the hybrid approaches. they are combinations of a local optimizer and a global ( in this case a Gene optimización technique). one very useful that I personally use a lot is particle swarm optimization.

Scott, according to my experience parameters typically can be sorted into classes by their "importance": simply put, more "important" parameters affect the performance more significantly than less "important". Then most of the time if you run genetic optimization including parameters of all classes, you get long series of results with just one value of an "important" parameter and numerous values of "less" important which essentially don't affect the performance. So, I solve this problem only by understanding why this parameter is more important than another from "physical" standpoint if you want. Otherwise if I can't suggest a reasonable explanation, I'd rather stay away from such a strategy.

True Alex .. I am using MC for 6 years.. same time genetic options bring same not logical values.. and forward testing in this case is a must..

I agree with you on understanding your parameters; in this case it was when I was working for a fund that liked to throw things at the wall and let the optimizer sort out which parameters were "important". I consider that experience to likely be the genesis of why I am so adamantly in the camp of understanding my trades vs. something akin to machine learning.

@Alex....I see that what you call a parameter is the feature of the system. I see a parameter as a part of the feature, such as the period component in a sine wave.

https://math.stackexchange.com/questions/751781/how-to-parametrize-a-curve-by-its-arc-length

Importance weighting is a form of dimension reduction.

@Scott..Optimization tools can be used as an exploratory fashion to discern latent characteristics. Start with unique parameter sets to apply to each technology. Clearly, the genetic family is in the evolutionary group. Utilize a single technology and change the seeds or other inputs. Compare outputs of different technologies. We know the pros and cons of each technology, so any significant differences are revealing. They can also be used to model the physics of the phenom.

@Alex...Maybe we could persuade you to explore how multicollinearity affects and is expressed in the genetic family of EAs. Thanks a bunch.

Hi Alex, thank for your interesting article. It covers once again the issue of "optimisation" of a strategy. Optimizing an existing strategy implies that you are no longer happy with the one you are trading. But when is that ? When do you decide to optimize (and throw away the strategy with its parameters you were so confident of that you were actually trading it live). How much discretionary judgement is allowed or is this optimization process also a fully autonomous function of the strategy. Should some underperformance not be allowed as we know that drawdowns are an integral part of trading ? Thanks for your views

John, I actually treat these terms mostly like you do. To use your example, for me period, phase and amplitude are parameters. But when you work with a mixture of sine waves then periods (frequencies), amplitudes and phases make totally different impact on the resulting waveform. This is what I call a "parameter class". I understand that my terminology may be different from a standard because I am not a professional mathematician and especially in this domain, so if you could suggest a better, official term I'd be grateful.

Marc, there is a great confusion which appears once and again, and it relates mostly only to terminology. Optimization is simply an iteration over all (or a certain subset) of parameter values combinations. Therefore it can be used as during live trading, as well as during the r&d phase. Thats why I prefer to say "genetic search" instead of "optimization" to emphasise that we can use the same tools (the same platforms, the same backtesters) during the research phase.

@Alex - I do remember a product like that but never used it in any way. I once wrote something like what you described - it was a bit mask over a 32 bit integer and by simply turning bits off and on via a random number generator you could mimic a genetic algorithm (not quite that simple but close).

I prefer to make my money the old fashion way - don't do stupid stuff and you get to keep most of what you win. Let randomness work for you.

@Scott, when you say randomness.. what do you mean? A random system is supposed not to work..

@David, I think of signals in the negative; I can enter almost anywhere, in any direction as long as I don't have a signal in the opposite direction of the entry or a "don't trade in any direction" signal. My particular emphasis is on when not to trade vs looking for an entry signal.

For example, I measure price stability (it isn't important how) and if price is unstable, the algos don't trade until it becomes stable again. At that point, a random entry in a random direction will likely move 1 - 4 ticks in my direction just as it will likely move that much against me. By trading with a slightly negative risk/reward ratio and very exacting trade management sub algos, virtually any trade can win. That doesn't mean every trade wins, just that if you eliminate the ones that have little chance of winning, it leaves you with a good trade set.

I decided to use genetic optimization as the main topic for discussion in the upcoming webinar, anyone interested in attending please register at https://goo.gl/Vpj0Bg

Site down

Which one?