A recent study has shown the power of the neural network in many research areas. It has been proven that neural network can generalized and learn better from raw datasets. In recent years, RNNs is captivating as it can take arbitrary input opposed to static input datasets by standard feed-forward networks. Unlike standard feed-forward networks, the recurrent neural network has the ability to store relevant information over time, which makes RNN more powerful than other neural networks. The ability to store relevant information over time makes RNNs very useful in many applications like prediction, music, and speech. However, RNNs have a limitation in their memory and it cannot handle the long time lags between the past and present events. The long time lags problem of RNNs is overcomes by LSTM (“Long Short Term Memory”) networks. The LSTM can handle more than 1000 steps of time lags and have the ability to learn when to store and when to access the information.　　In this thesis, an LSTM networks is used as forecasting method, and method is developed integrating Independent Component Analysis to elevate the forecasting accuracy for time series. LSTM have been successfully used in many applications such as grammar prediction, speech recognition, and handwriting recognition. However, it has limited work in the field of time series prediction. The neural network has overcome the limitation of traditional methods and can handle missing or corrupt data, temporal dependence, and noisy data. Here we use the LSTM network with independent component analysis (ICA) for the prediction of stock prices. Independent component analysis in our work is used for the preprocessing of raw datasets before it is fit to the LSTM networks for forecasting. The independent component analysis is used to estimate hidden independent components of the prediction variables. After that independent component are used as an input variable to LSTM networks for the prediction. For the experiment and performance evaluation here we used 7 years of historical data of a daily transaction from DOW30 stock index. We used historical daily data of four company namely Cisco, Microsoft, IBM and Apple of dow30 stock index. In our work, we used 85%of data samples for training LSTM networks and remaining data samples are used for the testing of the model. Here we predict the opening prices of the target stock exchange data in one day advance and the performance of the LSTM-ICA based method is compared with single LSTM method, and sparse method. The data used in single LSTM model and sparse model are raw input variables. The quantitative metrics for the performance evaluation of model used is the root mean square error values.　　Our experimental results show that LSTM with ICA data pre-processing generates the lowest root mean square error values compared to single LSTM and sparse methods. The integration of ICA improves the performance of LSTM model compared to single LSTM and sparse models. Also from our results, it can be noted that compared to sparse model the rmse error produced by single LSTM is less. We can say that even from raw datasets LSTM can make a more accurate prediction then sparse model.