Machine Learning and AI in Finance:
Whence LASSO? A Rational Interpretation with Wen Chen and Liyan Yang
Management Science (2025) - Special Issue on AI for Finance and Business Decisions
Based on robust optimization under fat-tailed model risk, we provide an economic rationale for adopting machine learning techniques such as LASSO and Elastic Net in arbitrage trading. Our model explains the robust out of sample performance of LASSO and offers a solid, economic interpretation rooted in Bayesian rationality. We show that LASSO-based strategies can enhance and sustain traders’ total profit by mitigating competition, even as the number of traders approaches infinity. This reveals a new mechanism for limits to arbitrage.
*Presented at WFA, EFA, CICF, AI and Big Data in Accounting & Finance Conference, CFEA Special Session on Machine Learning in Finance, etc.
How Does Benchmarking Affect Market Efficiency? The Role of Learning Technology
with Wen Chen and Yajun Wang, Journal of Financial and Quantitative Analysis, 2nd Revise & Resubmit
This paper applies information theory to a model of attention allocation and portfolio choice, examining separative versus integrative learning adopted by asset managers with benchmarked incentives. Our analysis shows that under integrative learning (which is optimal and dimension-reducing), stronger benchmarking demands on the riskier asset can increase that asset’s price informativeness and even improve overall market efficiency, contrary to the implications of separative learning typically assumed in the literature. AI technologies such as large language models rely on advanced attention mechanisms to process large volumes of information. These tools can facilitate integrative learning, enabling investors to extract portfolio-wide rather than asset-specific signals, thereby improving market efficiency.
*Presented at EFA (European), MFA, CICF, among others. Featured in 2024 news coverage of Phys.org
Seeing is Believing: Annual Report Visuals and Stock Returns (under review)
with Wesley Deng, Lei Gao, and Guofu Zhou
This paper develops a novel model to guide AI-powered empirical analysis of firms’ use of visuals in annual reports. We document that firms earn 3–5% abnormal returns after adding visuals, accompanied by a surge in institutional investor attention and holdings. Effects are strongest for visuals related to innovation and technology. Firms adopting R&D-focused visuals experience increases in patent grants and innovative output, suggesting visuals mitigate investor inattention and effectively convey nuanced fundamental information to markets.
*Presented at EFA (European), SFS Cavalcada (Asia-Pacific), CICF, AI in Finance Conference, Bretton-Woods Accounting & Finance Conference.
Market Structure and Frictions:
Dynamic Duopolistic Competition with Sticky Prices with Steven L. Heston
Operations Research (2025)
This paper addresses a long-standing paradox in the textbook model of Fershtman and Kamien (Econometrica, 1987). We show that FK’s continuous-time formulation of dynamic duopoly prevents proper analysis of the frictionless limit, yielding their paradoxical conclusion that frictional effects persist even as frictions vanish. Our discrete-time formulation corrects this erroneous result which has repeated in literature and propagated across textbooks: Mehlmann (1988), Dockner et al. (2000), Engwerda (2005), Kamien and Schwartz (2012), Lambertini (2018). We propose a general, rigorous approach for limit analysis of economic models with small frictions (e.g., transaction costs). This is essential for evaluating the impact of policies that deliberately introduce small frictions as “sand in the wheels of commerce”.
Do Position Limits on Futures Trading Benefit Commodity Markets?
with Wen Chen and Yajun Wang
Regulators believe that imposing position limit on speculators would dampen futures price volatility and prevent market manipulation. Yet, we show that these do not hold because of two unintended consequences. First, speculative position limits are more likely to bind than regulators anticipate, because such constraints can serve as a coordination device for speculators to amass market power and hence harm hedgers — the very group this rule intends to protect. Second, position limits reduce liquidity in the futures market even when not binding. This has a spillover effect, through the information channel, which can exacerbate volatility and reduce efficiency in commodity markets.
Algorithmic Arbitrage with Fat Tails
with Wen Chen and Liyan Yang
This paper studies algorithmic traders who use simple threshold-based momentum strategies for low latency and ease of implementation. When liquidity conditions permit, an insider can strategically exploit these algorithmic traders, causing sharp price overshooting, volatility spikes, and temporary mispricing — patterns reminiscent of momentum ignition, flash crash, and quant meltdown in modern markets. The model justifies widely used momentum trading algorithms and illustrates how crowded high-frequency trading can create predictable, exploitable price distortions. Our results offer insights into market instability and inform the regulation of manipulative trading practices.
*An earlier version of this project, circulated under the title "Statistical Arbitrage with Uncertain Fat Tails", was presented at the 2019 NYU Stern Microstructure Meeting, the 4th PKU-NUS Annual International Conference on Quantitative Finance & Economics, and 2021 WFA.
Informed Trading and Internal Capital Markets: An Information-Theoretic Perspective
This paper applies information theory to a rational-expectations equilibrium model with informed trading and internal capital allocation. Informed traders extract monetary (private) value from private signals, while a manager can learn from prices to reallocate capital across projects, generating a distinct real (public) value of information for common shareholders. These two values move in opposite directions: more informative prices lower private trading rents but improve capital allocation. Managerial inattention, adjustment costs, and financial constraints weaken this transmission channel and can increase private profits by preserving exploitable heterogeneity. The model yields testable predictions that link informed speculation in financial markets to its real feedback effect on corporate investment.
Fundamental Asset Pricing Theory:
What if the Long Forward Rate is Flat? (under review)
This paper draws on Dybvig, Ingersoll, and Ross (1996) to show that the long forward rate is constant in asymptotically stationary and ergodic economies, extending the applicability of long-run pricing kernel factorizations by Alvarez and Jermann (Econometrica, 2005) and Hansen and Scheinkman (Econometrica, 2009). We identify a limitation of a popular term-structure modeling method, which requires a path-independent pricing kernel and hence implicitly restricts the market price of risk to equal the long-term bond volatility. In contrast, the general equilibrium of Cox, Ingersoll, Ross (Econometrica, 1985) is free of this restriction because its pricing kernel is path-dependent.
Limits to Leverage in Frictionless General Equilibrium with Albert S. Kyle
We show that in frictionless general equilibrium, an arbitrary constant beta asset (e.g., the money market fund) cannot in general serve as a valid numéraire that guarantees martingale pricing. We derive a tractable condition that links the admissibility of a candidate numéraire to the boundary accessibility of state variables under the associated change of measure. Our analysis reveals that financial leverage can be endogenously limited, even absent frictions such as trading costs, borrowing constraints, or default risk. This limit arises since excessively leveraged assets with free redemption options can violate the first-order condition for intertemporal substitution. Our results challenge the classical view that leverage can be scaled arbitrarily in frictionless economies, providing a unified explanation for why leveraged funds or ETFs often underperform their target multiples and why even safe assets may deviate from their stated values under liquidity stress.
On the Puzzle of Bond Pricing in Cox-Ingersoll-Ross Model
An unresolved puzzle in the CIR term structure model is that multiple plausible bond price solutions arise when the risk-neutral measure fails to be an equivalent martingale measure (EMM). We derive the unique economically relevant bond price and show that even when risk-neutral pricing breaks down, there are neither arbitrage opportunities nor price bubbles: if the correctly priced T-bond is used as numéraire, the induced T-forward measure is indeed an EMM that rules out arbitrage. This result addresses a pervasive misconception in financial literature and textbooks that equates risk-neutral pricing (a computational device) with the absence of arbitrage (a first principle). In particular, we show that the long forward measure is always an EMM and yields a more reliable no-arbitrage pricing framework. This demonstrates the advantages of the long-run pricing kernel factorization theory by Hansen and Scheinkman (Econometrica, 2009).
— Last updated: December 01, 2025 —
