Can investors learn from patent documents? Evidence from textual analysis

Abstract This paper examines the role of patent texts in the stock market valuation of patents. Utilizing the large language model BERT (Bidirectional Encoder Representations from Transformers) to summarize contextual information within patent texts, I find that patent texts explain 31.5% of the variation in the stock market valuation of patents and provide large incremental explanatory power beyond other structured patent characteristics, firm characteristics, and technological trends. Additionally, patent texts significantly predict the level, volatility, and cumulation speed of future earnings, suggesting they contain genuine information about firms' performance. However, investors do not fully incorporate such information within patent texts into stock prices, as evidenced by the predictive power of patent texts for future stock returns. This underreaction is diminished after the pre‐grant publication of patent applications is mandated. My findings underscore the value of patent texts as a source of information on internally developed intangibles and have implications for academics, practitioners, and regulators.