Research

A substantial share of public projects worldwide experience cost and schedule overruns, and how contract design shapes these outcomes remains a core challenge for project governance. We develop a data-driven machine learning framework to predict, interpret, and infer project performance using U.S. federal contract data, combining high-performing predictive models with interpretable machine learning and causal-inference-oriented methods. Models tailored to the structure of procurement data significantly improve prediction accuracy for both cost and schedule overruns. A novel aggregation metric identifies which contract attributes drive adverse outcomes and their severity. Estimating causal effects, we find performance incentives are largely ineffective under fixed-price contracts and significantly increase the likelihood of both cost and schedule overrun, while under cost-type contracts they significantly mitigate cost overrun with no robust effect on schedule. Results are robust to double machine learning estimators and high-dimensional propensity weighting.

Widely adopted metrics of scientific success such as the Journal Impact Factor capture scholarly influence within academic communities but are less informative about how science contributes to commercialization. We analyze a citation network of more than 272 million scholarly works, within which more than 190,000 works underlie patents subsequently absorbed into U.S. industry, assigned to corporate firms and to public and non-profit institutions. By measuring each paper's influence toward these industry-absorbed works, we introduce the Commercialization Impact Factor (CIF), a network-weighted measure of a publication venue's contribution to commercialization, capturing both direct industry-facing impact and diffusion impact as upstream science. Through an analysis of deviance holding venue profiles controlled, conventional prestige metrics account for only a small fraction of the variation in venues' CIF. The metric offers a complementary framework for research policy, promotion, tenure, and funding decisions aimed at real-world impact.

• Develops a game-theoretic model of low-end disruption in which the disruptor enters with an asymmetric accessibility advantage and can close the quality gap once the incumbent is displaced from the low end.
• Studies the incumbent's incremental innovation and product strategies as defenses against disruption, jointly with its anticipation of the disruptor's post-displacement quality catch-up.
• Identifies conditions under which an incumbent must be both informed and aggressive in going low to achieve both later low-end displacement and higher profit together; and quantifies the cost of misjudging the disruptor: for an uninformed incumbent no product strategy is both profit-maximizing and displacement-deterring, and it is misled toward conventional responses and a sub-optimal product strategy.

Securing intellectual property protection is vital to ventures large and small but is complex and expensive, and many ideas go unpatented because of it. We study whether algorithmic approaches can aid the patent application preparation process by predicting the odds of approval of individual patent claims, saving applicants time and legal resources. We find that even the best deep learning algorithms are seriously limited in predicting patent approval, given the contextual challenges of patent assessment. Combining human domain expertise with deep learning significantly improves the claim approval prediction rate; specifically, we present a graph-based approach offering a substantial improvement.