Predicting the obvious : a machine learning approach to superstar inventions

Abstract

While patent citations are a common way to measure innovative output, their use as a measure of invention quality involves a paradox. How can we use an ex-post measure of impact (the number of received citations) to identify the ex-ante quality of a given innovation? This paper proposes a novel method of measuring patent quality using patent text and sections of the patent citation distribution with the highest signal to noise ratio. We provide empirical evidence that the bias from using citations to measure quality varies by location in the patent distribution and, contrary to what one might expect, superstar patents are the most predictable while patents in the middle of the distribution are most contaminated with noise. We show predictability of patents increases monotonically over the patent distribution - with the most valuable being the most predictable - and removing the middle of the distribution has little impact on accuracy. We also provide suggestive evidence on the importance of patent text in measuring quality and conclude with suggestive geometric evidence we are capturing differences in underlying patent characteristics. As our model demonstrates, our empirical results generalize to other situations involving highly skewed processes observed with noise. This paper also has implications for empirical work using citation weighted metrics.