that priceaction post doesnt make any sense. why would price action not generate curve fitting? its the same kind of static data.
That 'priceaction' blog is always getting quoted as some sort of gold standard in terms of the pitfalls of back-testing. Let's face it, if the guys product did what he says it does, then why bother to sell it? Just use it to make money, no need to deal with irate customers or write a blog. I think Occam's razor applies here.
why would you not start out in R with a proof of concept and only when the blueprint and rough sketch looks worthwhile to pursue, only then would you pursue to code up a system on such scale in C. It makes no sense to get a bunch of programmers together with no one having a keen understanding of genetic algorithms and a very specialized knowledge in machine learning. Programmers alone unfortunately are never the ones that make the money and never the ones that bring aboard the crucial skills and knowledge to really get a product off the ground. It is the marketers (-> Facebook & Co), traders (->algo trading) that you need initially, not the programmers. Just my 2 cents
exactly, and domain specific knowledge here is not programming skills, it is someone with an expert knowledge in statistics, math, and machine learning. Programmers are completely useless at the stage you are at.
Very interesting discussion. Does anyone know rich statistician or mathematician who made fortune in the markets? In my experience the only readable texts in stats/math/ml are written by people who also know programming. Otherwise they live in their own abstract world trying to invent next great formula that explains everything and produce logically convoluted texts with the sole purpose of impressing fellow statisticians or mathematicians and hiding true logic and simplicity of given solution to the problem. I would rather prefer programmer who specializes in algorithms or has background in science. I like the idea of using R (or Python) for initial proof of concept instead of going C from start building GP algorithm from scratch. Once the project matures existing libraries might be too limiting and it warrants rolling your own or modifying existing library. It seems that R is extremely simple to start with (simpler than Python in my opinion) and both have extensive libraries. It is possible to integrate R with C++ (rcpp). I agree with others that using GP might not be very rewarding for market data/time series analysis but it is worth a try in step wise fashion instead of venturing into big software project from start. By the way why limit development to C if new C++11 has async and threading as part of the standard library. In addition high performance computing is done mainly on GPU so why go OpenMP, Pthreads, Cilk route?
yes, quite a number, but you could start with James Simons of Renaissance. Pretty much any top trader has an intricate knowledge of statistics and probabilities. But to create quantitative strategies that do not fall apart the second you trade them live requires more than just an understanding of probabilities and statistics. You clearly do not understand much about the underlying math, so I repeat it does not matter how good a programmer you are, your excitement about C++ 11, GPU computing, parallelization, and the like does not help you whatsoever in solving problems through genetic programming. Machine learning is an equally misunderstood concept that some in the retail crowd suddenly seem to be drawn to like flies to a pile of sh...judging from the many blog posts, it seems that everyone nowadays can easily become a quant. Running couple commands in R and actually truly understanding what one is doing, math and stats wise, are entirely different worlds. And I claim you require the latter if you want to create profitable trading algorithms.
The discussion was not really about intrinsic math/stats knowledge which is obvious but rather how to use available tools and programming technologies in design phase. Computer algorithm design and implementation is very tedious and difficult in team setting. Many projects fail because of poor planning and design in initial phase. Renaissance had good returns in golden years of bull market but their strategies do not perform good in in today's market. I wonder how much was it marketing of models created in house by scientist and run by computers plus old time tested reliance on insider information versus actual performance of mathematical predictive models created by them. I agree that creating profitable strategies will take more than using preexisting libraries but using those provides valuable insight into data, speeds up algorithm design and refinement.
as it's been said above, starting on a high level language like R, Python, Octave, etc... and putting together a working prototype of the system is probably the best road. If it works you'll have a clear blueprint to build into C and to know which parts you can run on MPI or OpenCL... If it doesn't work, then you can discard it at a low cost (in terms of hours put into the system)
the project is set to fail because there is nobody with a keen understanding of machine learning and genetic programming experience. Its like you want to start a lobster farm but do not know what a lobster is, does not matter your funding, and how much land/sea you can secure, you gotta know about your core product otherwise you are set to fail. Re Renaissance, please do not publish conjecture or theories, you clearly have no idea of the inner workings of this firm nor its performance metrics. Even hinting at the possibility of insider trading is ludicrous.
It is naive to thing that Wall St. makes money using superb skills of professionals in the field of math and stats. All those scientist have keen understanding of theories developed on completely different set of data and their experience is not applicable to market data directly and usually does not work in the long run. As to Renaissance I did not publish anything but rather asked valid question how this performance has been achieved in view of non-consistent returns in all market conditions. As we all know there are many ways of making money by hedge funds and the strategies become public only if someone gets caught. An example of one guy with math degree out of hundreds of thousands as a head of hedge fund does not prove much. Again there are no rich statisticians or mathematicians and those who managed to be lucky for some time were discredited later when market did not follow their models. Most of the "science" is used as cover up to show it to regulators or investors that there is risk management in place or as marketing tool. You need practitioner with analytic skills able to implement results of observations using science based tool set and not theoretical scientist. The project could fail if the initial hypothesis of achieving market edge by using GP is false. It also can fail by targeting wrong data sets with incorrect methods or the project is poorly designed/structured but definitely not because participants have no PhD degrees