Abstract:

Given the barriers to national and global climate change regulations, what are the direct and spillover effects of subnational regulations? Using the staggered adoption of U.S. state-level greenhouse gas (GHG) regulations as a natural experiment, I find significant reductions in corporate GHG emissions in adopting states. Affected firms also increase GHG emissions disclosures and reduce the number of GHG-related 10-K risk disclosures. Furthermore, I document large spillover reductions in GHG emissions for treated firms’ affiliated operations in non-adopting states. These findings suggest possible alternate pathways to reduce GHG emissions if national or supranational policy coordination is not possible.

Abstract:

This paper examines the predictive ability of ESG risk disclosures.  Using a textual analysis-based approach, we derive measures assessing firms’ material ESG risk disclosures with respect to two benchmarks: SASB recommendations, discerned using the Materiality Map applied to firms’ 10-K risk factors section; and industry best practice, discerned using the ESG risks disclosed by the industry’s largest firms.  Descriptively, we document (i) considerable variation in the consistency of firms’ ESG disclosures with SASB recommendations; (ii) that firms disclose as material a large number of ESG risk factors deemed immaterial per SASB recommendations; and (iii) that firms’ disclosures of material ESG risks are more consistent with industry best practice.  We find that two left-tail ESG outcomes—subsequent year ESG penalties and incidents—are predicted by firms’ material ESG risk disclosures that are consistent with industry best practice, but not those deemed as material per the SASB framework.  Overall, the results suggest that for left-tail ESG events the strongest predictive ability lies within ESG topics outside of those denoted as material per SASB, and within those denoted as material per industry leader firms.

Abstract:

Policymakers have expressed concerns about the impact of widespread fast-delivery promises in e-commerce, leading to inefficient shipping practices such as partially filled shipments. This may result in additional trips and higher transportation- and packaging-related emissions. In this paper, first, we estimate the causal effect of fast delivery promises on consumer outcomes and environmental emissions. Large-scale experiment conducted at online retailer Wayfair revealed unintended consequences when fast-delivery promises were removed. Specifically, we show that consumers were unable to easily access fast-delivered products located closer to them, resulting in a shift towards purchasing products located further away. This led to a 1% increase in environmental emissions, 2.4% longer delivery times, and 2.8% higher shipping costs. Next, considering the ineffectiveness of removing fast-delivery promises, we propose and evaluate a new product ranking policy to reduce emissions. Our proposed multi-armed bandit algorithm balances two goals: (i) it minimizes the shipping distance to each consumer by placing products located geographically closer to a consumer higher up in the ranking page results, and (ii) it maximizes the conversion probability. We show that the algorithm reduces emissions by 7.84% without compromising conversion or profit metrics.

Abstract:

We explore Large Language Models’ (LLMs) reliability for financial analysis.  Using three popular LLM-based tools (GPT-3.5, GPT-4.0, Bard), we assess the responsiveness and accuracy of these tools’ responses to queries for basic financial information.  Specifically, we submit prompts requesting basic financial statement elements (total assets and net income) for a random sample of 500 US firms from 2016–2020.  We document high response rates, varying from 80% to 100% usable numerical answers to queries across the three LLMs.  However, we also find response accuracy to be extremely low: less than 1% (5%) match Compustat total asset (net income) benchmarks, and errors are sizeable (exceeding 35% at the mean, and 22% at the median).  We observe that GPT-4.0 is the most accurate—and least responsive—model, suggesting potential trade-offs in model performance.  Of note, we find all three LLMs are more responsive and accurate for larger and profitable firms.  We conclude that LLMs exhibit low reliability for factual financial queries, that performance trade-offs exist between LLMs, and that performance varies predictably with firm characteristics.  Our results highlight the importance of training data, model parameters, and context when using LLMs, and indicate users should exercise caution when applying generalized LLMs for financial analysis purposes.