Twitter sentiment classification using machine learning techniques for stock markets

Sentiment classification of Twitter data has been successfully applied in finding predictions in a variety of domains. However, using sentiment classification to predict stock market variables is still challenging and ongoing research. The main objective of this study is to compare the overall accuracy of two machine learning techniques (logistic regression and neural network) with respect to providing a positive, negative and neutral sentiment for stock-related tweets. Both classifiers are compared using Bigram term frequency (TF) and Unigram term frequency - inverse document term frequency (TF-IDF) weighting schemes. Classifiers are trained using a dataset that contains 42,000 automatically annotated tweets. The training dataset forms positive, negative and neutral tweets covering four technology-related stocks (Twitter, Google, Facebook, and Tesla) collected using Twitter Search API. Classifiers give the same results in terms of overall accuracy (58%). However, empirical experiments show that using Unigram TF-IDF outperforms TF.