Research

This paper examines how uncertainty affects liquidity provision by incorporating macro, valuation, and execution uncertainties in a limit-order-book model. The model yields a spread decomposition and a hidden-share condition: hidden liquidity increases with macro uncertainty but decreases with valuation uncertainty because reduced time priority of hidden orders raises non-execution and selection costs. The macro effect is heterogeneous in equilibrium: when macro uncertainty is high, liquidity providers provide hidden liquidity to easier-to-value stocks and remove from harder-to-value stocks to minimize overall portfolio exposure. Consistent with these predictions, I show that increases in macro uncertainty raise hidden liquidity for easier-to-value firms and reduce it for harder-to-value firms. Robustness tests with alternative proxies and exogenous shocks from the COVID-19 pandemic and the temporary suspension of NYSE floor trading support a causal interpretation.

Nonbanks account for a substantial share of U.S. mortgage applications, a shift linked to technology and regulatory differences. We match Consumer Financial Protection Bureau complaint records to loan-level Home Mortgage Disclosure Act data for 2018 to 2023, covering 44.4 million mortgage applications. In a sample aligned with the Bureau’s institutional coverage, banks and their mortgage subsidiaries receive about 3.52 times as many origination-stage complaints per application as independent mortgage companies, conditional on borrower composition, product mix, and geography. The difference persists under alternative exposure measures and complaint definitions, although its magnitude is sensitive to large lenders. Banks originate a larger share of mortgages directly through retail channels, and the complaint-rate difference is smaller among predominantly retail lenders. Finally, among recorded origination-stage complaints, banks and their mortgage subsidiaries miss the response deadline on 0.3 percent of complaints, compared with 3.9 percent at independent mortgage companies.

We document substantial instability in the relation between equity-market policy uncertainty and stock price efficiency. Examining all U.S. common stocks over 2001–2024 using a validated daily price-delay measure, benchmarked against Hasbrouck pricing errors through 2015, we find that uncertainty was associated with higher price delay, wider bid–ask spreads, and lower turnover during 2001–2009. A one-standard-deviation increase in uncertainty raised aggregate price delay by roughly nine percent of its mean. Structural break tests date the disappearance of the relation to October 2009. The post-break coefficient remains episodically unstable for several years before settling near zero by late 2016, and the null persists through the COVID-19 pandemic. The efficiency and liquidity responses to uncertainty change around the same break, while the uncertainty index remains informative about market-relevant uncertainty. These patterns are consistent with a transformation in liquidity provision, although the aggregate evidence does not identify that transformation as the unique cause of the break. The SEC Tick Size Pilot's randomized market-design restrictions reduce liquidity and slow price discovery but do not restore uncertainty sensitivity. The findings demonstrate that pooled-sample estimates can obscure economically distinct market regimes and caution against treating uncertainty–market-quality relations as structurally stable.

Text-based measures of firm-level exposure to emerging risks are now standard, and are commonly validated by showing that they earn a return premium. We show that such evidence is insufficient. Two published measures of cybersecurity exposure — the 10-K disclosure-similarity score of Florackis et al. (2023, RFS) and the earnings-call keyword share of Jamilov et al. (2026, JF) — are nearly orthogonal, with a rank correlation of 0.13, yet both earn a quintile long–short six-factor alpha of approximately 5% per year. Only one of them behaves as a priced risk, and the difference derives from the source document rather than from the estimator. On days when cyber risk becomes salient, the two portfolios move in opposite directions, by 7.3 basis points a day (t = 2.51). The gap appears under all six definitions of a high-attention day, and reshuffling which days count as high-attention reproduces a gap this large in fewer than 2% of trials. Only the 10-K measure predicts which firms are subsequently attacked, with an odds ratio of 2.2 per standard deviation against 1.1, although the incremental predictive gain of either measure over firm size and industry is small. A keyword count of the same risk-factor text, constructed without reference to realized attacks, behaves like the 10-K measure rather than like the earnings-call measure that shares its method. The estimated incidence of the pricing across firm size reverses sign with the choice of instrument. A return premium therefore does not establish that a risk has been measured.

Using cross-platform measures of social-media activity, this study examines the relation between investor attention and corporate disclosure from 2012 to 2021. Firms are more likely to file an 8-K following increases in social-media attention, including filings classified conservatively as purely voluntary. Attention also predicts mandatory filings, indicating that the relation is not unique to discretionary disclosure. Pure voluntary filings released during high-attention periods are followed by lower short-horizon abnormal returns than those released during low-attention periods. Comparisons with matched high-attention non-filing days confirm a negative next-day return difference, although the corresponding five-day estimate is imprecise. The evidence is concentrated among small, low-priced firms with social-media coverage, which limits its broader applicability.

Investor disagreement is unobservable and its proxies correlate weakly, so a recent literature builds model-implied composites whose internal composition is set by theory rather than by fit. That composition has never been examined, because the latent variable such a measure estimates was itself unobservable. This paper shows that the DIS measure of Goulding et al. (2026, JFE) is algebraically invertible: released under five risk-aversion settings, its two structural blocks and the latent belief-precision term solve exactly, and three restrictions the mapping imposes hold against I/B/E/S, Compustat and CRSP data the recovery never uses, at correlations above 0.91. The composition is lopsided. One block carries 98.6% of the variance, and the analyst inputs - the only ones derived from what an identifiable population says about a specific quantity - supply 41% of belief precision in the lowest-disagreement decile but 13% in the highest, leaving the measure nearly inactive as a measure of beliefs precisely where it earns its return predictability. Those returns are themselves fragile. The 21.5% annualized alpha falls to 4 to $7% against realized volatility, and that residual is confined to 1994-2008: over 2009-2022 idiosyncratic volatility spans the composite rather than the reverse, while every compositional finding replicates in both halves.