Publications

How do government subsidies shape the information environment of public firms? Linking public subsidy data to analyst data and SEC Form 4 insider transactions for U.S. public firms over 2000 to 2023, we show that subsidies operate as attention events rather than information events. Analyst coverage intensity rises by approximately 4.8 percent in a panel OLS, and by 8.7 percent in a matched difference-in-differences design over the five-year award window following the first receipt. The effect builds for three years post-award and scales with award size. Yet the content of analyst output does not change. Consensus recommendations show a contemporaneous improvement, but no causal effect in the matched design, and forecast disagreement is flat once dispersion is scaled by the consensus mean. The unscaled increase using both I/B/E/S and Zacks data is a price-level artifact. Insiders respond decisively, with open-market trading volume rising 47 percent after subsidies, tilting toward net selling, and concentrating in opaque firms where outside evaluation is hardest. These post-award sales are less predictive of future returns than ordinary insider sales, consistent with opportunistic selling around subsidy-driven attention rather than informed trading. Government awards shift attention and liquidity rather than information, and corporate insiders, rather than analysts, appear to convert that attention into private benefits.

Large unrealized losses on long-duration securities held at amortised cost are a well-documented source of bank fragility. Much less is known about when those hidden losses show up in market-based measures of bank risk. We argue that hidden duration exposure raises crash risk only when a salient shock forces investors to revalue long-duration assets, and we test this with the September 2022 UK gilt crisis, a repricing shock that originated in a fiscal announcement and the pension sector rather than in banks. Across 42 banks in 13 jurisdictions, a one-standard-deviation increase in pre-shock gilt sensitivity, a market-revealed measure of duration exposure, raises the probability of a 5-day forward equity crash by 4.06 percentage points (p = 0.005); a placebo window one year earlier shows nothing, and the estimate is significant in every leave-one-country-out subsample. A machine-learning diagnostic applied to a daily panel of 106 U.S. banks (2010 to 2024) independently points to held-to-maturity exposure relative to equity as the relevant balance-sheet vulnerability after 2020, and U.S. stablecoin-stress episodes show the same pattern with the predicted sign, though with limited statistical power. A simple 22-day volatility signal matches gradient boosting out of sample, so the machine-learning contribution is diagnostic rather than predictive. The results identify the trigger that turns hidden duration losses into priced crash risk.

Reinforcement learning (RL) is increasingly proposed for financial decision making, yet evidence that it outperforms established methods remains fragmented and largely self-reported. We test the claim with controlled, out-of-sample experiments on real market data in two domains under one fixed, cost-aware protocol, so any difference is attributable to the decision rule alone. In equity portfolio optimization (Dow Jones constituents, daily 2006 to 2024), four deep RL algorithms compete against mean-variance, risk-parity, and naive 1/N diversification. Across the 2019 to 2024 hold-out, three universes, transaction-cost levels, market regimes, and walk-forward re-estimation, no RL algorithm robustly outperforms 1/N or a well-implemented mean-variance optimizer. In directional trading on real intraday crypto and equity data, hyperparameter-tuned RL fails to beat simple rule-based strategies and buy-and-hold, and an ablation shows that neither algorithm choice nor feature engineering makes RL competitive. Asked whether RL can augment classical methods, an RL meta-allocator improves substantially on pure RL and matches naive diversification, yet no RL design beats a well-implemented mean-variance optimizer. A controlled simulation explains the pattern: the sample a flexible learner needs to overtake a simple optimizer far exceeds available financial history. Even in a fair-shot test that feeds RL a survivorship-free CRSP cross-section with Compustat and IBES signals, no agent beats a simple characteristic-sorted benchmark. Across both domains, deep RL does not robustly outperform simple, well-implemented traditional methods, extending the classic 1/N diversification puzzle to the deep-RL era and cautioning against equating algorithmic complexity with financial performance.

A firm's reputational record and analysts' opinions about the firm are associated with its debt financing on two different clocks. In a panel of U.S. firms that links RepRisk incident data to Compustat, I/B/E/S and SEC Form 4 filings, net debt issuance is lower by 1.53 percentage points of lagged assets per standard deviation of the record's three-year accumulated state and by only 0.50 points per standard deviation of its current-year value, and in a distributed lag the association sits in the lags rather than in the current year. The opinion behaves the opposite way: issuance is lower by 1.00 points per standard deviation of its current value, and averaging it over three years shrinks the association to 0.55. The two signals enter with the same sign, and a regression on their difference imposes a restriction the data reject. Executives provide a benchmark. Chief executives and chief financial officers buy their own stock when analysts are pessimistic, while their trading shows no measurable association with the record: the chief executive's loading on the record is 0.010, with a 95 percent confidence interval from 0.038 below zero to 0.058 above and a minimum detectable effect of 0.069. Around a first severe incident, issuance does not move in the incident year and falls over the following three years, but the decline is no longer distinguishable from zero once the firms first treated between 2020 and 2022 are excluded and a placebo date reproduces it, so we report it as a descriptive time path. We document five designs that would identify an effect of the record and the number that closes each.