Analyzing Varied Approaches for Forecast of Stock Prices by Combining News Mining and Time Series Analysis

Forecasting of Stock Prices has become a trending topic of research relating to varied fields such as machine learning, economy and other fields. Most of the approaches to perform such a task make use of the time series data or news mining. Albeit several approaches have been proposed to combine the two but the choice among those remains a problem. Data mined from the News reports is analyzed with text mining techniques to obtain the mining results which are used for improving the accuracy of analysis algorithms. To perform the aforementioned task, there exist several approaches which may be linear, supervised or even unsupervised. This paper makes use of the financial data to predict the movement of the Dow Jones Industrial Average (DJIA) using varied methods and further delineates the comparison amongst the same. The text from the newspapers is represented using state-of-the-art Word Embedding to help resolve the information loss due to loss of context. Using these embedding, a neural network is trained to forecast markets. The predictions obtained are then evaluated and compared to standard Long Short Term Memory Networks (LSTM) based approach in tandem with popular Convolutional models and other neural network architectures. The results indicate that the neural network based approach out-performs the traditional methods and holds vast potential for further improvements.