AP @ HKUST (GZ)
leifuzhang 🌏 hkust-gz.edu.cn
hello.world 🌎 fin-tech.ai
​
FinTech [<finance, (t)cs> ≠ 0]
Information Economics in Financial Markets
Adversarial/Strategic ML/AI & Its Applications

If you are interested in working with me, please visit my research lab.
I am a financial economist with computer science training (in particular, machine learning)​. I did my postdoc in Chicago Booth's Finance Group (Supervisor: Zhiguo He (now at Stanford GSB))  and received a Finance PhD from Washington University in St. Louis (Committee: Jeremy Bertomeu, Jason Donaldson, Armando Gomes, Brett Green (Primary), Ohad Kadan, and SangMok Lee). I also studied at the Chinese University of Hong Kong (MPhil in Economics with a full scholarship; highest GPA award) and Peking University (Bachelor of Laws and Bachelor of Science; Guanghua scholarship).

​I currently focus on decentralized finance (DeFi). [I have learned a lot about cryptography from David Cash's course.] A recent paper of mine studies the oracle problem, perhaps the most fundamental issue in DeFi. [See my slides on two "celebrated" oracle manipulation attacks.] This paper makes two points: 1) there does not exist a perfect decentralized solution to the oracle problem; 2) machine learning can improve the current system dramatically by exploiting​ the high-dimensional structure inside oracles.

Research

Robust (Decentralized) Oracle Design
A perfect decentralized solution generally does not exist. Machine learning algorithms can dramatically improve the current system by exploiting the high-dimensional structure of oracles. [FinTech] [Machine Learning]

An oracle is a node that supplies blockchain smart contracts with external information that the blockchain itself cannot access. However, the oracle problem of ensuring that such information is accurate and resistant to manipulation remains a fundamental challenge. In this paper, I investigate the design of compensation mechanisms and aggregation methods for decentralized oracle networks consisting of adversarial and strategic nodes, to incentivize truthful reporting and robustly aggregate dispersed information into a consensus. I establish an impossibility result: Under a mild condition, fully decentralized robust compensation mechanisms—under which truthful reporting constitutes a strict Nash equilibrium for strategic nodes regardless of adversarial reporting strategies—do not exist. I then show that if the designer can occasionally access a verified signal about the ground truth, a simple squared-loss-based mechanism restores truth reporting as a dominant strategy. On the aggregation side, I show that the empirical coordinate-wise median widely used in practice may fail even without an adversary and ignores the intrinsically high-dimensional structure of decentralized oracle networks. By contrast, adversarial nodes that appear normal in every single dimension can be detected from a “global” view using high-dimensional robust machine learning algorithms, which yield a consensus with strong error guarantees. [Link to the Paper]

Selling Bubbles
​A secondary market joint with belief heterogeneity generates overpricing, encouraging entrepreneurs to reveal less information and sell bubbles. The model explains several features of the ICO market. [FinTech] [Information Economics]

How does an entrepreneur raise capital from investors with heterogeneous priors? When a secondary market exists, a price bubble arises and provides an incentive for the entrepreneur to manipulate investors’ beliefs through strategic communication. Under mild conditions, the amount of information disclosed decreases in disagreement. Without a secondary market, the price bubble vanishes, and the entrepreneur discloses more information. I discuss applications to emerging capital-raising methods, in particular initial coin offerings, and the regulatory implications for policymakers.​ [Link to the Paper]

Not Enough or Too Much? On the Financial System’s Equilibrium Transparency
A regulator without commitment power leads to excess transparency. [Information Economics]

In general, a regulator’s commitment power is necessary for implementing the optimal transparency policy in the financial system. The literature points out that lack of commitment power leads to excess opacity if investors additionally observe a sufficiently precise signal. Instead, this paper finds that an equilibrium featuring excess transparency always exists and survives natural refinements under certain conditions. Moreover, the optimal transparency policy and if the regulator needs commitment power to implement it are sensitive to the exogenous information structure. [Link to the Paper]

Mutual Fund Portfolio Constraints: Carrying Coals to Newcastle? (with H. Liu)
Investors impose portfolio constraints to restrain fund managers’ excessive risk-taking. These constraints do not appear binding because they transform local optimums to become constrained global optimums. [Information Economics] [Contract Theory]

Investors often impose short sale and no-leverage constraints on many mutual funds. However, we rarely observe these constraints bind in the portfolios of these funds. Are investors carrying coals to Newcastle? In a principal-agent framework in which the risk-averse agent (manager) is protected by limited liability, we find that such constraints are necessary to curb the manager’s excessive risk-taking due to limited liability. These constraints do not appear binding because there is an interior portfolio weight (between 0 and 1) that is locally optimal, and the constraints transform it to become a constrained global optimum. Moreover, this “non-binding puzzle” is exclusive for the “low-ability” manager, which suggests a new investment skill indicator. [Link to the Paper]

Robust Stress Test Design with Optimal Information Acquisition (Work in Progress)
Investors’ optimal information acquisition restores the optimality of a “pass/fail” test with one cutoff. ​[Information Economics]

​I study the optimal design of stress tests with adversarial coordination and optimal information acquisition. Investors may acquire additional information after observing the test result disclosed by the regulator. Information acquisition is “optimal” in the sense that investors have the “flexibility” to choose what information to acquire optimally. In contrast to the literature without optimal information acquisition, which usually obtains complicated optimal tests, I show a “pass/fail” test with one cutoff is optimal.