First, fixed the computer running the test by restoring the operating system to a point prior to the last update. Wonder if it was something I did or whether Microsoft will issue an "oops..." statement at some point.
Second, there are signs of a bug in the DLL code to do with normalization of the OHLC vector passed to it by the NT strategy. Proper normalization is absolutely vital to proper function and as a consequence the DLL code (NNPredictor.cs) is under the scalpel as we speak. Am using the opportunity to add support for networks with an expanded number of input OHLC vectors (up to 30 at the moment), since while the network development system (TDOHLC) can create such trained networks the DLL code is restricted to a single 4-feature input vector (topic alluded to by @sptrader). Also adding additional entry points, mostly to return system status and hence mostly to simplify debugging. Will list these explicitly when the updated DLL project is released, hopefully this week.
ETA: Should clarify that right now the NT strategy does not support strategy analysis (e.g., backtesting)--meant for forward, real time use only. However, after talking with my son (who implemented the strategy on his system and is providing invaluable feedback) backtesting suddenly makes infinite sense and have begun looking into it--hope next release will include that as well.
Last edited by bnichols; June 21st, 2012 at 08:35 AM.
The following 4 users say Thank You to bnichols for this post:
Howdy @NJAMC. I took a peek at the thread last week, about the same time I looked at the next gen of the NN code being developed here (not ready to post yet), and decided I wasn't ready to tackle either project. That will doubtless change as winter approaches and the walls start closing in.
At this stage I'm pretty much focused on improving my trading know-how, adding to the repertoire of AI techniques & optimizations not so much at the moment, first because as much as I love the AI concept so far every attempt I'm made to apply it to trading has produced the same well known result; namely, it works more or less if the problem one is trying to solve (e.g., predicting daily S&P close) provides enough significant independent variables, and not so well otherwise (performance in general less than so-so, for which problems, like a dog walking on its hind legs, if Sam Johnson were alive today might point out the surprise is not that NN does it badly, but that it does it at all). Unfortunately my current obsession (short time frame spot currency) falls somewhat into the latter category.
A second reason I'm more focused on trading than AI at the moment is that experience teaches without a profitable system & sound knowledge of what makes it profitable on the one hand, any AI technique on the other hand remains a solution looking for a problem.
I will look at the ABC technique at some point (which means code it in C# as a precursor to implementing in NT if I can't lay my hands on existing code (e.g., this code referred to by @gregid in post 11 of your thread), and run some tests) but can't say at the moment when that will be....might be sooner rather than later however, since although not afflicted by Attention Deficit Disorder as such, I'm easily distracted
In the meantime I look forward to following your thread !
PS: To come up to speed (PacRim currency market very slow) created 2 VS 2010 Ultimate C# projects to compare the console app code referred to above with your code as discussed in your thread and will take a look at the NT strategy you constructed.
Last edited by bnichols; August 28th, 2012 at 12:47 AM.
Reason: Add code reference
There are a few versions of NT trading strategies on that tread you can play with as well. Myself and @Luger have put source code out that autotrade based upon the Bee Hive output. Look forward to your insight in the future.
True, the biggest problem with the ABC system is defining the Nectar for the bees. Then what do you do with the solution once the bees find it?
The nice part is the ABC system can review large quantities of independent variables (assuming the hive can understand them) to produce a single solution and if configured correctly will adapt to the market conditions.
Howdy @NJAMC: Agree SourceForge documentation can be sparse. The "API" I use is DLLs created by a Microsoft code development product (currently Visual Studio 2010 Ultimate) and spend most of my time trying to interface the DLL with NT--my chosen battle being the NT DLL interface . Is this along the lines of what you're looking for? If so I can help with that.
The following user says Thank You to bnichols for this post: