Hello,Does anybody here use Neural networks in predicting the S&P 500 returns. I am trying to use the Matlab's neural network toolbox but I am having a hard time in making a good fit for the inputrs and targets. I took the daily returns of S&P 500 from 1969 to 1999 and split it into training and testing data. the training data consists of about 4000 points and the rest is for testing. I have lagged (8 lags) the returns and taking these lagged returns, I am trying to predict the original return. After reading a lot of papers on this topic, I came to conclusion that feed forward backpropogation networks are good for this purpose. I used MATLAB's NN toolbox and using their commands I gave the lagged returns as inputs and original returns as the target in the training set. After training the data, I generated the output giving the same inputs expecting to see the targets again. Unfortunately, It did not fit correctly. I was more interested in sign changes. So I counted them. It only fit 60% correctly even after I gave the same inputs that I used for training. Is this a fault with the NN toolbox of matlab or is it just that the relationship is so complex that it did not understand properly and may be a different network would work properly?? Please can somebody point towards a right direction??Is there any other freely available, simple to use NN software available??regards,monarc

Hi Monarc,I use Neural Network module from Data Miner - software by StatSoft.I think it's good for such problems.I use NN to speculate on EUR/USD rate, and I think that mulilayer feed forward backpropagations NN are really good.Best regards,Piotr

Monarc,how good is your model?what are your rates?

I don't know is it good from 1st January 2004 my NN earn 4131 pips.So It was 96 deals and 80% was positive.I take 2000$, and buy 1 NDF on EUR/USD, it was on 50.000$ and deposit was 500$, so 1500$ is free margin. After 1.5 year you have this 2000$ + around 20655$.I think that on S&P - it's very liquid, so I think you can find good NN and good input variables. Be patient and long hard, sometimes very boring work. Others details are in secret ;-)Best regards,Piotr

ups... it was question to Monarc, not for me... SORRY !!!

pnowy,Just to start with, if I want to take the 10 lagged returns of the S&P500 as inputs and tryu to predict the return that was lagged, a 10 input, 6-10 neuron hidden layer and 1 output neuron be good enough??? I am actually not getting more than 60% correct sigbn prediction even if I give the same training inputs to the network which is strange.I am thinking of switching to different software. probably S-Plus.

Hi Monarc,It's hard to say what architecture is good for S&P500 and what variable take to learn and predict.I think that predict (point prediction) the value of S&P500 it's very hard as any other index etc... If you want to use the NN to trade, I think you should focus on direction of change S&P500 - it's easier, so the costs are lower and profits can be higher ;-)I think that you should use lot's of variables e.g. : price of oil, interest rate, FX, DJIA and many others... You can as traders who trade on S&P500 what they use making decisions or try to trade on demo, to understand this market. After that it will be much more easier to choose input's variable.It's good to testing as many netork's architectures as much you can.Try not only with one hidden layer, but try with two hidden layers.Don't focus only on one netowrk's architecture and one combination of variables. Market is running, and after 3month it could be need to change network and combination of variables again... etc....Best regards,Piotr