Efficient Market Hypothesis Test with Stock Tweets and Natural Language Processing Models

Abstract

The efficient market hypothesis (EMH) plays a fundamental role in modern financial theory. Previous empirical studies have tested the weak and semi-strong forms of EMH with typical financial data, such as historical stock prices and annual earnings. However, few tests have been extended to include alternative data such as tweets. In this study, we use 1) two stock tweet datasets that have different features and 2) nine natural language processing (NLP)-based deep learning models to test the semi-strong form EMH in the United States stock market. None of our experimental results show that stock tweets with NLP-based models can prominently improve the daily stock price prediction accuracy compared with random guesses. Our experiment provides evidence that the semi-strong form of EMH holds in the United States stock market on a daily basis when considering stock tweet information with the NLP-based models.