I have decided to possibly try an insane move. And am in the process of fleshing out a Forex bot/EA. Right now I am in the process of running simulations on historical data, things are going pretty well and I think it’s time to take things to the next step. Mind you, I am more of an A.I programmer than a financial guru. MetaTrader currently seems like the best platform to write an EA on. The API seems rather open, and the code is C style and seems rather easy to code for. Before I take the dive, I just want to make sure that writing an EA using MetaEditor would be the best choice. Any other platforms that might be better, or is MetaTrader the best? As for the EA I’m working on, most of the logic/processing would happen outside of the EA, precalculated and precompiled. The EA would be just doing a few hundred calculations based on various factors.
Furthermore, what broker would you suggest for MetaTrader? This is a bit important, as it takes about an hour to do the logic processing, this is using OpenCL hardware acceleration combined with a relatively modern GPU capable of sustaining more than 1 teraflop per second. The output EA logic would be finely tuned based on the broker and various other factors. Switching brokers would require significant re-working of the EA, and would not be ideal at this stage.
Important things for an MT4 equipped broker, in order of importance:
Trustability; Doing a google search for MT4 brokers brings up all manner of horror stories.
Low minimal deposit; I plan on running multiple versions of the EA side by side with minor variations. I am willing to spend money, but considering the possibility of things going completely haywire I want to start out small.
High leverage; 100:1 or higher is ideal. Possibly a bit extreme, but might as well make things a bit interesting.
It’s not so much re-writing. I’m not re-coding anything, I just let it tune itself based on what works best. Like I said, I am more of an AI programmer. A year’s worth of Forex data at one minute intervals consists of about 300,000 lines of CSV data. Toss a bunch of data for processing, and let things go, letting it decide on what strategies work best. Different conditions, like spreads, leverage all can result in different strategies. A strategy that works for a high leverage, large spread broker may not work all that good for a lower leverage, lower spread broker.
I just found this out a few days ago but apparently IBFX which uses MT will let you open a $50 account. So, that might be a good size to test the efficiency of your EA.
On a side note, your concept- using OpenCL processing to do the grunt work is something I am also starting to work on. I’m not anywhere near the point you are at though- I am still trying to decide if I want code my processing algorithms raw or to put them in Matlab. And, apparently, there is a Matlab to C converter.
So are you a fuzzy logic or neural network AI programmer?
IBFX seems pretty good, I’ll have to take a look at it.
As for Matlab, I’ve never tried it. But OpenCL is a pain to debug. However if you need a serious amount of number crunching (IE, to run a few million rows of forex data at a time), the time spent coding for OpenCL will pay for itself pretty quickly. But if work you’re doing isn’t that logic intensive then OpenCL might be a bit too much overkill, and going for something easier to code might be a better idea.
I had originally started with open source NN based code. First attempt was to have the NN try and predict the future trading cost for stocks. The best I reached was 99.5% accurate for the next day’s closing value. Which sounds good. But the .5% was where all the profit went. I then tried to have it decide when to buy, feeding it previous history and trying to train it to buy when the value was expected to go above a certain figure. That was a bit more accurate, but the problem is that most NN software packages are focused more on average accuracy. Ergo, be as accurate as possible as much as possible, even if there are many false positives. Whereas for stock/currency trading, it is better to skip trades to avoid false positives. So I gave up on opensourced packages, coded stuff from scratch. What I have now is loosely based on neural network concepts. It would basically be described as a modified NN, which it’s training is defined by another NN, which itself is determined by a genetic algorithm. While I wrote the base code, the program itself decides the best way to do things.
glad somebody else is trying FuzzyLogic, respective NN.
I did that about 5 years ago since I could lay my hands on "testing with NN/FuzzyLogic, stocks at the time) and after 2 month not getting anywhere (except losses) I realized NN/Fuzzy is “FUZZY” and gave up on it.
I wish you good luck.
NN are interesting to be but I’ve heard Fuzzy Logic is a better way to go. I can’t speak for the success of either because I am still in the research phase.
I just graduated in Physics and Math this past Spring, so I am just beginning to read on some of the strategies using AI. It seems that most people who speak publicly fail at being successful with AI Programming and Financial forecasting. I personally think it’s because the successful people aren’t talking about their success.
As far as Matlab goes, I am considering using it because they provide two very useful frameworks- the NN framework and Financial Forecasting framework. Each of these frameworks lets you use optimized algorithms to analyze the financial markets using simplified NN’s. And, because my wife is still a student- I can purchase Matlab 10 and the the frameworks for a little under $200.
I’m glad there are other people pursuing the same goals as me, a little camaraderie always helps the process.
Yeah, I’m currently using GAs to backtest blends of systems, and have started toying with the Neuroph NN framework (one of my friends is a contributor to the project, so he’s a great resource for me). Would love to share more info with you guys as I go through this process.
Sorry for the late response. Honeymoon took precedent.
Neuroph is a great little collaboration, not to mention having the resource of a contributor.
I think we should start a new thread where we try to blend some of these ideas into a cohesive, free EA.
One question I am still seeking to answer as I look for a CUDA card: Do I need double precision processing power or can I get away with single precision? The 285GTX has almost a teraflop of single precision goodness…
