Research

How does AI manage household stock portfolios? We collect a prospective, daily time-series of stock recommendations using several LLM’s and study AI’s investment style. AI recommends undiversified portfolios that positively load on momentum, large companies, and low book-to-market firms. AI primarily recommends stocks based on how much media attention firms receive. Using multiple versions of queries and requests, we find that buy-and-hold and actively managed AI portfolios do not appear to earn statistically-significant abnormal returns based on the methodology of Daniel et al. (1997). 

This paper examines how unauthorized immigration affects the fiscal health of local governments. Using countries of origin and arrival dates, we isolate immigration driven by social, economic, and political conditions. We predict immigration using a shift-share instrument based on pre-existing population distributions. In areas with structurally tight labor markets,unauthorized immigration explains lower municipal bond yields. Areas with typical labor market conditions experience higher yields, as do areas with “sanctuary”status. These effects accompany increased unemployment rates and expenditures on public amenities, including welfare assistance, construction, education, and law enforcement. These expenditures are not offset by higher tax revenues. 

We use minute-by-minute Bloomberg online status data to characterize two important dimensions of sell-side equity analysts’ work habits: we estimate the average workday length (AWL) to proxy for analysts’ general effort provision, and we use the percentage of away days (PAD) to proxy for their private information collection. Both AWL and PAD improve forecast precision, a causal result that we confirm using the COVID lockdown as an instrument. Frequent traveling not directly related to firm events is positively correlated with the likelihood of becoming an All-Star analyst, suggesting that institutional investors also value private information transmitted through in-person interactions.

We examine whether wildfire smoke affects sell-side financial analysts’ earnings forecast pessimism. Using plausibly exogenous county-level variation in satellite-based smoke plume coverage matched to analysts’ office locations, we find that each additional smoke day around an earnings announcement lowers an analyst’s forecast relative to peer consensus by 0.083 percentage points—about 29% of the interquartile range of relative forecast optimism. Cross-sectional tests point to a salience mechanism: smoke-exposed analysts ask more climate-related questions on conference calls, and the effect is stronger for firms with greater climate-risk disclosure and recent environmental violations. Analyst workload does not moderate the effect, and controlling for air quality does not attenuate it, inconsistent with a physiological-impairment channel. The smoke effect diminishes with prior exposure, consistent with adaptation, and is robust to alternative smoke windows and outcome measures.