Deep Learning, Textual Sentiment, and Financial Market

In this paper, we apply the BERT model, a cut-edging deep learning model, to construct a novel textual sentiment index in the Chinese stock market. By introducing the market returns as sentiment labels, our BERT model successfully extracts useful sentiment-related information contained in asset prices. We find that the BERT-based sentiment has much greater predictive power than the traditional dictionary method has both in and out of sample, and the predictability becomes both statistically and economically significant. We further explore the source of the superiority of BERT-based sentiment. We find that BERT-based sentiment is more consistent with the theory, which shows an especially strong predictive power on market returns during economic downturns and can significantly predict some macroeconomic indicators. Besides, the BERT model has broader applicability than the dictionary method. We find that the BERT-based sentiment has a better performance on contexts with different language patterns and textual features than dictionary-based sentiment. Overall, our results suggest that the BERT model offers a more powerful, reliable, and theory-consistent measure of textual sentiment, highlighting the significant value of deep learning models, AI, and FinTech in the financial market