if your serious enough and have the skill to program, I can help with conceptualization and direction.
That is an interesting way to use nn. But I guess I will have tplay with nn designs to see how it fits into the equation. As far as amazon aws is considered, I'm even thing ou cluster. I think nn can benefit significantly n works. It would be very exciting to see it running on gpu I will keep you guys updated on my progress if you are interested.
That would be great. Fortunately, I have just enough experience to make something like this I've already started developing a prototype. Really just a proof of concept at this point. I only have about 2 hours a day to work on it for the next few days because of Sandy but I will keep you posted. The concept of nn and genetic algo is pretty straightforward to me but I need an advise on how they fit into the equation of market prediction model.
Evironment ( Price Feeds ... 1-50) Lifeforms (1-n) the lifeforms would need inputs from different price feeds. lifeform = capital allotted + Market(1)[(# contracts or #shares x (Price Entry - Current Price)] ........ Market(n) If capital <= 0 , lifeform(dead) If capital >= 0, lifeform(alive) If alive then buy/short (#shares/#contracts/Market(1)) #shares = capital/share price (no margin) #contracts = capital /(margin requirements/Market(1)) lifeform designation (multimarket) (market) If capital >= (2 x initial capital) then designate multimarket If capital <= (initial capital) then designate market If capital <= (initial capital) and <= (margin requirements) switch to market(no margin requirement) ie single stock If lifeform designate multimarket enter only noncorrellated markets Test for correllation (price feeds market 1, price feeds market 2)
lifeform components -capital (cash + profit or loss current holdings) -market or multimarket -price feed(s) (visual component) (number series with time) -tests (correllation)(test length n (random number generator 15-240 minutes) of two markets Actions: buy if linear regression slope is positive if LR slope timeframe(1)(n minutes) = LR slope timeframe(2)(n minutes) short if linear regression slope is negative if LR slope timeframe(1)(n minutes) = LR slope timeframe(2)(n minutes)
if profit after (n minutes) (random number generator) then close out position if loss after (n minutes) (random number generator) then close out position if multimarket duplicate after n days, duplicate with registry of: Correlation test length (n) LR timeframe(1) length and LR timeframe(2) length N minutes profit/loss check
Which market (1-N) ie. 1-5000? (all available price feeds with each listed stock or futures contract or currency pairs) rank each market based on price variance after n minutes, the lifeforms would only trade the market with price variance at the tails of a bell curve. thus the environment would be (tails of bell curve price variance), top 50 bottom 50.
i put together a simplified web version of what i'm working on. http://54.243.46.180 bear in mind that I made this primarily for myself and it has no domain or pretty ui. also, I do not guarantee any accuracy and use it at your own risk. don't click on it if you are going to troll or assert that ann cannot predict stock prices. the only reason why I decided to work on this is because there are research papers that state that there is a statistical significance that ann outperforms traditional trading strategies. this model copies the simplest and most common ann for stock prediction (except that this uses more advanced training method called iprop rather than backprop). i'm not sure if I'm going to implement the genetic algo for this specific web version (although I'm doing it for a server-side version), but I'm planning to add a feature that allows you to change the network design for fine-tuning. it predicts accurately sometimes, and sometimes it doesn't. but at this point, my goal is to have a model that can find a correlation with statistical significance (a proof that this thing somewhat works).
i think it's better to divide into a prediction module and a trading module, because otherwise overfitting would still be too high. that means the GA is independent of the trading module. or in other words, the fitness equation does not include a solution's profitability, but instead how well it predicts a price difference occuring between now and a certain time in the future. i've been following this idea in C/C++ for over 2 years