Quantifying The Effect Of The Rumor On The Stock Market

Access of daily data is available for the users of the Web media which scholars and companies always endeavor to contemplate with deep insight.But most of part of this data is unsynchronized and despite of this,the data contains significant information for conducting research.In the competitive area of Finance and Economics where transactions are automatically managed by machines,information asymmetry can easily either exacerbate the losses or augment the profits.Thus,ally financial rumor stemming from the web media and the stock prices seems actually crucial over using either source separately.Contemplating the impact of the rumor on the stock market is a tough problem because of the nature of the natural languages,and also because of the noisy and hardly predictable nature of the stock dataset which may reflect factors beyond the scope of the Finance and Economics.Research in this topic basically rely on several skills including the feature extraction,the Natural Language Processing and the predictive models.Actually the most spread feature extraction techniques include the word embedding,whereas the Recurrent Neural Networks based models are ideal for the tasks including the time series or any datasets,which samples are believed to maintain a certain degree of time dependencies.However,enormous literature has already proven that very long time series are problematic for the standard RNNs.The LSTM model is powerful enough without being provided with the time dependencies among the observations in the dataset.It needs to learn independently.Furthermore,many successful machine learning models don't even make the assumption of the time dependency among time series samples.The memory provided models always assume that the time gap between consecutive samples remains constant.Meanwhile,the rumors are free of any control,populate the web media nowadays and sometimes carry useful signals.This research explicitly contributes via an enhanced version of the LSTM has been formulated via new branching,some deep states and the time lag as an additional input.The LSTM has been modified such that the time lag computed is applied to both inputs and hidden state at each time step.The deep states make it possible for the predictions to no longer depend upon a single state but,on several states which can separately learn distinct patterns.The deep states act in a fashion similar to the way the feature maps of the Convolutional neural networks capture different facets of signals which is contained into the input dataset,the proposed model also implements a new timing function which scales and shifts the time lag from its original range to a [0,0.75] range concluding to be more convenient for the machine learning models.The research has unveiled that this model is best suited for solving the classification and regression problems.The proposed model performs splendidly when provided longer time series.This work made several contributions comprising the use of function composition rather than simple mathematical formula or Machine Learning activation function.Furthermore,a new branching which has less gates and input nodes was proposed to enhance the former LSTM cell.The model was totally built in TensorFlow which provided more control and opportunity to customize the programs.The proposed model has attained the accuracy and F score of 99.93% on the Classification test,proving that the prediction of the time series can further be improved to reach state-of-art performance.Finally,this work has developed an applicable model which can be executed to solve real life problems including prediction of the financial time series and it also shed lights on the importance of scaling the dataset by assessing two data scaling techniques.