Working Papers
The Realism Gap: Why Liability Frictions and Integration Costs Stalled the AI Boom [Download PDF] Despite the transformative potential of Generative AI, aggregate US automation rates have stagnated since 2022. I define this divergence as the "Realism Gap" and argue that Algorithmic Trust Frictions—specifically liability risks—act as a binding constraint on adoption. Using a novel LLM-verified index of algorithmic risk constructed from corporate 10-K filings , I document that simple OLS estimates are biased by reverse causality. I construct a shift-share (Bartik) instrument interacting historical liability sensitivity with aggregate risk shocks to identify the causal impact of friction. I find that exogenous trust shocks significantly retard automation (βIV=−0.22), imposing a "Trust Tax" that renders a meaningful share of technical capability economically unviable. Counterfactual simulations show that resolving liability uncertainty would recover productivity but unleash a targeted displacement effect in high-liability sectors.
Firm Risk Networks [Download PDF] This paper estimates a factor pricing model from firm disclosures. Using the latest techniques from natural language processing and machine learning, I isolate risk factors of firms from their disclosures. I then decompose these risk factors into those that are shared with other firms and those that are unshared risk factors. This sharing of risk naturally leads to the formation of networks of firms connected by the factors affecting them. From these networks, I build measures of risk at the firm level. I find that these text-based measures of shared risk achieve smaller pricing errors than the Fama-French three and five factor models. I also find that the text-based measure of unshared risk is not priced in the US stock market. I have built a website to help validate the clustering of topics. [Click here to interact with the data.]
Perceptions Matter: Evidence from the Global Media Sentiment Indices of the Chinese Economy [Download PDF] Drawing from state-of-the-art techniques from Natural Language Processing, we construct two new media-based Chinese economic sentiment indices using a large corpus of English and Chinese news articles and demonstrate that differences in perception matter for economic outcomes. Our sentiment classification models improve the accuracy of lexicon approaches by a factor of two. Consistent with the agenda setting theory in communication research, we find that information flows from the English to the Chinese media—with the latter tending to be more positive—and not the other way around, providing new evidence on international news diffusion. Moreover, the difference in sentiment between Chinese and English news outlets—which we term the “perception gap”—has widened in recent years. Evidence from a structural VAR suggests that positive sentiment shocks foreshadow increases in China’s policy rates and asset returns, as well as global commodity prices.
Work In Progress
Aggregate Effects of Acquisitions Mergers and acquisitions (M&A) have been a part of economic study through the lens of industrial organization where the focus is on understanding changes in market power and subsequent anti-trust policy measures. However, there is reason to believe in the importance of M&A from a macroeconomic perspective—particularly in the context of firm investment. In this paper, I demonstrate that acquisitions have the following characteristics: (a) have a strong positive relationship with firm sales; (b) are procyclical at the macro level; (c) are lumpy. Building upon a standard general equilibrium model of firm investment, I interpret acquisitions as a form of intangible investment built through organizational capital. I then estimate the model and develop economic insights for the role of acquisitions in the macroeconomy.