Rajib Oraon

How do incumbent workers adjust their human capital when a new technology changes the value of existing skills? I study this question using quarterly résumé histories for 1.5 million U.S. workers from 2018 through 2025 and the public release of ChatGPT as a common shock. Treatment intensity is the pre-release AI exposure of a worker's occupation. Workers in more exposed occupations become substantially more likely to earn AI-related credentials after ChatGPT. This increase does not represent a broad expansion of training: the same workers reduce certification in adjacent technical fields, including data science and cloud/IT, and in other discretionary subjects such as marketing and finance, while safety and compliance credentials do not move and total certification does not increase. The response is therefore primarily a reallocation of human-capital investment across skills. Adjustment is stronger among postgraduates, workers with technical backgrounds, employees at large firms, workers with shorter tenure, and workers already employed in technology industries. Employer demand for AI skills subsequently rises more in exposed occupations, but worker certification begins to increase before this rise in vacancy demand becomes clearly detectable. The findings are consistent with workers reallocating skill investment in anticipation of technological change before changes in posted labor demand are fully visible.

Educating children in developing countries continues to pose a challenge, and technology might be a potential solution. This paper explores the most basic technology: the availability of affordable, fast internet access and its impact on human capital outcomes. I exploit the phased rollout of BharatNet, India's national rural broadband programme, estimating 2SLS specifications with district and time fixed effects and instrumenting phased treatment with a district's terrain ruggedness index, since Phase 1 prioritized blocks requiring less optical fibre to connect and cable requirements are themselves a function of terrain. The sample combines three rounds of India's National Sample Survey on Household Social Consumption: Education, covering roughly 164,000 rural households across up to 651 districts, with district-level rollout data. I document a positive and significant impact of fast internet access on educational enrollment at the middle and secondary levels. There is a noticeable shift in major choice from humanities to science at the secondary level, along with an increase in preference for medicine courses and a decrease in IT and computer courses at the tertiary level. Lastly, I document an increased preference for private institutions.

with Kunal Garg

We compare a dynamic logit benchmark with machine learning for corporate default prediction on a point-in-time panel of 157,556 firm-year observations on U.S. non-financial, non-utility public firms (Compustat and CRSP, 1980–2022). Both models are trained on pre-2006 data and evaluated out of time across three non-overlapping forward windows: a discrete-time hazard model with the predictor set of Campbell, Hilscher, and Szilagyi (2008), against XGBoost tuned by time-series cross-validation. Both models discriminate well, with AUC-ROC between 0.90 and 0.93, but strong discrimination does not ensure reliable calibration. Both track realized default rates closely in ordinary years, yet underpredict defaults by roughly half in 2007–2008 and overpredict them by an order of magnitude in 2020–2022. Expected-loss implications follow through a Basel-style PD × EAD × LGD bridge, relevant to CCAR stress testing and IFRS 9 provisioning.