haha, I see. Could you elaborate on what you mean by the 'market-model side?' Like creating actual trading strategies?
I have a webinar on my website on the systems page, in which I describe the market-model used by CL151, and how I used it to identify & qualify "pattern 151". You can also search for the webinar I did in July with NinjaTrader, it is in the NinjaTrader add-ons partners channel on YouTube. This will give you one *example* of a market-model - of course, there are many other ways of modeling the market.
An example of a "legitimate economic phenomenon"? Trading system performance usually depends on serial correlation and doing k-folds may destroy that resulting in unrealistic results. See trend-following systems for example. It may work for medium frequency but still it is a dubious concept because the problem with trading systems is not validation but non-stationarity. Here are two options: (A) Train rats or monkeys to trade for you: http://www.artmarcovici.com/rat-traders (B) Precognition: http://bit.ly/1yHuOt2 http://bit.ly/1trXjTq
An example of a legitimate economic phenomenon could be, for example, the historical cointegration relationships among interest rates. With all due respect though, I'm not looking to design a valid strategy right now, I'm looking at designing the most rigorous framework in which to test any strategies. I will say though that most of my hypothetical strategies in the future will probably revolve around trading weak form stationary relationships though, so if you could explain to me a little more about why doing k-folds is dependent on serial correlation I'd be interested in understanding that a bit more. I don't know a lot about this technique, it's just something we're doing in my machine learning class. Thanks.
Modern markets are distorted by policy makers and the Fed, and moved to abstract space that has very little correlation with legitimate economic foundations. It makes market analysis more interesting but also quite difficult. Since the pace of this market distortions accelerated recently past data and known correlations may not be very helpful. Economy is still economy and will come back knocking on the door at some point but dealing with it using logic may not quite work today. At this stage I would not worry about over fitting but rather with getting better understanding how markets work and find set of quantitative tools that would be in sync with the markets. This would narrow list of testing hypothesis to work with and in turn make design goals of testing framework more concrete. Whether you will use statistical methods or machine learning or something else it seems to me that all of it should be analyzed from the perspective of probabilistic models. Single method approach will not work in my opinion. This would be one way to look at it and obviously there are different approaches as well. Few remarks on CUDA. There is C++ AMP available that can be used with native C++ and is a little easier to use. It is C++ not C. Consumer graphic cards are actually quite good and there is little advantage to pay for high end scientific card. Single precision operations are slower and double precision are very very slow plus small part of GPU is knocked off with fewer cores available. One way to go is to use integer operations (moving decimal point beforehand) whenever possible and the few remaining floating point operations that would be necessary will not slow program too much.
I tend to agree with you about the multidisciplinary approach. Going back to the type of trading opportunities that interest me you can see that they all pretty much originate from econometrics. Obviously thats not ebough though. Statistics and math make up the framework of any model, and I find machine learning most useful as an set of tools for optimization and problem solving. There's overlap and other sciences that can be applied to markets obviously (as bashatraders rat example demonstrates), but I'm just a pure math and CS major (+ a stats concentration). when I interviewed with that HFT firm they told me that no one in the industry really uses machine learning (and I quickly dropped the subject haha), but I think a) machine learning is a broad and ambiguous term, anyone who computes an MLE is basically doing statistical learning in my book b) there's more to trading than just HFT. That being said, I think machine learning probably can be one of the most dangerous approaches to trading in terms of the probability of being "fooled by randomness." Like I said, I think machine learning is cool and applicable, but personally I see it as more of a tool to be used in conjunction with others as you also alluded to.
Sorry dom993, something went wrong for sure as those were cjbuckley4' quotes. BTW I like your website. Very informative.
True, the most dangerous and for this reason the most lucrative too if used properly but only 1 in 1000 users know how.