Targeted Evaluation of Context-Sensitive Sentiment Analysis Algorithms for Prediction of Stock Trends

Abstract

Stock sentiment plays a very important role for analysis and prediction of stock trends. Data regarding stock sentiment is available via micro-blogging platforms like Twitter, Google reviews, news articles, etc. Sentiment analysis of this data can give an approximate idea about the underlying stock’s price movement. In order to perform this task, a wide variety of sentiment analysis models have been proposed by researchers. These models include dictionary-based models, aspect-based models, deep-learning models, etc. Each of these models has their own nuances, advantages and limitations. For instance, deep learning models like Word2Vec and GloVe are highly accurate and can incorporate synonym matching, but require large delays for sentiment computations. Simpler models like text-blob and dictionary-based matching have good performance for application specific datasets, but cannot be applied for General purpose text. Thus, it becomes ambiguous to select best suited model for the given stock type, which increases testing and evaluation delay while building stock-based sentiment analysis systems. In order to reduce this ambiguity, the underlying text evaluates performance of some of the most efficient sentiment analysis models in terms of delay and accuracy of sentiment evaluation for Apple, Reliance, Tata Motors, ONGC stock. The models evaluated include Rule based models and Word embedding models. It is observed that a trained Word2Vec model and Wordlist-model outperforms other models and can be used for high accuracy stock-based sentiment analysis.