Conference Agenda

Machine learning factor models of the cross section of stock returns have produced spectacular results, which are explained by two different ingredients: expanding the information set and allowing for more flexible function forms. We disentangle the value-added of each ingredient, considering a variety of empirical settings: from highly stylised to realistic. We show that the benefit of both ingredients declines when moving from standard settings in the literature to more realistic settings that exclude microcaps, remove look-ahead bias on yet-to-be-published factors, account for transaction costs, and exclude short positions. While the value of nonlinearity disappears even before imposing transaction costs or short-sale constraints, the value of an expanded information set is more persistent. Our feature importance analysis reveals that characteristics like firm size and short-term reversal - crucial predictors in standard settings - lose most of their value once investability constraints are considered. These findings challenge claims about the universal benefits of machine learning sophistication, demonstrating that real-world implementation constraints fundamentally alter which model ingredients improve portfolio performance.

Machine learning (ML) methods are highly flexible, but their ability to approximate the true data-generating process is fundamentally constrained by finite samples. We characterize a universal lower bound, the Limits-to-Learning Gap (LLG), quantifying the unavoidable discrepancy between a model’s empirical fit and the population benchmark. Recovering the true population R2 , therefore, requires correcting observed predictive performance by this bound. Using a broad set of variables, including excess returns, yields, credit spreads, and valuation ratios, we find that the implied LLGs are large. This indicates that standard ML approaches can substantially understate true predictability in financial data. We also derive LLG-based refinements to the classic Hansen and Jagannathan (1991) bounds, analyze implications for parameter learning in general-equilibrium settings, and show that the LLG provides a natural mechanism for generating excess volatility.

There is increasing awareness of how local media affects financial markets, but also of the endogeneity of media coverage. We separate the causal impact of local media on financial markets from selection effects using a new, hand-collected database of newspaper closures. We find that at least 29% of local newspaper closures are driven by distress, and thus, likely endogenous to local economic conditions. Return volatility and idiosyncratic risk decrease significantly after non-distress-driven newspaper closures, but increase after distress-driven closures, suggesting the presence of substantial selection effects. We find similar patterns for liquidity and trading. Once we account for selection, the estimated impact of local newspapers on volatility increases by over 40%. The reduction in volatility after non-distress-driven newspaper closures is larger for stocks subject to greater information frictions, lower national media coverage, firms located in remote areas, firms with a more concentrated geographic presence, and during recessions. All these tests suggest that investor information processing is the main channel that drives our results. Our findings highlight that the effect of media on financial markets may be larger than previously documented.

This paper studies how the exit of equity analysts from public-facing sell-side roles

affects investment behavior, price efficiency, and the firm-level information environment.

Using a novel dataset tracking analyst career transitions and employer characteristics,

we show that analysts who join buyside institutions tend to be more accurate and

influence portfolio allocation in stocks they previously covered. Despite this, price

efficiency deteriorates following analyst exits, especially for complex or thinly covered

firms. We also find increased earnings surprises and greater forecast dispersion. The

results highlight a tradeoff between private gains and public costs in the production of

financial information.

Large language models can trade: they execute coherent strategies, generate realistic market dynamics, and can be weaponized for manipulation. Using an open-source simulated stock market with persistent order books, short selling, dividends, and social feeds, I document three findings. First, LLMs consistently adhere to their assigned strategies, functioning as value investors, momentum traders, or market makers per their instructions. Second, their markets exhibit features of real financial markets, including price discovery and bubble formation. Third, when instructed to manipulate, they generate persuasive messages that create feedback loops, driving prices far from fundamentals. The framework enables testing financial theories that lack closed-form solutions and running experiments that would be too costly or unethical with human participants.

AI agents can emulate the beliefs of human survey respondents. We leverage this idea at scale to examine how investor disagreement emerges in response to firm news and to measure such disagreement at high frequency. We endow a local large language model (Llama 3) with over 200 demographically representative investor personas and elicit their sentiment toward S&P 500 firm-specific news headlines from 2010–2025. LLM personas disagree in economically meaningful ways: disagreement is largest for social and governance-related news, and smallest for hard news tied to firm fundamentals. The measure aligns with human survey evidence and traditional uncertainty measures, yet it is distinct from social-media disagreement. Turning to market outcomes, LLM-derived disagreement is strongly associated with same-day and next-day abnormal trading volume. Our main findings are stable across the model’s pre- and post-training windows.

We measure the elasticity of demand for Treasuries at auction directly from bidding data.

From 1992 to 2010, demand for Treasuries was surprisingly elastic: a 1% increase in the supply

of Treasuries relative to the amount outstanding corresponded to a 2 basis point increase in

long-term yields. Since 2010, demand has become almost five times more inelastic, implying

that yields now rise by 9 basis points per 1% increase in supply. This deterioration of demand

is also apparent in the secondary market. Prior to 2010, long-term yields declined on average

by 1.5 basis points after auctions and these declines have been concentrated in auctions with

strong investor demand. After 2010, this trend has reversed and yields no longer fall after

auctions. This weaker demand for Treasuries coincides with less foreign investor demand

and reduced secondary market liquidity.

We evaluate the enforcement of information barriers—China Walls—within conglomerates. Our setting is the 23 million trades in 2019-2024 in the Israeli Shekel market, where the US SEC imposes China Walls around dealers. Our difference-in-differences design compares the trade volumes and profits of funds that are affiliated with, clients of, or entirely unrelated to a dealer around the days when the dealer is especially likely to hold valuable information. Dealers never trade or share information with their affiliate funds, despite that they do share information with their clients, and funds within the same conglomerate do so among themselves. Our findings persist in crisis and noncrisis periods and across granular cells of fund and asset characteristics. From a back-of-the- envelope calculation, imposing China Walls around funds would eliminate $23.7 billion in trades. We reveal a remarkable regulatory capacity to control information flows within conglomerates.

We study how shifting data ownership from lenders to borrowers would affect data production, fintech entry, and resource allocation efficiency. We categorize borrowers' data into portable and non-portable elements. Portable data, which requires active production and is transferable, will provide more accurate information. In our model, an incumbent bank, endowed with non-portable data, determines the amount of portable data produced, while a potential fintech entrant is more efficient at analyzing it. Our findings indicate that the shifting of data ownership will reduce portable data production, thereby diminishing resource allocation efficiency, despite encouraging fintech entry. The bank's non-portable data endowment and fintech's analytical efficiency have non-continuous, non-monotonic effects on portable data production.

This paper studies the impact of digitized invoices on small business lending through China’s Golden Tax Phase III Project, which established a unified tax-reporting system and digitized VAT invoices. Using the project’s staggered provincial rollout as a natural experiment and a difference-in-differences (DID) design, I find that small businesses in treated provinces obtain substantially larger credit lines, pay lower interest rates, and enjoy longer maturities. Banks also report higher measured incomes and assign higher internal credit scores to these firms. I trace these improvements to the elimination of fraudulent invoices, which endows banks with more reliable hard information for income assessment and risk-based pricing. As a result, digital invoices mitigate information asymmetries in lending, allowing banks to extend more and cheaper credit to firms lacking hard information. Consistent with this mechanism, the largest credit gains occur in invoice-intensive industries; the most opaque borrowers (e.g., individually owned businesses) and small banks experience the greatest benefits. Finally, macro-level credit data from the People’s Bank of China corroborate my micro-level findings. Overall, this study highlights the critical role of digital tax infrastructure in alleviating financing frictions for small businesses and underscores its broader importance for financial inclusion and economic growth.

Traditional Pigouvian and Coasean approaches to regulating externalities falter under uncertainty and disagreement, as they require precise knowledge of externality costs or frictionless bargaining. We develop a system of outcome-contingent "Coasean transfers" that leverage uncertainty and agents' heterogeneous information to achieve efficient outcomes without requiring disclosure of private information. These transfers operate through payments consisting of a quantity component (the gap between an agent's action and the market average) and a price component (tied to publicly observable aggregate outcomes). We prove a market-equivalence result: the optimal transfer pricing schedule corresponds to the equilibrium price in a hypothetical Coasean market for the externality. The equilibrium allocation under Coasean transfers is team efficient and strictly dominates traditional tools like Pigouvian taxes. These transfers are budget-balanced and informationally light, requiring only normative objectives, not private information or signal structures. They incentivize information acquisition, remain robust when agents distrust each other's information, and are politically viable, receiving ex-ante unanimous support.

We show that even when firms' sustainability claims are costless and non-verifiable, they can still retain (partial) credibility. Non-babbling equilibria arise in a standard cheap-talk model when investors face uncertainty about both cash flows and the sustainability attribute. In such equilibria, sustainability claims are endogenously negatively correlated with cash flows and positively correlated with the sustainability attribute, even if the two are ex ante independent. The negative cash-flow implication of a claim serves as an endogenous cost, enabling credible communication. We characterize how disclosure behavior and credibility vary with the level of cash-flow uncertainty and provide novel empirical evidence supporting this mechanism. Our results shed new light on firms' communication of sustainability information in the absence of regulatory interventions.

Using the staggered adoption of data technologies providing firms with novel insights about their customers, we show that their investment becomes significantly less sensitive to non-fundamental stock price movements after adoption. The effect is consistent with data technologies improving managers’ internal information about future demand, thereby reducing their reliance on stock market signals. This “replacement effect” is robust to alternative explanations and reverses under data-privacy restrictions.

Our findings suggest that the diffusion of data technologies weakens the informational role of stock prices in guiding real investment decisions.

Alternative data has reshaped financial markets, yet it arises outside traditional disclosure channels, beyond market participants' control. We exploit Apple’s App Tracking Transparency as a privacy-driven shock to alternative data generation, revealing a previously overlooked vulnerability in financial markets. The policy weakens the predictive power of mobile traffic for firm performance and mutual fund trading. Funds reliant on such data lose their edge in exposed stocks, while analyst forecasts citing mobile traffic become less accurate and elicit weaker reactions. Firms more dependent on these market participants face greater information frictions.

This paper examines how effortless intuition (System 1) and deliberative reasoning (System 2)

jointly influence investor decision-making. We construct a novel dataset of livestream

promotional events linked to the initial offerings of mutual funds in China between 2020 and

2024. These events occur before any performance records or portfolio disclosures become

available, providing a clean setting to isolate the effects of different cognitive processes. We

find that investors’ intuitive responses—elicited by the dynamic emotional displays of

livestream presenters, including vocal tone, facial expressiveness, and body movement—are

positively and significantly associated with fund subscriptions. However, this effect weakens

when the livestreams convey richer information or feature fund managers, conditions that

engage more deliberative reasoning. Our study provides novel evidence on how intuitive

(System 1) and deliberative (System 2) processes interact to influence investor choices,

offering a foundation for developing a positive theory of investor behavior.

We study whether AI-extracted facial features from borrowers’ photos can serve as a scalable proxy for “soft” information missing from traditional credit models, such as conscientiousness, patience, and self-control. These traits influence financial behavior but are rarely captured in administrative data. Linking LinkedIn photos and employment and education records to voter registration and Experian data for over one million U.S. borrowers, we find that facial embeddings add significant predictive power for default risk beyond standard observables such as as credit scores, gender, and race. The incremental value is largest for younger, lower-income, and thin-file borrowers, where traditional credit scoring technology is least informative. A separate model mapping facial images to perceived Big Five personality traits reveals personality as one mechanism through which images proxy for soft information. These results suggest that facial embeddings capture stable behavioral traits absent from standard credit data, as well as perceived attributes that may influence how individuals are treated by others. While such models offer new insight into the role of personality and soft information in credit markets, their use in screening raises important concerns about fairness, privacy, and autonomy.