Working paper
As AI tools become common inputs in economic decisions, public actions
become increasingly difficult to interpret. A sequence of identical choices
may reflect a cluster of independent data or merely repeated reliance on the
same AI signal. This creates a fundamental trade-off: shared AI usage
correlates actions and accelerates herding, but it also dilutes the
informational value of public histories, pushing decision-makers back to
their own private signals. I formalize this trade-off in a BHW-style social
learning model where some agents observe a shared AI signal and others
receive independent signals. With unobserved agent types, public history
endogenously reveals both the state of the world and the underlying sources
of past actions. I find that AI transforms social learning in three regimes:
low-precision AI exacerbates the risk of incorrect cascades; high-precision
AI dictates long-run social outcomes; and intermediate precision induces
non-monotonicities, meaning better AI can locally degrade social learning.
Together, the results show that evaluating AI by precision alone is
incomplete.
Draft available upon request. Slides
Working paper
Shared Private Information and Information Aggregation
Low-cost shared private signals, such as recommendations from FinTech
platforms, robo-advisers, or generative-AI tools, are privately accessed yet
commonly received. We embed such a signal into a Grossman-Stiglitz (1980)
market in which traders choose among an independent signal, a shared signal,
and remaining uninformed. Three findings frame the analysis. First, a
cheaper or more precise shared signal robustly improves market depth in our
numerical experiments, but it does so at the cost of greater price exposure
to errors in the shared signal, a liquidity-shared-signal-error tradeoff.
Second, equilibrium is unique whenever some traders remain uninformed; on
the fully informed boundary, multiple equilibria can arise when the shared
signal is relatively noisy, while precise shared signals restore uniqueness
under mild conditions. Third, price informativeness is regime-dependent: in
the partial-adoption regime, a cheaper shared signal leaves it unchanged; on
the fully informed boundary, it can fall, rise, or be non-monotone depending
on the shared signal's precision relative to independent research.
Draft available upon request. Slides
Master thesis
Cyber Risk and Interbank Network
This research extends Allen and Gale's (2000) work to explore the interplay
between cyber risk and financial networks, with a focus on the liquidity
channel through interbank deposits. It reveals coordination failures and
underinvestment in cybersecurity in certain equilibria, emphasizing the need
for minimum cybersecurity requirements. Additionally, in the presence of
significant cyber risk, the interbank network is unable to effectively share
liquidity risk. The study further highlights that underinvestment in
cybersecurity increases the likelihood of the financial network deviating
from its optimal financial structure.
Draft available upon request.