Research

Working papers:

Sources of Technological Novelty and Organization Types: A simple Theory

Co-authored with Ashish Arora (Duke) and Reinhilde Veugelers (KU Leuven)

Abstract: This paper investigates the incentives and capabilities for producing technological novelty in bio-pharmaceuticals. Organizations differ in their efficiency in searching for novelty and in the benefits they derive from novelty. We combine this with the perspective that experience with technologies has different effects on the ability to produce novel and incremental inventions. We develop a simple model to guide hypothesis development and test its predictions in a sample of drug-related biotechnology patents filed between 1985 and 2000. In an important empirical advance over the literature, we study both the generation of novelty as well as the economic value of the novel inventions. This enables use to distinguish between factors that differentiate organizations in the benefits they derive from novelty from factors that differentiate organizations in the efficiency with which they can generate technological novelty.

Enchanted Wanderer or Stone Guest? On Field-original Knowledge and the Novelty and the Creativity of Inventors.

Abstract: Do inventors with knowledge that is uncommon in their field of work create more novel and highly valuable inventions? The answer is 'yes', with reserves. Using data on the careers of 100,000 inventors over 15 years of corporate U.S. patents and an inventor-firm fixed effects panel, I document that inventors with field-original knowledge (i.e., uncommon among others who work in the same field) produce inventions which are on average more novel and more likely to be exceptionally valuable, but more likely to be failures as well. Additionally, estimates from structural equation models illustrate that novelty is at the same time an outcome of the search process and a contributor to the creation of economic breakthroughs. Caution is due because the rate of failure increases as well, suggesting that greater uncertainty, rather than recombinant fertility or displaced expertise, may be a driving mechanism.

Online supplements:

Fig: The relationship between field-original knowledge ('distance'), novelty ('new combinations') and the breakthrough rate.

Exploring of the Commercial Value of Technological Trajectories: The Roles of Scientific Knowledge and Technological Complexity

Abstract: Are science-based novel technologies more or less likely to spawn innovations with commercial value? Using a novel dataset matching the trajectories of 1,369 drug-related corporate inventions in biotechnology to their approval for market launch, I document that when the novel recombinant technology is built with science-based components, its trajectory---the set of inventions that incrementally build on the novel technology---has a larger number of commercially successful followers on average (proxied by patents of FDA-approved medicinal drug). The relationship intensifies as complexity increases (when recombination is more difficult, measured by greater interdependence of the components). In the context of technological innovation, a larger science base may imply that more knowledge is available about the components, in addition to it being formalized, codified and explicit. Results suggest that greater scientific understanding of newly combined components helps overcome the hurdle of complexity in the search for novel commercially valuable innovations.

Some Inventions are More Equal than Others: Assessing and Developing Patent Indicators that Signal Radical Inventions

Co-authored with Paul-Emmanuel Anckaert (SKEMA), Francesco Appio (Pôle Universitaire Léonard de Vinci), Cindy Lopes-Bento (Maastricht University), Bart Van Looy (KU Leuven) and Dennis Verhoeven (KU Leuven)

Abstract: Early and accurate prediction of radical inventions has considerable strategic and financial importance for firms. In this work we examine a set of novel ex-ante indicators built from patent data and assess their ability to predict patents linked to such radical inventions. Ex-ante indicators rely on information available up to the time of the patent grant only, and thus hold great predictive potential. We combine these indicators with insights from patent analytics, and develop predictive machine learning models for the most important inventions from the early phase of the biotechnology industry (1985-2001). To evaluate the contribution of our approach, we benchmark our models against the 'state of the art': logistic classifiers with the standard indicators in the patent literature. Our best model, a random forests classifier, delivers 69% recall and 100% precision (AUC=0.84), improving recall and precision by 15% and 405%, respectively, against the baseline models. We also analyzed our models' false positives and negatives to derive feasible suggestions on how to improve ex-ante patent-based indicators.

Online supplements:

  • Jupyter notebook: explaining the machine learning classification/prediction tasks and their performance (my main contribution to this joint work).

Fig: 2-class precission-recall curves from the baseline model (blue, only existing indicators) vs our improved predictive models (orange & green)

Miscelaneous projects:

Recondita armonia: Using machine learning to measure the distance between composers' music based on harmonic novelty

I am a huge 'classical' music fan. As a hobby, I combine machine learning and approaches common in patent analytics and bibliometrics to explore large datasets of symbolic music and composer metadata. My goal is to create new algorithms to suggest classical music to hardcore fans like me. You can read about it here.