Research

We study the connection between Generative AI and capital structure. We use 350 million LinkedIn job postings from Revelio Labs to construct a firm-level measure of Generative AI adoption based on hiring for "integrator" roles that build or deploy large language model systems. Using a staggered difference-in-differences approach, we find that firms adopting Generative AI exhibit reduced leverage and improved debt-service capacity. Results are consistent and robust across alternative leverage measures and modern staggered-adoption estimators, and withstand identification checks including entropy balancing, stacked-cohort matching, parallel-trends tests, synthetic DiD, and two instrumental variable strategies based on the evolution of LLM capabilities and geographic proximity to AI research universities. One mechanism reveals that Generative AI adoption shifts firms toward intangible, less collateralizable capital and raises technology risk, consistent with lower debt capacity. Second, adoption increases mass layoffs, especially among firms with high ex ante generative-AI exposure, and deleveraging is stronger in low-union industries, consistent with reduced reliance on debt as a labor-discipline device.

We examine how antitrust enforcement shapes product market dynamics using U.S. government procurement as a laboratory. Leveraging large language models to systematically identify DOJ procurement cases, we document a trade-off between broader market participation and procurement performance. Non-defendant contractors, particularly small, women-, and minority-owned businesses, gain market share due to reduced entry barriers and the temporary exclusion of defendants. However, enforcement also impairs procurement performance in affected markets, with adverse effects concentrated in complex projects. Our findings provide new insights into how antitrust actions influence market structure and performance outcomes.

Do banks accumulate knowledge about corporate technology, and does it matter for their lending? To answer this question, we combine corporate innovation with syndicated loan data. We find that loans to firms sharing similar technologies with banks’ prior borrowers obtain lower loan spreads. We can rule out product market competition, the value of their technology and ability to innovate, and/or numerous other firm characteristics as alternative explanations. By estimating a structural bank-borrower matching model and exploiting the consummation of bank mergers and acquisitions, we can show that shocks to banks’ technology knowledge causally affect loan spreads.

We develop a deep learning Co-Lending Graph Neural Network to extract forward-looking information about banks’ future credit deterioration and loan loss recognition from syndicated lending networks. The model integrates bank characteristics, loan attributes, and directed co-lending links to generate a Co-Lending Network (CLN) spillover risk score for each bank-quarter. Using quasi-natural experiments around bank credit-rating downgrades and the collapse of Lehman Brothers, we show that the CLN score captures directional risk spillovers transmitted through actual co-lending relationships rather than merely shared borrower defaults. The CLN score robustly predicts future accounting outcomes up to two years ahead, including loan loss provisions and non-performing loans, and also forecasts market-based default probabilities, idiosyncratic stock return volatility, and profitability. Predictive power is concentrated in revolving credit facilities, consistent with spillovers arising from weakened monitoring by lead arrangers that transmit risk to participant banks. Predictive performance is strongest for complex, small, and private banks, where market signals are weaker and reporting opacity is greater. Our deep learning framework provides a scalable early-warning tool that complements accounting-based monitoring of interconnected banks.

Gender disparities in science shape who produces knowledge, who receives credit, and whose discoveries translate into downstream impact. We construct a global researcher-level panel spanning 1900–2024 by linking 80.4 million unique publications to 14.7 million unique researchers and connecting scientific outputs to grants and patent citations to non-patent literature. Our findings reveal that despite dramatic progress in women’s entry into science, gender gaps persist and re-emerge at multiple stages of the scientific pipeline. Near-parity among recent entrants coexists with lower female representation among active researchers, reflecting differential career survival and slow-moving cohort replacement. Large cross-field differences suggest powerful sorting across disciplines. Modest early-career productivity gaps widen after the first five years and coincide with higher female attrition and continued underrepresentation in leadership and credit-intensive roles. Women’s share of grant funding rises but remains below their share of active researchers. Cross-country evidence further suggests that convergence is more closely associated with educational gender parity and digital connectivity than with aggregate increases in government R&D spending. Women also remain underrepresented among breakthrough discoveries and are less likely to produce work cited in patents. Taken together, these findings indicate that closing gender gaps requires attention to mid-career dynamics, funding design, credit allocation, and translation pathways, not entry alone.

‹

›