You can use the classical UCI databases: iris, pima and so on. There are other databases for time series prediction too. But your NN can't go any further than data does and, indeed, stock prediction (time series, in general) is a difficult problem so my hunch is that other training data will not solve your problem.

But i am un aware to use these databases. Can u plz explain about these db's.??

Your data must i) be as pure as possible (if possible some very good simulation or a control sample in data) ii) not be fully used, always split them in 2 parts: one for the training and one to check the performance of the training.

One which basis , one can select training and testing data.??? I have past stock data (date, high price , low price, accuray and volume) in it. Would you guide mein now plz.

i think both training and testing data sholud be the same, well i mean if you are giving to ANN date, high price ,low price as training set inputs , also you sholud give same forecasted inputs(date,h-price,l-price) for your testing input data. Also, your target data is important for your training set. By the way what is your target data for ANN?

bests

i am going to use ANN for Stock Price Prediction. i have downloaded past stock data in form of file.csv . now the problem is that i m not getting is how will i divide this data into testing data, training data , and validating data. would u plz guide me about this division.

I have tried different neural networks but problem remains same.

Take your data and split them in 2 equivalent parts, for instance if you have time-serie data between 2000 and 2010, use 2000-2005 and 2005-2010 (but they don't need to be of the same size).

Train the NN with the 2000-2005 data, and then run your trained NN on the 2005-2010 data (testing): if your NN has been correctly trained and if your data are similar, you should get similar outputs for the NN.

for instance if you have 4 months data (=120 days), you can use 110 days for the training set and you can test or predict rest of ten days. you can use back-propagation method http://www.mathworks.com/help/nnet/ug/multilayer-networks-and-backpropagation-training.html for ANN network your train data should be long as possible as.

Split your data set as randomly as possible accommodating at least 70% of your whole data set for training and remaining 30% for testing / validation. If it is a time series problem consider at least first three fourth of your data for training as larger the training set better would be the efficiency of NNs in studying the patterns.