How Can Innovation Screening Be Improved? A Machine Learning Analysis With Economic Consequences for Firm Performance

Using USPTO patent application data, I apply a machine-learning algorithm to analyze how the current patent examination process in the U.S. can be improved in terms of granting higher quality patents. I make use of the quasi-random assignment of patent applications to examiners to show that screening decisions aided by a machine learning algorithm lead to a 15.5% gain in patent generality and a 35.6% gain in patent citations. To analyze the economic consequences of current patent screening on both public and private firms, I construct an ex-ante measure of past false acceptance rate for each examiner by exploiting the disagreement in patent screening decisions between the algorithm and current patent examiner. I first show that patents granted by examiners with higher false acceptance rates have lower announcement returns around patent grant news. Moreover, these patents are more likely to expire early. Next, I find that public firms whose patents are granted by such examiners are more likely to get sued in patent litigation cases. Consequently, these firms cut R&D investments and have worse operating performance. Lastly, I find that private firms whose patents are granted by such examiners are less likely to exit successfully by an IPO or an M&A. Overall, this study suggests that the social and economic cost of an inefficient patent screening system is large and can be mitigated with the help of a machine learning algorithm