Deep Learning, Text, and Patent Valuation

This paper uses deep learning and natural language processing (NLP) methods on the US patent corpus to evaluate their predictive power in estimating two measures of patent value: (i) investor reaction to patent announcements as measured in Kogan et al., 2017 and (ii) forward citations. While forward citations have traditionally been used as measures of economic value in the literature, their utility is mainly retrospective. Contemporaneous predictions of patent value, as embodied in investor reactions to patent grants, can be important for managers and policy-makers for prospective decision making. We compare the prediction performance of models using the structured features of the patent (number of citations, technology class, etc.) to deep learning and NLP methods. Relative to linear regression models using the same features, deep learning models reduce mean absolute error (MAE) by approximately 32%. Incorporating patent text further lowers the MAE by 13%.