Beyond Patent Ownership: Learning About Technological Usefulness
Abstract: Technology is central to economics, but current datasets lack the scale, scope, span, and specificity economists need. We apply natural language processing and positive-unlabeled machine learning to descriptions of patents and U.S. public firms over nearly three decades to create a firm-level technology dataset that is unmatched in its combination of scale, scope, span, and specificity. For the first time, we reveal the core technologies of non-patenting firms. We estimate the value these firms derive from technological innovation, and argue that stock market participants inefficiently process their technological information, enabling profitable trading strategies.
Figure: Positive-Unlabeled Learning Classification Results
Notes. The scatter plot shows a class-balanced random sample of 250 positive and 250 unlabeled firm-patent pairs, highlighting the separation between these groups. Firm-patent pairs where the firm owns the patent and our model predicts usefulness are marked with transparent black circles with solid black centers, while those predicted useless are marked with transparent black circles only. Pairs where the firm does not own the patent and our model predicts usefulness are marked with transparent yellow circles with solid yellow centers, while those predicted useless are marked with transparent yellow circles only (on hover). The horizontal axis represents SBERT cosine similarity scores, and the vertical axis represents TF-IDF cosine similarity scores. Kernel density plots above and to the right show the class-conditional distributions of each feature for the random sample. While the scatter plot demonstrates the predictive results of the simple two-feature model to illustrate class separation in two dimensions, our preferred model incorporates additional features and interactions, achieving significantly better performance.
BibTeX Citation
@article{Gorrin2025BeyondPatent,
title={Beyond Patent Ownership: Learning About Technological Usefulness},
author={Jesus Gorrin and Rory Mullen},
journal={Working Paper},
year={2025}
}