Kevin, I think you have the spelling wrong. It is Gene Simmons. He was with KISS and, more recently, on the Celebrity Apprentice. crisis
Nothing set in stone, just looking for a positve expectation, low relative variance, and low or negative correlation with currently traded models. If I could do an 80% win rate in liquid futures with losers not much larger than winners I would be worth $ billions in very short order. I would not be posting on ET. Hence my skepticism regarding your claims. I favor kernel density estimators. Several types of neural nets can be configured to act as KDE's. Another poster on this thread mentioned support vector machines. I would be very leery of that method, You are basically building your classifier on extreme values (the outliers) in your training set.
I'm going to call bullsh*t on you. I've worked on projects relating to almost all the keywords you've used, and they don't fit together the way you seem to imply they do. Neural networks can't generate code. They are just big curve fitters. GA can't generate code either. It's just a search strategy. Maybe you mean EA (evolutionary algorithm)? I've done those to. In fact, for the very purpose of generating predictive codes. It takes many very large, powerful computers several days to come up with anything resembling coherent attempts. You can do it in real time? Not a chance. Moreover, if you can somehow do all of this, you can't figure out how to use IB's API? That's just not internally consistent with the skill level you represent.
Kevin, OK, thanks. Ones eventual net profit is more correlated to trade frequency than W/L ratio. If you can trade every 10 bars at 60% you'll make more, depending on the impact of drawdownâs, than trading at a much higher percent but only every 500 bars. You might have noticed in developing models that as you push the accuracy higher you reduce trade frequency. It's a bit like an economic version of the Heisenberg Uncertainty Principal in Quantum Physics....the more you know about a particlesâ location the less you know about its velocity and visa versa. I've never worked with KDEs, so I can't comment on them. Jerry
Sjfan You said that youâve used predictive software on projects and said "Neural networks can't generate code." What kind of projects, in what industries, and what grade of software? I'll grant that an application you buy on the Internet for a few hundred dollars or an Excel Add-in version can't do much but rest assured of two things: 1) Large corporations do use NNs for certain applications, like making a calculation of the probability of a credit card transaction being fraudulent and deciding to approve it or not. 2) For this kind of time critical, distributed application they do NOT have a Geek in cubicle J-2534 type in purchase specifics and then click the mouse to fire up a fraud modeling application to run a NEW NN to decide on approval for your credit card transaction. I can say this as I used to work for a company that created the NN apps that major credit card companies use to screen for fraud. Now that was a few years back. Perhaps in a cost cutting move they may have recently dumped the architecture of distributed high speed transaction processing using defined fraud models in favor of hiring several hundred thousand illegal alien type Geeks to sit in sweat shop cubicles in some third world county to run a new fraud model for each credit card transaction.....but I doubt it. Jerry Jerry
How very presumptious of you. They were proof-of-concept forecasting projects, in finance, custom built apps designed to see whether any of this was practical. It's one thing to use NN for fraud detection. Sure. Facial recognition, no problem. It's good for stationary patterns. Non-stationary time series? not a chance. And in anycase, it doesn't generate code, just a bunch of weights. You want the thing to be 1 Gig? Seriously? You specifically said execution code. Even if the code is machine generated (via your GA, which I very much doubt), it can't be 1Gig. That's a LOT of op codes. More to the point, if you in fact do have most of these details figured out to any degree, how to use a simple IB API is the last question you would have. I don't believe you. I think you are a liar.
Sjfan You said when I asked what your background was: "How very presumptious of you." *** I was just trying to figure out your level of expertise as you didn't appear to have much understanding of how Data Mining is done in sophisticated implementations in the corporate or government arena. You Profile on Elite Trader is pretty empty so I couldn't get much from that. *** Then you say: "They were proof-of-concept forecasting projects, in finance, custom built apps designed to see whether any of this was practical. " *** Well there is you answer. The custom built app was apparently not designed to generate output so that the results of the modeling process could be used in a production mode by another application. Do you really think that tool like IBM Intelligent Miner, PSS, S-Plus, SAS and a bunch more only function so statisticians can model and explore data sets and look at nice graphs and charts but have no further use? Like save the results so they can be used by the rest of the corporation in a daily or minute by minute basis to process new data. Here is a bit form an S-Plus web site: "Deployment and Scoring Web-ready graphical reports HTMLmodel summary exports Non-interactive batch execution of all components (Server version) Model ports support automatically-updating scoring components Score custom predictive models created using S-PLUS on very large databases Predictive Model Markup Language (PMML) model import and export Generate C code for run-time model scoring (Server version)" I draw your attention to the last line "Generate C Code". They seem to be able to do this even if the application you worked on could not. We can all learn something new every day, I guess. *** Then you go on to say: "It's one thing to use NN for fraud detection. Sure. Facial recognition, no problem. It's good for stationary patterns. Non-stationary time series? not a chance. " Do you say that because you tried and failed or because you know it to be true? Do yourself a favor before making presumptuous sounding statements, Google the terms "neural network" "time series" "stock market" and "pdf". With the pdf in term you'll get a lot of detailed academic studies on the subject. Some of which will mathematically disprove the Random Walk Theory of market behavior. If you like I'd be more than happy to educate you on this: give me a way to send you files and I'd be happy to share part of my research library on the subject. *** Then you say: "And in anycase, it doesn't generate code, just a bunch of weights. " *** Then you should call up the major credit card companies and tell that that the software provided to them for credit card fraud detection by companies like Fair Isaacs and other firms is a massive fraud since the NN fraud models they though they were using in their transaction applications don't actually exist, since it's impossible to convert a series of weights into code. You might get a reward for uncovering such a massive and long running fraud! *** Then you say: "You want the thing to be 1 Gig? Seriously? You specifically said execution code. " *** Correct in terms of a trading system. That includes a whole parliament of models. I've given the details a few days ago on this thread. Please reread as I'm not going to retype it. *** Even if the code is machine generated (via your GA, which I very much doubt), it can't be 1Gig. That's a LOT of op codes. *** Yes, but Neural Network, I don't know where you got GA from *** More to the point, if you in fact do have most of these details figured out to any degree, how to use a simple IB API is the last question you would have. *** Why? What does a predictive model have to do with the best application to use for automated execution? *** I don't believe you. I think you are a liar. ** Considering you understanding of how advanced software is used in distributed applications, time series forecasting and how to covert weights into C++, I'm not surprised, but at the same time neither worried nor offended. Somebody once said something about any sufficiently advanced technology will appear to be magic until the observer can learn enough about it to understand its basis of operation. Why don't you download the demo version of a high end modeling application, run a NN model and then export it to C++ yourself, and then you can say WOW it does work! Or in terms of time series predictability, send me a time series data set that you have tried to use to predict future market characteristics and found it impossible, I'll run it and return the model output. As Bob Dylan once said "if you're not busy born born (learning) youâre busy dying" Jerry
sjfan, The academic studies section at this site should get you started on using neural networks for time series forecasting. http://www.optimaltrader.net/neural_network.htm#studier Let me know when you are ready for more material. Jerry
Good Response Jerry, I'm relatively new to this site, though I'm a long time trader. I've noticed a 'mean spirit' to a few of the posts such as the one that started this thread. At the end of the day we can disagree without making it personal. Your taking the 'high road' convinced me to stick around for a while. BTW it was Einstein who made the magic remark about technology.