Thanks Euler, truth be told, I'm a bit envious of all the tools this generation has available at their disposal. The Red Queen is alive and well-- and for the poster that mentioned kurzweiil, you better hope your wetware plasticity can keep up with the exponential learning curve. Sucks getting older.
This is a very important concept. Most profitable trading strategies are quite straightforward... Usually some variation on classic Market Making or Relative Value Arbitrage. But what makes it profitable is EXECUTION... Infrastructure, low costs, discipline, etc. If you cannot explain in 60 seconds... (a) what specific "market inefficiency" you are exploiting and (b) why YOU have this competitive advantage... Then you have nothing.
I agree with this comment. You would end up overfitting the past, that is fantastic results with past data and vast losses in real trading. This because this is not a realm where NN can be applied meaningfully. Here there is nothing to predict, simply because the underlying phenomenon is unpredictable. It's similar to when you compute a regression between 2 unrelated variables: you sure get a number for the correlation coefficient, but its meaningless... Tommaso ------------------------------------------ The G-BOT Project - Automated Day Trading System http://www.datatime.eu/public/gbot/
I have read and I like this thread and agree with what is said above. Which technology to use in order to optimize your model is only one part, and not the most important one. Since we use both Genetic Algorithms (GA) and Neural Networks (NN) in our trading systems I may have some input on the issues with using NN:s. The major problem with all optimization for trading is, as been stated many times, the problem of overfitting the model parameters to the data. Specific issues we have found with NN:s are how to create a good error measurement (at least for back-propagation networks), how to create a goal function that makes sense for trading and how to select a training algorithm. If all of these are not created carefully you can be sure that the optimization algorithm, whatever technology you are using, will find ways to exploit any problem with your model in order to satisfy your goal function, often leading to bad results. One simple example is that the optimization algorithm could become obsessed with including a few very good trades in the output even though the results are purely news-driven. This means that the NN training will try to isolate a single super trade by tweaking the parameters of the solution space no matter what. This will in all likelihood not provide a good generalization of what constitutes a good trading opportunity. All in all, it may not be the NN itself that is the problem, but the complete optimization set-up. We found that some working non-NN trading models we created could easily be converted into a standard feed-forward NN. Now, when we use NN in our hybrid models we use the GA for training it. GA:s are bad at training a NN, but it gives us larger freedom in creating our goal functions. GA:s have problems of their own, but that's another story... /Hugin
Thanks. Are you using the GAs primarily for NN weight optimization? If not, do you mind if I ask what your GA fitness function is optimizing? ----------------------------- P.S. To BigDaveDiode. That book you rec'd was fantastic! Best book rec I've seen on these boards. Best practical book I've seen in a long time and way ahead of its time considering publication date. I was able to get the used version for $6 on amazon and the software is fully functional with PCA, LDA, morlet wavelets, ARMA, filters, etc in DSP/time domain. Very good. Way better than any of ehlder's stuff. The only drawback is if you don't have much experience in basic multivariate stats, NNs, and time series; it's a bit more of a practical viewpoint from someone experienced in these areas. Don't expect to be bogged down in mathematical details either; written by a mathematician, yet very hands on and practical. How did you find that book, dave?
Don't get it without the CD. The CD is imperative. That is the best part of the book. Also, don't worry about DOS or NT specs, it works fine on XP. Only drawback is it is a bit slow on training nets and it is not a simple and intuitive GUI interface, but pretty easy to use once you get used to it. I haven't played with the step size parameters, yet, but the NN examples are very slow to converge as they are.