Jasper Pan

Working Papers 

Equity Prices the Opportunity, Debt Prices the Risk: Generative AI and Corporate Credit Spreads (with Priyank Gandhi, Juntai Lu, Alberto Plazzi and Jia Wei)   

How do corporate bond markets price workforce exposure to generative AI? We find that firms with greater workforce exposure to generative AI experience credit spreads nearly 8 basis points wider per standard deviation of exposure following the release of ChatGPT, well exceeding the 4 – 6 basis point effects for policy uncertainty shocks documented in the literature. This finding appears paradoxical: high-exposure firms simultaneously experience rising equity valuations and wider credit spreads. We argue that ChatGPT’s release created a real option to restructure labor whose fundamental parameters — adoption costs, success probabilities, and competitive consequences — were unknown at the outset. This parameter uncertainty, typically absent from prior models of labor-technology substitution, simultaneously raises expected asset value and generates relatively wider credit spreads for high-exposure firms. A model with parameter uncertainty rationalizes this pattern, and cross-sectional tests confirm its predictions: the effect is larger for firms with greater credit fragility, labor intensity, and weaker governance.

Conference Presentations: Wharton-Chicago-Harvard Insolvency and Restructuring Conference (2026)*, Ohio State Fisher AI in Business Conference (2026)*, Indian School of Business Summer Research Conference (2026)*, Concordia University AI in Finance Conference (2026)*, Financial Management Association Emerging Scholar Initiative (2026), Southern Finance Association (2026), C.r.e.d.i.t. Conference (2026)*, Eastern Finance Association Early Career Forum (2026), RCF-ECGI Corporate Finance and Governance Conference (2026)

Invited Seminars: Drexel University (2026), Rutgers University Camden (2026), California State University Fresno (2026)

Do Measures of Bank Diversification “Measure Up”? (with Priyank Gandhi and Darius Palia)   

We analyze the effectiveness of several widely used measures of bank business line diversification in capturing the diversification effect (i.e., the idea that diversification reduces the idiosyncratic volatility of a portfolio of financial products). Contrary to the implications of the modern portfolio theory, current measures of bank business line diversification are poorly or positively correlated with idiosyncratic volatility. We instead propose the correlation-adjusted entropy (CAE) measure which accounts for number of business lines, income shares, and the imperfect correlations among business line incomes. CAE accurately captures the diversification effect, and variation in CAE coincides with exogenous shocks to bank business line diversification (the passage of Dodd-Frank and Economic Growth Acts), providing an important external validation for our measure. When we use CAE to revisit the question of how bank business line diversification impacts performance and risk, we find that diversified banks are more profitable, have lower probability of bankruptcy, and lower tail and systemic risk, in contrast to findings in the literature that document mixed results.

Conference Presentations: Financial Management Association (2026), FDIC Bank Research Conference (2026)*, Midwest Finance Association (2026)*, AFA Junior Faculty Mentoring Program (2026), Financial Engineering and Banking Society Conference (2026)*, Southern Finance Association (2026), Eastern Finance Association (2025), Financial Markets and Liquidity Conference (2025)*, RCF-ECGI Corporate Finance and Governance Conference (2025), International Moscow Finance Conference (2025), Contemporary Issues in Financial Markets and Banking Conference (2025), Friends of Women in Finance Symposium (2025), Sydney Banking and Financial Stability Conference (2024)*, Pacific Basin Finance, Economics, Accounting, and Management Conference (2024), AFA Committee on Racial Diversity Mentoring Program (2023)

Invited Seminars: Central European University (2025)*, University of New South Wales (2024)*, University of Melbourne (2024)*, University of Sydney (2024)*, Monash University (2024)*, Victoria University of Wellington (2024)*, Auckland University of Technology (2024)*

Awaiting the Prompt? Generative AI Uncertainty and Firm-Level Investment (with Priyank Gandhi, Simi Kedia, Juntai Lu and Jia Wei)

We document that firms with relatively greater workforce exposure to GenAI experience a decline in investment in the period following the release of ChatGPT on November 30, 2022. A one standard deviation increase in workforce exposure lowers the ratio of firm-level investment to lagged total assets by 0.57 percentage points, or nearly 6% of the sample mean. We argue that ChatGPT's release created a real option for firms to reorganize production around GenAI whose adoption cost, success probability, and competitive consequences were unknown at the outset. This parameter uncertainty, which scales with the firm's workforce exposure to GenAI, raises the option value of waiting and depresses current investment, when investments are irreversible. A real option model of firm investment with parameter uncertainty calibrated to bond market data (Gandhi, Lu, Pan, Plazzi, and Wei, 2026) reproduces the empirical investment response. Cross-sectional tests confirm the model's prediction that the decline in investment is concentrated in physical capital, persists through fiscal year 2025 without reversal, and is smaller for firms that have more redeployable assets. Additionally, we rule out that our results are driven by productivity substitution, talent hoarding, payout reallocation, an asymmetric-information flip, or managerial agency frictions.

Conference Presentations: Financial Management Association (2026)*, Southern Finance Association (2026)

Ambiguous Fedspeak and Risk Allocation in Corporate Bond Markets (with Tim Dong) 

We study how ambiguity in monetary policy communication a!ects institutional trading in corporate bond markets. We use large language models (LLMs) to read FOMC policy statements and generate buy, hold, or sell decisions for a representative bond trader, where the cross-model average captures the directional signal and cross-model disagreement measures policy-text ambiguity. We develop a simple equilibrium framework in which ambiguous central-bank communication reduces the risk-bearing capacity of short-horizon investors, creating price concessions that long-horizon insurers absorb. Consistent with this mechanism, we find that LLM disagreement, rather than the directional signal, predicts post-announcement insurer demand. Using 1.06 million cleaned NAIC insurer bond transactions around 76 FOMC policy statements from 2015 to 2024, a one-standard-deviation increase in disagreement is associated with approximately 5.7 percentage points more net buying by insurers over business days +1 to +5 relative to the pre-FOMC window. The effect is robust to controls for time, COVID-19, and the 2022–2023 hiking cycle, absent in placebo pre-event windows, and strongest for investment-grade, public, non-144A U.S. bonds and for life insurers. Overall, the results show that ambiguity in central bank communication not only a!ects prices but also reallocates risk across investors, and that LLM disagreement provides a useful empirical measure of this ambiguity.

Data Readiness and Firm Adjustment to Generative AI (with Tim Dong) 

The release of ChatGPT created a common shock to firms whose workforces were exposed to generative artificial intelligence. This paper asks whether exposed firms with stronger data complements adjusted differently after the shock. Using the workforce-exposure design of Gandhi et al. (2026), we compare prepared and unprepared firms within the high-exposure sample over fiscal years 2020–2024. Prepared firms are high-exposure firms with above-median GPT dataasset scores before the post-ChatGPT period. Difference-in-differences estimates show that prepared firms reduce employee growth by 1.9 percentage points in the specification with firm fixed effects, industry-year fixed effects, and standard firm controls. At the same time, prepared firms exhibit higher operating profitability: return on assets rises by 2.5 percentage points, gross margin rises by 7.7 percentage points, and EBITDA rises by about 302 million dollars in the same specification. The sales response is more sensitive to controls, so the main scale result is concentrated in labor rather than broad revenue contraction. A mechanism exercise shows that acquisition- and goodwill-based empire-building premiums that existed before GenAI are no longer statistically significant after the shock. The evidence suggests that data readiness changes the real effects of GenAI exposure by shifting firms toward leaner labor growth and higher operating efficiency.

*  Presented by coauthor.