Multi-day Window for Stock Movement Prediction and Financial News Classification for Predicting Market Sentiments

Abstract

There are many factors that influence the stock market prices and the general sentiment of the market. It has been difficult for a very long time to predict the stock prices in an accurate way using an automated system because stock prices are extremely volatile in nature as they depend on a variety of factors like financial/political news, historical prices, and market sentiments. In this paper, the proposed individual models take into consideration the above-mentioned factors to predict the future stock price movement and market sentiment. This is done by gathering all credible financial news data and historical stock price data. The proposed News Sentiment model uses deep learning to predict the next day's market sentiment i.e. bullish/bearish. This four-class news sentiment detection Four Class - News Sentiment Detection (FC-NSD) model has a testing accuracy of ~43\% and a R2 score of ~0.25 which is higher than previously published models. The proposed classification model (CDWM - Convolutional Dual Window Model) leverages one-dimensional convolutional layers to remove noise and learn patterns. This coupled with a unique pipeline that uses multi-day window prices for training is implemented to predict the future stock movement with higher accuracy. The proposed LSTM classification model achieves a state-of-the-art classification accuracy of 68.75\% in the binary buy-sell classification problem which is ~8\% more than the existing state of the art models.