Research


Research Interest

Publications 

Papers

We introduce a general approach for analyzing large-scale text-based data, combining the strengths of neural network language processing and generative statistical modeling to create a factor structure of unstructured data for downstream regressions typically used in social sciences. We generate textual factors by (i) representing texts using vector word embedding, (ii) clustering the vectors using Locality-Sensitive Hashing to generate supports of topics, and (iii) identifying relatively interpretable spanning clusters (i.e., textual factors) through topic modeling. Our data-driven approach captures complex linguistic structures while ensuring computational scalability and economic interpretability, plausibly attaining certain advantages over and complementing other unstructured data analytics used by researchers, including emergent large language models. We conduct initial validation tests of the framework and discuss three types of its applications: (i) enhancing prediction and inference with texts, (ii) interpreting (non-text-based) models, and (iii) constructing new text-based metrics and explanatory variables. We illustrate each of these applications using examples in finance and economics such as macroeconomic forecasting from news articles, interpreting multifactor asset pricing models from corporate filings, and measuring theme-based technology breakthroughs from patents. Finally, we provide a flexible statistical package of textual factors for online distribution to facilitate future research and applications.

Keywords: Textual factors, LLMs, Topic models, Word embedding

We argue that firm-level agency conflicts not only counter the role of network structure in the propagation of shocks but they can have a significant impact on system-wide behavior that differs from those predicted based on network structure alone. This implies that corporate governance can play an important role in macro fluctuations. We consider a collection of firms linked through equity cross-holdings whose managers can make investment decisions in response to an exogenous shock. Prior work concludes that more integrated networks amplify shocks. We find that if managers are subject to default costs or limited liability, this effect is reversed because their investment decisions mitigate the spread of an initial shock. In the face of moral hazard, however, their investment choices amplify an initial shock. In particular, when the network is fully diversified, the aggregate effect of idiosyncratic shocks does not diminish as received wisdom would suggest.

Keywords: Agency conflicts, Financial networks, Ownership networks, General equilibrium

The centrality in a network is a popular metric for agents' network positions and is often used in regression models to model the network effect on an outcome variable of interest. In empirical studies, researchers often adopt a two-stage procedure to first estimate the centrality and then infer the network effect using the estimated centrality. Despite its prevalent adoption, this two-stage procedure lacks theoretical backing and can fail in both estimation and inference. We, therefore, propose a unified framework, under which we prove the shortcomings of the two-stage in centrality estimation and the undesirable consequences in the regression. We then propose a novel supervised network centrality estimation (SuperCENT) methodology that simultaneously yields superior estimations of the centrality and the network effect and provides valid and narrower confidence intervals than those from the two-stage. We showcase the superiority of SuperCENT in predicting the currency risk premium based on the global trade network.

Keywords: Supervised learning, Network regression, Two-stage estimation, Measurement errors

The finance–growth nexus has been a central question in understanding the unprecedented success of the Chinese economy. Using unique data on all the registered firms in China, we build extensive firm-to-firm equity ownership networks. At the end of 2017, there are 5.6 million firms belonging to at least one network, and 35 million out-of-network firms. Entering a network and increasing network centrality leads to higher firm growth, with the effect of global centralities on growth strengthening over time. The positive effects of network positions become more pronounced for more productive firms and those with more financial constraints, but weaker for SOEs. Firms’ network positions promote growth through both financing and resource-sharing channels. The 2008-09 global financial crisis (GFC) led to a sudden slump in China’s export sectors and exogenous shocks to the equity networks. Non-export firms connected with the export sectors in the equity networks before the GFC gained centrality relative to firms from the same industries but unconnected with export-sector firms in the networks, and greater centrality resulted in higher growth in the post-GFC period. Our evidence also shows that the RMB 4 trillion stimulus launched by the Chinese government in response to the GFC partially “crowded out” the positive network effects.

Keywords: Ownership network; Equity capital; Firm growth; Bank credit

We find that good news extracted by ChatGPT from the front pages of Wall Street Journal can predict the stock market and is related to macroeconomic conditions. Consistent with existing theories, investors tend to underreact to positive news, especially during periods of economic downturns, high information uncertainty and high novelty of news. In contrast, the negative news is only associated with contemporaneous returns. Traditional methods of textual analysis, such as word lists and large language models like BERT, can barely find any predictability. In short, ChatGPT appears the best AI in discerning economic-related news that drive the stock market.

Keywords: LLMs, ChatGPT, Textual analysis, NLP, Return predictability

We re-examine the state sector and its role in the Chinese economy through the equity ownership networks of 40 million firms from 1990 to 2017. Our measure uncovers a broader array of state-owned enterprises (SOEs) than the existing measures. State ownership witnesses dual trends: decentralization—the scale of capital from the central government has been declining—and indirect control—a rising number of SOEs with provincial and city government ownership stakes and increasing hierarchical distance between existing and new SOEs from their ultimate state owners. The increase in hierarchical distance and expansion of SOE networks can be explained by the openings of new high-speed rail routes. Firms with minority state ownership stakes and greater hierarchical distance to the central and provincial government owners have better performance as measured by higher profitability and productivity.

Keywords: state-owned enterprises; state capital; decentralization; ownership network; hierarchical distance.

Using business registry data from China, we show that internal capital markets in business groups can play the role of financial intermediary and propagate corporate shareholders’ credit supply shocks to their subsidiaries. An average of 16.7% local bank credit growth where corporate shareholders are located would increase subsidiaries investment by 1% of their tangible fixed asset value, which accounts for 71% (7%) of the median (average) investment rate among these firms. We argue that equity exchanges is one channel through which corporate shareholders transmit bank credit supply shocks to the subsidiaries and provide evidence to support the channel.

Keywords:Tiered intermediation,Equity networks, Equity investment, Investment, Bank credit shock

The speed at which the US economy has recovered from recessions ranges from months to years. We propose a model incorporating the innovation network, the production network, and cross-sectional shocks and show that their interactions jointly explain large variations in the recovery speed across recessions in the US. In the model, besides the production linkages, firms learn insights on production from each other through the innovation network. We show when the innovation network takes a low-rank structure, there exists one key direction: the impact a shock becomes persistent only if the shock is parallel to this key direction; in contrast, the impact declines quickly if the shock follows other directions. Empirically, we estimate the model in a state-space form and document a set of new stylized facts of the US economy. First, the innovation network among sectors takes a low-rank structure. Second, the innovation network has non-negligible overlap with the production network. Third, recessions with slow recovery are those witnessing sizable negative shock to sectors in the center of the innovation network. Such network structures and the time-varying sectoral distribution of the shocks can well explain the large variation in the recovery speed across recessions in the US. Finally, to emphasize the prevalence of the channel, we explore the application of the theory in asset pricing.

Keywords: Business cycles, Innovation networks, Production networks, Sectoral shock, Persistence,

This paper provides evidence that network complexity limits investors' ability to process non-local information, through the lens of return cross predictability. Using firm-to-firm citation networks, we find that the non-local indirectly-linked firms can well predict the return of the focal firm, while the predictability of the local directly-linked firms is weak. A long-short strategy using the indirect links yields a risk-adjusted monthly alpha of 198 (164) basis points with equal (value) weights. We further find that (i) the indirect citation links are much more complex than direct ones, (ii) the magnitude of cross predictability increases with the degree of link complexity, (iii) institutional investors don't adjust their positions in a stock with complex links, but in one with simple links immediately, (iv) firms with more complex links receive more public attention, are much larger in size, and exhibit less idiosyncratic volatility than those with simple links, (v) there is little role of the usual proxies for limited investor attention and arbitrage cost in explaining our anomalies, once controlling for the linking complexity, and (vi) there are no differences in expected returns of stocks with various link complexity.

Keywords: Return cross predictability, Citation networks, Information complexity, Link complexity, Limited investor attention

We study the impact of exposure to robotics and automation on corporate innovation, which informs how innovation begets innovation. Using advanced language models, we document that firms with high automation exposure witness a decline in technology similarity over time and significantly shift innovative activities toward AI-related ones, which intuitively complement automation. The shift is more pronounced for firms with greater AI-related research experience or generate more data. Furthermore, AI patents are more costly than non-AI patents in various dimensions, including team size, researchers' labor input and inventors' originality. Consequently, firms with high automation experience a significant rise in R&D expenditure but a decline in the number of new patents in subsequent years. Against the backdrop of rising automation, this explains the puzzling observation in the literature that at the aggregate level, firms seem to become less innovative in recent years despite greater R&D expenditures. Finally, we present a simple dynamic equilibrium model to rationalize such innovation shifts.

Keywords: AI, Corporate innovation, Language models, Technology, Robot

We consider the dynamic joint assortment and pricing problem for retailers who aim to maximize cumulative revenue at regular time intervals. %(e.g., weekly or monthly). The main question investigated in this paper is, in addition to customer's choice behaviors, how to account for varying customer arrivals (market size) influenced by assortment and pricing decisions---a factor that traditional MNL models assume to be fixed, often leading to suboptimal decisions.  To address this problem, We propose a generalized multinomial logit (MNL) model that combines the discrete choice model  {capturing} customer preferences \zwR{and} a nonhomogeneous Poisson model capturing the influence of the offered products and prices on customer arrivals. Building on this model, we develop a new online joint assortment-pricing policy, Poisson-MNL ($\PMNL$), based on the upper confidence bound (UCB). We establish its (near) optimality by proving a non-asymptotic regret bound of $\sqrt{T}\log{T}$ and providing a matching lower bound (up to $\log{T}$). The proof requires technical innovations to extend existing MNL-bandit results to sub-exponential random variables. Simulation studies highlight the importance of accounting for the endogeneity \zwR{of}customer arrival rates, assortment, and pricing. The results show that $\PMNL$ effectively learns customer choice and arrival rates and provides joint assortment-pricing decisions that outperform other \zwR{models with} fixed arrival rates.

Keywords: Contextual bandits; Assortment problem; Pricing problems; Reinforcement learning, Customer arrival; On-line decision-making.

The emergence of Large Language Models (LLMs, e.g., ChatGPT) is believed to revolutionize writing. We investigate the impact of generative-AI-assisted revisions on academic writing, focusing on researchers' heterogeneous usage and their convergence in writing. Using a dataset of over 627,000 academic papers from arXiv, we develop a framework for identifying ChatGPT-revised articles by training a set of prompt- and discipline-specific language models ourselves. We first document the widespread usage of GPT, with significant heterogeneity across disciplines, gender, ethnicity, and academic experience, as well as a rapid evolution in writing style over time. Moreover, our analysis reveals that LLM usage significantly improves writing outcomes in terms of clarity, conciseness, adherence to formal writing rules, etc., and the improvements depend on nuanced types of usage. Finally, a difference-in-difference analysis suggests that while the birth of LLMs leads to a convergence in academic writing, adopters, males, non-native speakers, and junior researchers adjust their writing style the most to resemble that of more experienced scholars

Keywords: Generative AI, LLMs, Technology adoption, Writing style.

Selected Project in Progress


Media Reports

Academic Services

To students

I hope the following provides insights into my research interests and agenda.

1. The first delves into the tangible and intangible linkages between firms, examining their implications on corporate finance, governance, monetary policy, and the broader economy, an area in which I have a special interest

2. The second explores the employment of statistical learning, deep learning, and reinforcement learning techniques in portfolio or asset management, enriched by regular interdisciplinary discussions with my co-authors from fields like finance, statistics, and computer science across various institutions.

3. The third investigates the interplay between non-structural data (like text, video, and graph) with deep learning and asset management.

My students have been placed in the very top institutions like UPenn Wharton, Princeton, MIT, Morgan Stanley, Jane Street, Ubiquant Investment, Optival etc.,