Abstract:
In this century, deep learning has been applied for many research areas after accomplishing considerable improvements of prediction in solving complex problems in computer vision and computational linguistics area. However, some issues such as time series prediction are still limited to benefit from the deep networks because of its difficulty in gathering enough data for learning process. Specially after arrival of graphical processing unit deep learning get intention from both researcher and developers. Prediction on time series data has always been challenging task because habitually it is nonlinear, chaotic and non-stationary. Traditional statistical time series and machine learning models for prediction may not satisfy the much higher demand of exactness in time series prediction. Other than accuracy, another problem is stability means when missing values in temporal data. In this study modified form of deep recurrent neural network is used, Long Short-Term Memory cell are capable to learn pattern in short term and long term sequences. Also it is helpful even when some data points is missing in data. Through feature engineering famous technique feature split use in this thesis to make new features and they helpful for better learning. LSTM use to perform pattern learning since they learn a representation of their inputs at hidden layers which is eventually used for regression. Stock prices are represented as time series data and neural networks are trained to learn the patterns from trends. Experiments for this study are performed on stock market dataset using deep learning model. On this data set also apply some famous time series and machine learning techniques for comparison. Results are promising for other time series applications as well.
