Sentiment analysis of Twitter messages is a challenging task because they contain limited contextual information. Despite the popularity and significance of this task for financial institutions, models being used still lack high accuracy. Also, most of these models are not built specifically on stock market data. Therefore, there is still a need for a highly accurate model of sentiment classification that is specifically tuned and trained for stock market data.

Facing the lack of a publicly available Twitter dataset that is labeled with positive or negative sentiments, in this paper, we first introduce a dataset of 11,000 stock market tweets. This dataset was labeled manually using Amazon Mechanical Turk. Then, we report a thorough comparison of various neural network models against different baselines. We find that when using a balanced dataset of positive and negative tweets, and a unique pre-processing technique, a shallow CNN achieves the best error rate, while a shallow LSTM, with a higher number of cells, achieves the highest accuracy of 92.7% compared to baseline of 79.9% using SVM. Building on this substantial improvement in the sentiment analysis of stock market tweets, we expect to see a similar improvement in any research that investigates the relationship between social media and various aspects of finance, such as stock market prices, perceived trust in companies, and the assessment of brand value. The dataset and the software are publicly available. In our final analysis, we used the LSTM model to assign sentiment to three years of stock market tweets. Then, we applied Granger Causality in different intervals to sentiments and stock market returns to analyze the impact of social media on stock market and visa versa.