Corporate Climate Risk: Measurements and Responses

Abstract 摘要

This paper conducts a textual analysis of earnings call transcripts to quantify climate risk exposure at the firm level. We construct dictionaries that measure physical and transition climate risks separately and identify firms that proactively respond to climate risks. Our validation analysis shows that our measures capture firm-level variations in respective climate risk exposure. Firms facing high transition risk, especially those that do not proactively respond, have been valued at a discount in recent years as aggregate investor attention to climate-related issues has been increasing. We document differences in how firms respond through investment, green innovation, and employment when facing high climate risk exposure.
本文对盈利电话记录进行了文本分析，以量化公司层面的气候风险。我们构建了分别衡量自然气候风险和过渡气候风险的词典，并识别出积极应对气候风险的公司。我们的验证分析表明，我们的测量方法捕捉到了公司层面各自气候风险暴露的差异。近年来，随着投资者对气候相关问题的总体关注度不断提高，面临高过渡风险的公司，尤其是那些不积极应对的公司，其估值出现了折价。我们记录了企业在面临高气候风险时如何通过投资、绿色创新和就业做出应对的差异。

Climate change poses severe challenges to businesses and society at large. Scientists predict that climate change will lead to increased incidence and severity of both chronic and acute climate and weather events, leading to unprecedented risks and disruptions that will affect corporations, the financial system, and the aggregate economy (Litterman et al. 2020). Following the pioneering work of Nordhaus (1977), many economists have studied interactions between climate change and the economy (e.g., Golosov et al. 2014; Nordhaus 2019); however, climate finance topics, such as how to assess, mitigate, and hedge climate risk across firms and asset classes, have received limited attention until recently. A major challenge to advancing this research agenda is the lack of credible measures of climate risk exposure across asset classes, in particular measures of equity assets (Hong, Li, and Xu 2019; Engle et al. 2020; Giglio, Kelly, and Stroebel 2021).
气候变化给企业和整个社会带来了严峻的挑战。科学家预测，气候变化将导致慢性和急性气候和天气事件的发生率和严重程度增加，从而导致前所未有的风险和破坏，影响企业、金融系统和整体经济（Litterman et al. 2020）。继Nordhaus（1977）的开创性工作之后，许多经济学家研究了气候变化与经济之间的相互作用（例如，Golosov et al.2014；Nordhaus 2019）；然而，气候融资课题，如如何评估、缓解和对冲不同企业和资产类别的气候风险，直到最近才得到有限的关注。推进这一研究议程的一个主要挑战是缺乏跨资产类别气候风险暴露的可靠衡量标准，尤其是股票资产的衡量标准（Hong、Li 和 Xu 2019；Engle et al.2020；Giglio、Kelly 和 Stroebel 2021）。

Several factors contribute to the above-mentioned lack of measures of firm-level climate risk exposure. First, in spite of stricter mandates imposed by regulators and investor demand, firms remain reluctant to disclose their climate risk exposure. For example, the most-common carbon emissions data have been available for only a limited number of traditional sectors (e.g., manufacturing and utilities), and firms often omit the indirect costs of carbon in supply chains (Shapiro 2021). Second, climate change is ever evolving, and it remains unclear how the climate will eventually change and affect firms, thus introducing significant uncertainty in government and corporate decision-making (Barnett, Brock, and Hansen 2020). Third, while historical emissions data are needed to assess a firm’s past business models, data capturing forward-looking views will be more useful in evaluating the firm’s climate exposure and adaptability in the transition toward an environmentally sustainable economy, an important goal for climate finance research (Giglio, Kelly, and Stroebel 2021).
造成上述缺乏公司层面气候风险暴露措施的原因有几个。首先，尽管监管机构和投资者提出了更严格的要求，但企业仍不愿披露其气候风险敞口。例如，最常见的碳排放数据仅适用于少数传统行业（如制造业和公用事业），而且企业往往忽略供应链中的间接碳成本（Shapiro 2021）。其次，气候变化是不断变化的，目前仍不清楚气候最终将如何变化并影响企业，因此给政府和企业决策带来了很大的不确定性（Barnett、Brock 和 Hansen 2020）。第三，虽然需要历史排放数据来评估企业过去的商业模式，但获取前瞻性观点的数据将更有助于评估企业在向环境可持续经济转型过程中的气候风险和适应性，这也是气候融资研究的一个重要目标（Giglio、Kelly 和 Stroebel 2021）。

In this paper, we fill this gap by quantifying, for the first time, climate risk exposure at the individual firm level, using earnings call transcript data for U.S. public companies. We use textual information from earnings calls in our analysis for several reasons. First, the vast majority of U.S. public firms hold regular earnings conference calls with their analysts and investors to discuss performance and factors related to performance, and, a point that is critical to this study, earnings calls contain detailed discussions with valuable and insightful information about the climate risks a firm faces beyond those that stem from public sources.¹ Second, unlike other firms’ disclosures, such as regulatory filings that are highly scripted and may lack informativeness and timeliness (e.g., Brown and Tucker 2011), the content contained in quarterly earnings transcripts is timelier and could vary significantly from quarter to quarter, allowing us to measure climate risk more accurately in real time. Third, discussions in earnings calls are inherently weighted by importance as an earnings conference call is a relatively short meeting where various parties can discuss only what they view as material factors—a feature that is key to measuring the importance of climate risks to firms. Finally, earnings calls also include discussions on how firms respond to climate risks, which enables us to capture firms’ proactiveness in addressing climate issues—a unique and important innovation in our study.
在本文中，我们利用美国上市公司的盈利电话记录数据，首次量化了单个公司层面的气候风险暴露，从而填补了这一空白。我们在分析中使用收益电话的文本信息有几个原因。首先，绝大多数美国上市公司都会定期与分析师和投资者召开财报电话会议，讨论公司业绩和与业绩相关的因素，而这一点对本研究至关重要，财报电话会议包含详细的讨论，其中有关于公司所面临的气候风险的有价值、有洞察力的信息，而这些信息超出了来自公开渠道的信息。¹ 其次，与其他公司披露的信息不同，如监管机构的文件，这些文件都是高度脚本化的，可能缺乏信息性和及时性（如：¹ ）、Brown 和 Tucker 2011）不同，季度收益记录中包含的内容更加及时，而且每个季度之间可能会有很大的变化，这使我们能够更准确地实时衡量气候风险。第三，由于财报电话会议是一个相对较短的会议，各方只能讨论他们认为重要的因素，因此财报电话会议的讨论本身就具有重要性加权--这一特点是衡量气候风险对公司重要性的关键。最后，财报电话会议还包括关于企业如何应对气候风险的讨论，这使我们能够捕捉到企业在解决气候问题方面的主动性--这是我们研究中独特而重要的创新。

We measure the climate risk faced by a given firm at a given time based on the share of earnings calls conversations that are centered on physical climate risk and transition risk, respectively. Our approach is similar to those used by prior studies (e.g., Baker, Bloom, and Davis 2016; Hassan et al. 2019, 2023, 2020). More importantly, we also measure whether or not the company’s attitude or response is proactive regarding the rise of climate risk by analyzing the verbs used in climate risk discussions. To do so, we overcome several challenges in applying standard textual analysis methods. The first is that any such analysis must account for multiple categories of climate risk (e.g., Giglio, Kelly, and Stroebel 2021; Stroebel and Wurgler 2021), which can be broadly classified as (a) physical climate risks, which are related to the physical impacts of acute climate events (e.g., hurricanes and wildfires) or chronic conditions (e.g., abnormal winter) and (b) transition risks. Given the multifaceted nature of climate risk, it is challenging to create a single measure that can capture all aspects of a firm’s climate risk exposure. Instead, we measure distinct climate risks separately using a dictionary-based approach.
我们根据分别以实际气候风险和过渡风险为中心的财报电话会议份额来衡量特定公司在特定时间面临的气候风险。我们的方法与之前的研究（例如，Baker, Bloom, and Davis 2016；Hassan et al.2019、2023、2020）。更重要的是，我们还通过分析气候风险讨论中使用的动词来衡量公司对气候风险上升的态度或反应是否积极主动。为此，我们克服了应用标准文本分析方法所面临的几个挑战。首先，任何此类分析都必须考虑气候风险的多个类别（例如，Giglio, Kelly, and Stroebel 2021；Stroebel and Wurgler 2021）、(b) 过渡风险。鉴于气候风险的多面性，创建一个单一的衡量标准来捕捉公司气候风险暴露的所有方面是具有挑战性的。因此，我们采用基于字典的方法分别衡量不同的气候风险。

The second challenge faced when measuring climate risk is that a well-constructed dictionary of climate-related keywords is not readily available in the literature, and a significant number of false positive and false negative cases arise if we apply a set of commonly known weather or climate keywords to a large set of transcripts. We adopt the dictionary approach over the machine learning (ML) method, with careful human supervision to minimize the occurrence of false positives and negatives. This approach allows researchers to make careful and deliberate judgment calls when classifying text based on complex concepts, such as climate risks, while preserving transparency and replicability.² Through careful selection over many iterations, we construct three comprehensive dictionaries consisting of over 1,600 climate keywords that are not directly related to either energy costs or general environmental risks.
在测量气候风险时面临的第二个挑战是，文献中并没有现成的气候相关关键词的完善词典，如果我们将一组众所周知的天气或气候关键词应用到大量的文字记录中，就会出现大量的假阳性和假阴性情况。与机器学习（ML）方法相比，我们采用了字典方法，并通过仔细的人工监督来尽量减少假阳性和假阴性的出现。² 通过多次迭代的精心选择，我们构建了三个全面的词典，其中包括 1600 多个气候关键词，这些关键词与能源成本或一般环境风险都没有直接关系。

To construct climate risk measures, we require the respective physical climate risk keywords to appear in the vicinity (±1 sentence) of at least one risk synonym to ensure that firms are indeed exposed to uncertainty related to climate-related events (as in Hassan et al. 2019).³ Transition risk differs in that it may not materialize in the short term and is thus measured based on discussions of keywords in our transition risk dictionary without having to appear near a risk synonym. Our approach produces three climate risk measures for each firm at quarterly frequency. In addition, using a list of verbs that capture firms’ proactive attitudes when discussing transition risk, we decompose our transition risk measure into proactive and nonproactive components.
为了构建气候风险度量，我们要求相应的物理气候风险关键词至少出现在一个风险同义词的附近（±1 句），以确保企业确实面临与气候相关事件有关的不确定性（如 Hassan et al.³ 过渡期风险的不同之处在于，它可能不会在短期内实现，因此根据我们的过渡期风险词典中的关键词讨论来衡量，而不必出现在风险同义词附近。我们的方法以季度为频率，为每家公司生成三个气候风险衡量指标。此外，我们还使用了一个动词列表来捕捉企业在讨论转型风险时的积极态度，从而将我们的转型风险度量分解为积极和非积极两个部分。

After establishing our measures, we conduct a battery of analyses to validate that they indeed capture a firm’s exposure to climate risks. First, we examine the list of most frequently discussed keywords in each of the measures and find that the patterns are consistent with intuitions. Second, we examine the time-series patterns as well as industry and firm-level variations in the climate risk measures. While relative industry rankings vary across different types of climate risks, they all exhibit significant variations that are consistent with industry-level exposure to climate risks. Third, in our validation analysis using various external benchmarks, we further demonstrate the validity of our climate risk measures. Our analysis shows that the presence of natural disasters in a local area is associated with a significant increase in both acute and chronic climate risk measures for firms headquartered in that area over the subsequent quarter.
在确定衡量标准后，我们进行了一系列分析，以验证这些标准确实反映了企业所面临的气候风险。首先，我们检查了每个衡量指标中最常讨论的关键词列表，发现其模式与直觉一致。其次，我们研究了气候风险度量的时间序列模式以及行业和公司层面的变化。虽然不同类型气候风险的相对行业排名各不相同，但它们都表现出显著的差异，这与行业层面的气候风险暴露是一致的。第三，在利用各种外部基准进行的验证分析中，我们进一步证明了气候风险度量的有效性。我们的分析表明，对于总部位于当地的企业来说，当地发生自然灾害与该地区在随后一个季度的急性和慢性气候风险指标的显著增加有关。

Validating the transition risk measure, we examine its correlations with two sets of existing external benchmarks: (1) firm-level MSCI Climate Change Index (CCI) and (2) industry-level carbon dioxide (CO₂) intensity constructed by Shapiro (2021) and firm-level CO₂ intensity based on the U.S. Environmental Protection Agency’s (EPA) emissions data. First, we find that our transition risk measure is positively and significantly correlated with MSCI CCI. Second, we find a strong and positive correlation between the average transition risk and CO₂ intensity as measured by Shapiro (2021) at the NAICS six-digit level for the manufacturing sector. Finally, analyzing firm-level emissions data, we find that our transition risk measure—albeit only its nonproactive component—is positively correlated with a firm’s CO₂ intensity in subsequent years. This relationship is significant in only one direction, suggesting that firms that face higher transition risk but proactively respond to such risks are indeed more active and effective in reducing their carbon footprints.
为了验证过渡风险度量，我们检验了它与两组现有外部基准的相关性：(1) 公司层面的 MSCI 气候变化指数 (CCI)；(2) 由 Shapiro (2021) 构建的行业层面的二氧化碳 (CO₂) 强度和基于美国环境保护署 (EPA) 排放数据的公司层面的 CO₂ 强度。美国环境保护署 (EPA) 的排放数据。首先，我们发现我们的转型风险衡量标准与 MSCI CCI 呈显著正相关。其次，我们发现平均转型风险与 NAICS 六位数级别的制造业二氧化碳₂ 强度Shapiro（2021）之间存在很强的正相关性。最后，通过分析企业层面的排放数据，我们发现我们的转型风险度量--尽管只是其非主动部分--与企业随后几年的二氧化碳₂ 排放强度正相关。这种关系只在一个方向上显著，表明面临较高转型风险但积极应对这种风险的企业在减少碳足迹方面确实更积极、更有效。

While maintaining high correlation when overlapping, our newly developed measures provide improved coverage and quantification of firm-level exposure to climate risk compared to existing measures. Compared with ESG ratings, our measures are available at the quarterly level for 4,719 public firms over a long period of time, and are less prone to the selection bias that occurs commonly with ESG data. Unlike the EPA’s plant-level CO₂ emissions data, which are limited only to firms that operate in the manufacturing, mining, and trade sectors, our measures cover all sectors where earnings call data are available, thus offering a comprehensive assessment of climate risk exposure across the economy. Of all public firms with earnings call data available, about 61.8% (2,918 firms) show at least one positive value in the transition risk measure, which corresponds to 34.7% of the firm-years that have positive values in transition risk. Even when considering the years when MSCI CCI data become available, our measure, on average, provides coverage of transition risk to an additional 952 firms with nonmissing values and 480 firms with positive values. Furthermore, we show in a variance decomposition analysis that the majority of variations in our three climate risk measures occur at the firm level, capturing not only cross-firm but also within-firm variations in climate risk exposure.
与现有指标相比，我们新开发的指标在保持高度相关性的同时，还能更好地覆盖和量化公司层面的气候风险。与环境、社会和公司治理评级相比，我们的衡量标准可长期提供 4,719 家上市公司的季度数据，不易出现环境、社会和公司治理数据常见的选择偏差。美国环保署的工厂级二氧化碳₂ 排放数据仅限于制造业、采矿业和贸易行业的企业，而我们的衡量指标则涵盖了所有有盈利电话数据的行业，因此可以全面评估整个经济中的气候风险暴露。在所有可获得电话收益数据的上市公司中，约有 61.8%（2918 家公司）的过渡风险指标显示了至少一个正值，这相当于 34.7% 的公司年度过渡风险指标显示了正值。即使考虑到 MSCI CCI 数据可用的年份，我们的衡量方法平均也能覆盖额外 952 家未缺失值公司和 480 家正值公司的过渡风险。此外，我们的方差分解分析表明，我们的三个气候风险衡量指标的大部分变化都发生在公司层面，不仅捕捉到了气候风险暴露的跨公司变化，也捕捉到了公司内部的变化。

Having established the validity of our measures, we next study one of the most important issues in the climate finance literature—the extent to which climate risk, especially transition risk, is priced in capital markets (e.g., Bolton and Kacperczyk 2021a; Giglio, Kelly, and Stroebel 2021). We first relate the firm-level transition risk measure to a firm’s market valuation measured by Tobin’s q, and find that our transition risk measure is negatively correlated with a firm’s Tobin’s q, suggesting that the firm’s transition risk exposure is priced in equity markets. Second, we find that this relationship has only become significant since 2010, likely because of rising aggregate investor attention to climate risk (e.g., Choi, Gao, and Jiang 2020; Engle et al. 2020), as well as climate-related initiatives and regulations implemented around this time.⁴ Third, when analyzing the relative effects of the proactive and nonproactive components of the transition risk measure, we find that only the nonproactive component has a significantly negative relation with Tobin’s q, suggesting that equity markets appear to discount only firms that do not actively manage their transition risk, while not penalizing those that address risk proactively. Importantly, these findings remain robust even after controlling for firm fixed effects, providing additional support for the idea that changes in climate risk discussion correlate with changes in Tobin’s q.
在确定了我们测量方法的有效性之后，我们接下来研究气候融资文献中最重要的问题之一--气候风险，尤其是过渡风险在资本市场中的定价程度（例如，Bolton and Kacperczyk 2021a；Giglio, Kelly, and Stroebel 2021）。我们首先将公司层面的转型风险度量与以托宾 Q 衡量的公司市场估值联系起来，发现我们的转型风险度量与公司的托宾 Q 负相关，这表明公司的转型风险敞口已在股票市场上定价。其次，我们发现这种关系自 2010 年以来才变得显著，这可能是因为投资者对气候风险的总体关注度不断提高（例如，Choi、Gao 和 Jiang 2020；Engle et al.4第三，在分析转型风险度量中主动和非主动部分的相对影响时，我们发现只有非主动部分与托宾 Q 显著负相关，这表明股票市场似乎只对那些不主动管理转型风险的公司打折扣，而对那些主动应对风险的公司并不惩罚。重要的是，即使在控制了公司固定效应后，这些发现仍然是稳健的，这为气候风险讨论的变化与托宾 Q 变化相关的观点提供了更多支持。

Further analysis shows that our measures capture unique information that is useful in studying the pricing effects of climate risk based on horse-race regressions with various alternative measures. In particular, we consider (1) a transition climate risk measure constructed with the same dictionary but using textual information from firms’ 10-K/10-Q filings, (2) a transition risk measure constructed based on climate-related company news from Dow Jones Newswires, (3) MSCI CCI or ESG ratings, and (4) measures constructed by Sautner et al. (2023) using different climate dictionaries and methods. In all of these tests, the coefficients for our transition risk measure and its nonproactive component remain negative and significant at the 1% level, confirming the unique value added by both the earnings calls data and our construction method. In summary, our transition risk measure generates new and valuable information that is not already available in other public sources and also provides comprehensive coverage over a large sample of public firms from 2002 onward.
进一步的分析表明，我们的衡量标准捕捉到了独特的信息，这些信息对于研究气候风险的定价效应非常有用，而这些定价效应是基于与各种替代衡量标准的赛马回归得出的。具体而言，我们考虑了：（1）使用相同词典但使用公司 10-K/10-Q 文件中的文本信息构建的过渡气候风险度量；（2）根据道琼斯通讯社中与气候相关的公司新闻构建的过渡风险度量；（3）MSCI CCI 或 ESG 评级；以及（4）Sautner et al.(2023) 使用不同的气候词典和方法构建的指标。在所有这些测试中，我们的过渡风险度量及其非积极成分的系数仍然为负，并且在 1%的水平上显著，这证实了盈利调用数据和我们的构建方法所带来的独特价值。总之，我们的过渡风险度量方法提供了其他公开来源所没有的新的有价值的信息，并且全面覆盖了 2002 年以来的大量上市公司样本。

In the last set of analysis, we explore how firms respond, in terms of investments, innovation, and employment, to transition risk exposure. Our results show that firms’ attitudes toward climate issues—their proactiveness—matter significantly in how they respond to climate risk along these dimensions. First, we find that, while there is no significant relation between transition risk and investment as measured by total capital expenditures (CapEx) in nonproactive firms, firms that proactively respond to climate risk tend to increase their investment subsequently. Second, we find a negative relation between transition risk and subsequent R&D expenditures, a finding that is driven entirely by nonproactive firms. In contrast, proactive firms innovate more actively by producing more green patents in subsequent years. Given this relationship, we conduct further analysis to explore the attributes of proactive firms and their potential differential impact on firm valuation. We find some evidence that the equity markets tend to value proactive responses to transition risk from green patenting firms more than nongreen proactive responses. Finally, our employment analysis shows that firms that do not proactively respond reduce employment following a rise in transition risk, while the firms that proactively respond to transition risk do not reduce employment subsequently. Taken together, our measures are useful not only for understanding the pricing of transition risk in capital markets, but also for predicting real outcomes as firms proactively respond to changes in climate risk.
在最后一组分析中，我们探讨了企业如何在投资、创新和就业方面应对转型风险。我们的结果表明，企业对气候问题的态度--企业的主动性--对企业如何在这些方面应对气候风险具有重要影响。首先，我们发现，虽然在非积极主动的企业中，过渡风险与以资本支出总额（CapEx）衡量的投资之间没有显著关系，但积极主动应对气候风险的企业往往会随之增加投资。其次，我们发现转型风险与后续研发支出之间存在负相关关系，这一发现完全是由非主动型企业驱动的。与此相反，积极主动的企业在随后的几年中会产生更多的绿色专利，从而更加积极地进行创新。鉴于这种关系，我们进行了进一步分析，以探讨积极主动企业的属性及其对企业估值的潜在不同影响。我们发现一些证据表明，与非绿色的积极应对相比，股票市场更看重绿色专利企业对转型风险的积极应对。最后，我们的就业分析表明，不积极应对转型风险的企业会在转型风险上升后减少就业，而积极应对转型风险的企业则不会随之减少就业。综上所述，我们的措施不仅有助于理解资本市场对过渡风险的定价，而且有助于预测企业主动应对气候风险变化时的实际结果。

1 Related Literature 1 相关文献

Our paper contributes to the literature by constructing firm-level climate risk measures. Properly measuring climate risk exposure across assets is critical to any study of climate risk and its impact on the underlying assets. A growing body of literature studies the effects of climate change on real estate assets and housing markets using properties’ exposure to physical climate risk factors, such as projected sea-level rise (SLR), flooding, and hurricanes (e.g., Bernstein, Gustafson, and Lewis 2019; Baldauf, Garlappi, and Yannelis 2020; Goldsmith-Pinkham et al. 2023; Keys and Mulder 2020; Giglio et al. 2021).⁵ With regard to equity assets, however, the literature still lacks a set of measures with which to measure firms’ exposure to climate risks systematically, and researchers must use alternative measures, for instance, CO₂ emissions data or ESG ratings (e.g., Engle et al. 2020)⁶ despite concerns about their coverage and reliability (Stanny 2018; Berg, Koelbel, and Rigobon 2022). As a result, Giglio, Kelly, and Stroebel (2021) conclude in their survey that there is “substantial scope for improvements of the measures of climate risk exposure, in particular for equity assets.” Our paper represents valuable progress toward developing new ways to quantify firms’ climate risk exposure.
我们的论文通过构建公司层面的气候风险度量，为相关文献做出了贡献。对气候风险及其对相关资产的影响进行研究时，正确衡量不同资产所面临的气候风险至关重要。越来越多的文献利用房地产对物理气候风险因素（如预计的海平面上升（SLR）、洪水和飓风）的暴露程度来研究气候变化对房地产资产和住房市场的影响（如......）、Bernstein, Gustafson, and Lewis 2019；Baldauf, Garlappi, and Yannelis 2020；Goldsmith-Pinkham et al.2023; Keys and Mulder 2020; Giglio et al.2021 ）。⁵ 然而，在股权资产方面，文献仍然缺乏一套系统衡量公司气候风险暴露程度的指标，研究人员必须使用替代指标，例如二氧化碳₂ 排放数据或 ESG 评级（例如：^{5 ）、Engle et al. 2020）6 尽管对其覆盖面和可靠性存在担忧（Stanny 2018；Berg, Koelbel, and Rigobon 2022）。因此，Giglio、Kelly 和 Stroebel（2021 年）在他们的调查中得出结论："气候风险暴露的测量方法还有很大的改进空间，尤其是对股票资产而言。我们的论文在开发量化公司气候风险暴露的新方法方面取得了宝贵的进展。}

More broadly, our paper adds to the climate finance literature in several ways. First, our measures can be used to study how capital markets price climate risk. Several studies examine whether equity markets price risks related to long-run temperature shifts, drought, sea-level rise, or carbon emissions (e.g., Hong, Li, and Xu 2019; Bolton and Kacperczyk 2021a,b; Hsu, Li, and Tsou 2023; Ilhan, Sautner, and Vilkov 2021). Other evidence points to climate risks affecting fixed-income and real estate markets.⁷ Different from all these studies, we show, using our novel firm-level climate risk measures, that climate risk is priced in equity markets, especially following a rise in aggregate investor attention in recent years. We also document that firms’ proactiveness attenuates the discounting of high climate risk in equity markets. Second, our measures could help investors implement effective hedging strategies, which is of great importance considering that many effects of climate change will manifest far into the future and neither financial derivatives nor insurance markets is available to directly hedge those long-horizon risks. Engle et al. (2020) propose an approach to dynamically hedging climate risk using historical responses of individual stocks to their “Climate News Index.” Our firm-level climate risk measures, along with their proactive component, also can be used by investors to assess, construct, and hedge portfolio exposure to aggregate climate risk in accordance with their risk tolerance.
更广泛地说，我们的论文在几个方面为气候融资文献增添了新的内容。首先，我们的措施可用于研究资本市场如何为气候风险定价。一些研究考察了股票市场是否会对与长期气温变化、干旱、海平面上升或碳排放相关的风险进行定价（例如：Hong, Li, Xu, et al、Hong, Li, and Xu 2019；Bolton and Kacperczyk 2021a，b；Hsu, Li, and Tsou 2023；Ilhan, Sautner, and Vilkov 2021）。⁷ 与所有这些研究不同的是，我们利用新颖的企业级气候风险度量方法表明，气候风险在股票市场上是有定价的，尤其是在近年来投资者的总体关注度上升之后。我们还发现，企业的主动性会降低股票市场对高气候风险的贴现。其次，我们的措施可以帮助投资者实施有效的对冲策略，考虑到气候变化的许多影响将在未来很长时间内显现，而金融衍生品和保险市场都无法直接对冲这些长期风险，因此这一点非常重要。Engle et al.(2020) 提出了一种利用个股对其 "气候新闻指数 "的历史反应来动态对冲气候风险的方法。我们的公司级气候风险度量及其主动成分也可用于投资者根据其风险承受能力评估、构建和对冲投资组合中的总体气候风险。

Our study is closely related to a contemporaneous paper by Sautner et al. (2023). While both papers propose firm-level measures of climate exposure using earnings call data, there are major differences in both the methodology and the scope of the economic questions explored. Unlike Sautner et al. (2023), who use an ML algorithm, we construct climate-related dictionaries manually through careful human supervision and iterative testing. Like that of Loughran and McDonald (2011) and Baker, Bloom, and Davis (2016), our approach is more transparent and less sensitive to initial inputs and parameter choices than ML algorithms, providing us with what we consider as a necessary and effective tool given the complexity of climate issues. More importantly, the scope of the economic questions we explore in our study is quite different from theirs. While they focus primarily on economic factors that correlate with firms’ climate change exposure, we explore whether transition risk and, especially, firms’ proactiveness in addressing it, are priced in equity markets as well as how firms respond to transition risk. Our paper is unique as the first in the literature to measure firms’ proactiveness in addressing climate issues. One of our key contributions lies in documenting that proactive attitudes are priced in equity markets and that proactive firms respond, in terms of investment, green innovation, and employment, differently to rising transition risk.
我们的研究与 Sautner et al.(2023).虽然这两篇论文都提出了公司层面的气候风险度量方法，但在方法论和探讨的经济问题范围上都有很大不同。与 Sautner et al.(2023) 使用的是 ML 算法，而我们则是通过仔细的人工监督和反复测试手动构建气候相关词典。与Loughran和McDonald（2011）和Baker、Bloom和Davis（2016）的方法一样，我们的方法比ML算法更透明，对初始输入和参数选择的敏感度更低，为我们提供了我们认为在气候问题的复杂性下必要且有效的工具。更重要的是，我们在研究中探讨的经济问题的范围与他们大不相同。他们主要关注的是与企业气候变化风险相关的经济因素，而我们探讨的则是转型风险，尤其是企业应对转型风险的主动性，是否在股票市场上被定价，以及企业如何应对转型风险。我们的论文独树一帜，是文献中第一篇衡量企业在应对气候问题方面的主动性的论文。我们的主要贡献之一在于记录了股票市场对积极主动态度的定价，以及积极主动的企业在投资、绿色创新和就业方面对不断上升的转型风险做出的不同反应。

2 Data 2 数据

2.1 Earnings calls 2.1 盈利电话

To measure firm-level exposure to climate risk, we use as our primary data source transcripts of earnings calls involving all U.S. public firms obtained from Thomson Reuters’ StreetEvents database. These transcripts record discussions between a public company’s management team, industry analysts, investors, and the media regarding the company’s corporate strategy, operating conditions, and financial performance for a given quarter. The same data are used in several other papers, for example, Hassan et al. (2019), who study corporate exposure to political risk, and Li et al. (2021), who create novel measures of corporate culture. Firms typically hold one conference call in each fiscal quarter following their earnings releases. Thus, we conduct most of our analysis at the firm-quarter level. One important benefit, among others, of using the earnings calls data is that, because the data are available for almost all public firms, we can construct climate risk measures that place all public firms on a level playing field, as opposed to using ESG scores only or other measures that are available for only a small subset of firms that may be subject to selection bias.⁸
为了衡量公司层面的气候风险，我们使用从汤森路透的 StreetEvents 数据库中获取的所有美国上市公司的盈利电话会议记录作为主要数据来源。这些记录了上市公司管理团队、行业分析师、投资者和媒体就公司战略、运营状况和特定季度的财务业绩所进行的讨论。其他几篇论文也使用了相同的数据，例如，Hassan et al.(2019) 研究了企业面临的政治风险，Li et al.(2021)，他们创建了新的企业文化衡量标准。企业通常会在每个财政季度的财报发布后召开一次电话会议。因此，我们的大部分分析都是在公司季度层面进行的。使用财报电话会议数据的一个重要好处是，由于几乎所有上市公司都有这些数据，我们可以构建气候风险度量，使所有上市公司处于公平竞争的环境中，而不是只使用 ESG 分数或其他仅适用于一小部分公司的度量，这样可能会出现选择偏差。

We use all earnings call data from January 2002 through the first half of 2018 in our analysis, and extract the texts of entire conference calls from the raw XML transcript files using Python, which includes both presentations by management and subsequent Q&A sessions. We also extract firm identifiers (e.g., firm names, tickers, CUSIP numbers) and earnings call information (e.g., date and time) from the transcript files.
我们在分析中使用了 2002 年 1 月至 2018 年上半年的所有财报电话会议数据，并使用 Python 从原始 XML 转录文件中提取了整个电话会议的文本，其中包括管理层的发言和随后的问答环节。我们还从记录文件中提取了公司标识符（如公司名称、股票代码、CUSIP 编号）和财报电话会议信息（如日期和时间）。

2.2 Firm-level financial data
2.2 公司层面的财务数据

We obtain firms’ financial data from Compustat. We use Tobin’s q as the main measure of a firm’s market valuation to examine whether the stock market has priced the climate risks captured by our measures. To study a firm’s responses to climate risk, we consider CapEx, R&D, and employment as outcomes. Other firm-level attributes, such as total assets, property, plant, and equipment (PPE), and the book leverage ratio, are used as control variables. All the firm-level attributes are available at the quarterly level, except for employment data, which are available only annually. Information about firms’ stocks is obtained from the Center for Research in Security Prices (CRSP).
我们从 Compustat 中获取公司的财务数据。我们使用托宾 Q 作为衡量企业市场估值的主要指标，以考察股票市场是否对我们所衡量的气候风险进行了定价。为了研究企业对气候风险的反应，我们将资本支出、研发和就业作为结果。其他公司层面的属性，如总资产、不动产、厂房和设备（PPE）以及账面杠杆比率，都被用作控制变量。除了就业数据只能按年度获得外，所有企业层面的属性均可按季度获得。公司股票信息来自证券价格研究中心（CRSP）。

We match the earnings call data with other firm-level data using firm identifiers and apply several filters. First, because many financial firms, especially insurance companies, sell insurance products to others to hedge climate- or disaster-related risks, we exclude financial firms (North American Industry Classification System or NAICS 52) from our main analysis. Second, we exclude firms whose headquarters are located outside the continental United States. Our sample includes 4,719 unique firms and 139,959 firm–quarter observations. Table 1 presents summary statistics for Tobin’s q, CapEx, R&D expenditures, Property, Plant, and Equipment (PPE), book leverage, return on assets (ROA), employment, and total assets. CapEx, R&D expenditures, and PPE are all scaled by a firm’s total assets in the preceding quarter.⁹
我们利用公司标识符将盈利电话会议数据与其他公司层面的数据进行匹配，并采用了几种筛选方法。首先，由于许多金融公司（尤其是保险公司）向他人销售保险产品以规避气候或灾害相关风险，我们在主要分析中排除了金融公司（北美行业分类系统或 NAICS 52）。其次，我们将总部位于美国大陆以外的公司排除在外。我们的样本包括 4,719 家独特的公司和 139,959 个公司季度观察值。表 1 列出了托宾 Q、资本支出、研发支出、不动产、厂房和设备（PPE）、账面杠杆率、资产回报率（ROA）、就业率和总资产的汇总统计。资本支出、研发支出和财产、厂房和设备均按公司上一季度的总资产缩放。

2.3 Additional textual data
2.3 补充文本数据

We also use textual information from firms’ regulatory filings, in particular 10-K and 10-Q filings, as alternative data sources to construct our climate risk measures. We focus on the two most relevant sections in 10-K/10-Q filings: (1) management discussion and analysis (MD&A) and (2) Item 1A “Risk Factors.” MD&A section contains management discussions of firms’ performance, risks, and future plans. The risk factors (RF) section provides information about the risk factors a firm identifies that might influence the company or its equity return. MD&A section is available for our entire sample period, from 2002 through 2018, while RF section is available only from 2006 onward following the implementation of Regulation S-K Item 105.
我们还使用公司监管文件中的文本信息，特别是 10-K 和 10-Q 文件，作为构建气候风险度量的替代数据来源。我们重点关注 10-K/10-Q 文件中两个最相关的部分：(1) 管理层讨论与分析 (MD&A) 和 (2) 第 1A 项 "风险因素"。管理层讨论与分析部分包含管理层对公司业绩、风险和未来计划的讨论。风险因素 (RF) 部分提供了公司确定的可能影响公司或其股本回报的风险因素的相关信息。从 2002 年到 2018 年的整个样本期间都有 MD&A 章节，而 RF 章节只有在《S-K 法规》第 105 条实施后的 2006 年以后才有。

We use publicly available company news as another source of textual data that we can use to construct firms’ climate risk measures. We obtain such data from RavenPack, which provides a comprehensive sample of firm-specific news stories from Dow Jones Newswires.¹⁰ To identify news stories about specific firms, we use relevance scores from RavenPack; these scores range from 0 to 100, capturing how closely the underlying news is related to a particular company. We identify relevant news stories for a given firm by requiring the relevance score to be 75 or above, as recommended by RavenPack.¹¹ We also exclude repeated news using the event novelty score provided by RavenPack so that our data capture only fresh news about a company. Finally, we use the same transition risk dictionary to determine whether a specific news story about a given firm is related to transition risk.
我们将公开的公司新闻作为另一个文本数据来源，用来构建公司的气候风险度量。¹⁰ 为了识别有关特定公司的新闻报道，我们使用了 RavenPack 的相关性分数；这些分数从 0 到 100 不等，反映了相关新闻与特定公司的密切程度。我们根据 RavenPack 的建议，要求相关性得分达到或超过 75 分，从而确定特定公司的相关新闻报道。¹¹ 我们还使用 RavenPack 提供的事件新颖性得分排除重复新闻，从而使我们的数据只捕捉到有关公司的新鲜新闻。最后，我们使用相同的过渡风险词典来确定关于特定公司的特定新闻报道是否与过渡风险有关。

2.4 Other external firm data
2.4 其他外部公司数据

To analyze the firm-level response to climate risk through green innovation, we obtain patent data from the Global Corporate Patent data set.¹² We follow Cohen, Gurun, and Nguyen (2020) and Haščič and Migotto (2015) and classify green patents as those containing environment-related technologies, such as emissions abatement technologies, renewable energy, and energy storage. The patent data are available for U.S. firms from 2002 through 2017. We calculate the number of green patents produced by each firm in a given year and define two measures to capture the intensive and extensive margins of firms’ green innovation activities: (1) an indicator that equals one if a firm has been granted at least one green patent in a given year, and zero otherwise and (2) the ratio of green patents to the total number of patents granted to the firm in that year. The first measure is available for all public firms, while the second measure is available only for firms that had at least one patent granted in a given year.
为了分析企业通过绿色创新应对气候风险的情况，我们从全球企业专利数据集中获取了专利数据。¹² 我们仿效Cohen, Gurun, and Nguyen (2020)和Haščič and Migotto (2015)，将绿色专利归类为包含环境相关技术的专利，如减排技术、可再生能源和能源存储。专利数据来自 2002 年至 2017 年的美国公司。我们计算了每家公司在给定年份中产生的绿色专利数量，并定义了两个指标来反映公司绿色创新活动的密集边际和广泛边际：(1) 如果一家公司在给定年份中至少获得了一项绿色专利，则该指标等于 1，否则等于 0；(2) 绿色专利与该公司当年获得的专利总数之比。第一个指标适用于所有上市企业，而第二个指标仅适用于在某一年至少获得一项专利授权的企业。

We obtain several external data sets to validate the new climate risk measures. The first data set contains natural disaster data from the Spatial Hazard Events and Losses Database (SHELDUS) that has been used in the economics literature (e.g., Barrot and Sauvagnat 2016) to examine the effects of natural disasters. These data record the counties, beginning/end dates, event names, main causes of damage (e.g., flooding, hurricanes), and the estimated economic losses. We match these data with our sample using firms’ headquarters locations, and we use the natural disasters as an external benchmark for validating our physical risk measures.
我们获取了几个外部数据集来验证新的气候风险度量。第一个数据集包含来自空间灾害事件和损失数据库（SHELDUS）的自然灾害数据，经济学文献（如Barrot 和 Sauvagnat 2016）曾使用该数据库研究自然灾害的影响。这些数据记录了县、开始/结束日期、事件名称、造成损害的主要原因（如洪水、飓风）以及估计的经济损失。我们将这些数据与企业总部所在地的样本进行匹配，并将自然灾害作为外部基准来验证我们的有形风险度量。

Our second external benchmark comprises several external ESG index or ratings. These scores measure how well a company manages ESG risks and opportunities based on information published in news coverage and/or corporate disclosures, such as sustainability reports and corporate websites, surveys, and information provided by other stakeholders, such as regulatory agencies and industry associations (e.g., Berg, Koelbel, and Rigobon 2022; Christensen, Serafeim, and Sikochi 2021). We obtain ratings from three sources (MSCI, RepRisk, and Refinitiv), and these ratings include overall scores as well as three individual scores (environmental, social, and governance) at the monthly or annual level. We use the MSCI CCI—a climate change theme score that is directly comparable to our climate risk exposure measures—as the main external benchmark. We note that the environmental components of ESG ratings provided by rating agencies focus on environmental risk that is entangled with, but different from, climate risk. Nevertheless, we conduct supplemental validation exercises using the RepRisk or Refinitiv Environmental Scores.¹³
我们的第二个外部基准包括若干外部 ESG 指数或评级。这些评分根据新闻报道和/或企业披露的信息（如可持续发展报告和企业网站）、调查以及其他利益相关者（如监管机构和行业协会）提供的信息（如Berg、Koelbel 和 Rigobon 2022；Christensen、Serafeim 和 Sikochi 2021）来衡量企业对 ESG 风险和机遇的管理程度。我们从三个来源（MSCI、RepRisk 和 Refinitiv）获得评级，这些评级包括月度或年度级别的总体得分以及三个单项得分（环境、社会和治理）。我们使用 MSCI CCI 作为主要的外部基准--这是一个气候变化主题评分，与我们的气候风险暴露度量直接可比。我们注意到，评级机构提供的环境、社会和治理评级中的环境部分主要关注与气候风险相关但不同的环境风险。不过，我们使用 RepRisk 或 Refinitiv 环境评分进行了补充验证。

Our third external benchmark consists of CO₂ emissions data from the EPA’s Greenhouse Gas Reporting Program (GHGRP) as an additional benchmark for our transition risk measure. Since October 2009, the GHGRP program has mandated that sources that emit 25,000 metric tons or more of CO₂ greenhouse gases per year must report their emissions, and the data are made publicly available on an annual basis starting in 2010 at the plant level; and these data include plant identity, geographic location, parent company, industry (NAICS), and greenhouse gas emissions. Following Bartram, Hou, and Kim (2021), we obtain plant-level emissions data from the EPA and match them with firm-level data from Compustat based on the names of parent companies.
我们的第三个外部基准包括来自美国环保署温室气体报告计划（GHGRP）的二氧化碳₂ 排放数据，作为我们衡量过渡风险的额外基准。自 2009 年 10 月起，GHGRP 计划规定，每年排放 25,000 公吨或更多 CO₂ 温室气体的排放源必须报告其排放量，并且从 2010 年开始每年公开工厂一级的数据；这些数据包括工厂身份、地理位置、母公司、行业（NAICS）和温室气体排放量。根据Bartram、Hou 和 Kim（2021），我们从美国环保署获得了工厂级排放数据，并根据母公司名称与 Compustat 的公司级数据进行了匹配。

3 Measuring Climate Risk at the Firm Level
3 在企业层面衡量气候风险

3.1 Constructing climate dictionaries
3.1 构建气候词典

We follow the recent literature that exploits textual information in earnings call data to identify risks (e.g., Hassan et al. 2019, 2023, 2020) to construct our firm-level climate risk measures. We must overcome several challenges in applying the textual analysis method to the construction of climate risk measures.
我们借鉴了最近的一些文献，这些文献利用盈利电话数据中的文本信息来识别风险（例如，Hassan et al.2019、2023、2020）来构建公司层面的气候风险度量。在应用文本分析方法构建气候风险度量时，我们必须克服几个挑战。

First, as pointed out by Giglio, Kelly, and Stroebel (2021), when studying climate risk and its impact on underlying assets, it is important to note the several categories of climate risks and that these distinct risks often do not materialize at the same time. Broadly speaking, climate-related risks can be classified into two major categories: (1) physical risks, which are related to the physical impacts of climate events, and are either acute (e.g., droughts, floods, extreme precipitation and wildfires) or chronic (e.g., rising temperatures and an accelerating loss of biodiversity), and (2) transition risks, which are caused by not responding to climate change and improving how businesses operate as society moves toward adopting sustainable practices (ie, low-carbon manufacturing). Transition risks are primarily influenced by policies and regulations and by societal expectations and market pressure. Given the multifaceted nature of climate risk, it is challenging to create a single measure that captures all aspects of a firm’s climate risk exposure. Instead, using a dictionary-based approach, we measure three climate-related risks separately: (1) acute physical risk, (2) chronic physical risk, and (3) transition risk. Given the complexity and multifaceted nature of climate issues and the importance of generating replicable results, we believe, for several reasons, that the dictionary approach is a better choice in this context than ML methods. First, ML methods are not as transparent as the dictionary approach because many ML algorithms function as black-box models. Second, ML methods are sensitive to initial inputs and parameter choices. Third, the accuracy of ML predictions depends heavily on constructing a large, representative training data set that is not readily available in the context of complex and multifaceted climate issues.
首先，正如Giglio、Kelly 和 Stroebel（2021）所指出的，在研究气候风险及其对相关资产的影响时，必须注意气候风险的几个类别，而且这些不同的风险往往不会同时出现。从广义上讲，与气候相关的风险可分为两大类：(1) 物理风险，与气候事件的物理影响有关，要么是急性的（如干旱、洪水、极端降水和野火），要么是慢性的（如气温升高和生物多样性加速丧失）；(2) 过渡风险，是由于没有应对气候变化，没有在社会采用可持续做法（即低碳制造）的过程中改进企业运营方式而造成的。过渡风险主要受政策法规、社会期望和市场压力的影响。鉴于气候风险的多面性，创建一个单一的衡量标准来捕捉企业气候风险暴露的所有方面是具有挑战性的。相反，我们采用基于字典的方法，分别衡量三种与气候相关的风险：(1) 急性物理风险，(2) 慢性物理风险，(3) 过渡风险。鉴于气候问题的复杂性和多面性，以及产生可复制结果的重要性，我们认为，出于几个原因，在这种情况下，字典方法比 ML 方法是更好的选择。首先，ML 方法不如字典方法透明，因为许多 ML 算法都是黑箱模型。其次，ML 方法对初始输入和参数选择很敏感。 第三，人工智能预测的准确性在很大程度上取决于能否构建一个大型的、有代表性的训练数据集，而在复杂和多层面的气候问题中，这个数据集并不容易获得。

Second, unlike using preexisting training libraries (as in, e.g., political or accounting textbooks), developing climate-related keywords requires considerable human effort. We detect two important issues once we apply a set of commonly known weather or climate keywords to a large set of transcripts. First, a significant number of false positive cases will arise in which keywords are used to describe issues that are entirely unrelated to the climate (e.g., “business climate,” “public cloud,” “economic storm”). A second issue is that weather and climate irregularities are commonly expressed using combinations of contrasting keywords (e.g., “warm winter,” “unseasonably cold,” “cool summer”). If we rely on a dictionary that consists entirely of unigrams, it is unlikely that we can include unigrams, such as “winter” or “warm,” thus generating many false negatives. We address these issues by manually constructing a hybrid dictionary consisting of both unigrams and bigrams (adjacent two-word combinations) to reduce both false positives and false negatives.
其次，与使用已有的训练库（如政治或会计教科书）不同，开发与气候相关的关键词需要大量的人力。一旦我们将一组众所周知的天气或气候关键词应用到大量的记录誊本中，我们就会发现两个重要问题。首先，会出现大量的假阳性情况，即关键词被用于描述与气候完全无关的问题（如 "商业气候"、"公共云"、"经济风暴"）。第二个问题是，天气和气候的不规则性通常使用对比性关键词的组合来表达（如 "温暖的冬天"、"反常的寒冷"、"凉爽的夏天"）。如果我们依赖完全由单字组成的字典，就不太可能包含单字，如 "冬天 "或 "温暖"，从而产生许多错误的否定。为了解决这些问题，我们手动构建了一个由单字词和双字词（相邻的两个单词组合）组成的混合词典，以减少误报和误判。

Specifically, our method builds on the premise that no algorithm understands the context of a human conversation better than human beings do.¹⁴ We start our dictionaries with a list of unigrams that we extract from the following sources: (a) disaster “incident-type” indications in the Disaster Declarations Summary of Federal Emergency Management Agency (FEMA), (b) Wikipedia’s list of severe weather phenomena,¹⁵ and (c) additional seed words that we added manually, namely, “temperature,” “cold,” “unseasonable,” and so on. We use this list to obtain all bigrams that contain at least one of the unigrams from the entire sample of earnings call transcripts. We then manually screen, for each unigram, the top-500 associated bigrams. If the top-500 associated bigrams are unambiguously used in the context of climate-related conversations, we then include the corresponding unigrams in the unigram dictionary. If not, we include the top-500 associated bigrams in the bigram library pending further screening. To reduce the incidence of false negatives, we supplement the bigram library with climate-related bigrams extracted from additional sources: (a) white papers and reports on climate issues mentioned by Engle et al. (2020), (b) news articles posted by The Weather Channel, and (c) an undergraduate textbook on meteorology (Ahrens 2008). Lastly, we screen the library through many iterations to eliminate false positives and include false negatives.
14我们从以下来源提取的单字词表开始我们的词典：(a) 联邦紧急事务管理局（FEMA）灾害申报摘要中的灾害 "事件类型 "指标，(b) 维基百科的恶劣天气现象列表，¹⁵ 以及 (c) 我们手动添加的其他种子词，即 "温度"、"寒冷"、"反常 "等。我们使用该列表从整个盈利电话记录样本中获取至少包含一个单字的所有大词。然后，我们对每个单字词组手动筛选出前 500 个相关的双字词组。如果排名前 500 位的关联大词被明确用于气候相关对话的语境中，我们就会将相应的单词纳入单词词典。如果不是，我们就将前 500 个相关的词条纳入词条库，等待进一步筛选。为了减少假阴性的发生，我们用从其他来源提取的与气候相关的大词汇来补充大词汇库：(a) Engle et al.(2020), (b) The Weather Channel 发布的新闻文章，以及 (c) 一本气象学本科教科书 (Ahrens 2008)。最后，我们通过多次迭代对库进行筛选，以消除误报，并纳入误报。

We distinguish between climate risk and other risks in building our dictionaries. First, companies may discuss their climate topics that are related to changes in energy prices, but the latter not exclusively related to climate risk. To ensure that our climate risk measures are not driven by energy prices, our climate dictionaries do not contain any keywords related to energy prices or costs.¹⁶ Instead, we construct a firm-specific, time-varying energy-price exposure index and include it as a control variable in our main analysis. Furthermore, companies’ environmental responsibility and greenhouse gas emissions efforts are likely correlated, but not equivalent. We thus remove any keywords on general environmental risk (e.g., air pollution, environmental issues, EPA, sulfur dioxide) from the climate dictionaries.
我们在建立词典时对气候风险和其他风险进行了区分。首先，公司可能会讨论与能源价格变化相关的气候话题，但后者并不完全与气候风险相关。16相反，我们构建了公司特定的、随时间变化的能源价格风险指数，并将其作为主要分析的控制变量。此外，公司的环境责任和温室气体排放努力可能相关，但并不等同。因此，我们从气候词典中删除了任何有关一般环境风险的关键词（如空气污染、环境问题、环保局、二氧化硫）。

Our final dictionaries consist of 37 unigrams and 1,649 bigrams: the acute physical risk dictionary contains 21 unigrams and 350 bigrams; the chronic physical risk dictionary contains 16 unigrams and 977 bigrams; and the transition risk dictionary includes 322 bigrams. The majority of the dictionaries consist of bigrams, reflecting our deliberate effort to achieve accurate text identification and quantification, as prior research shows that text classification accuracy improves when applying bigrams of words as opposed to unigrams (e.g., Tan, Wang, and Lee 2002; Bekkerman and Allan 2004).
我们的最终词典由 37 个单字和 1,649 个双字组成：急性生理风险词典包含 21 个单字和 350 个双字；慢性生理风险词典包含 16 个单字和 977 个双字；过渡风险词典包含 322 个双字。词典的大部分内容都是由大词组成的，这反映了我们为实现准确的文本识别和量化所做的努力，因为先前的研究表明，与单字相比，使用大词可以提高文本分类的准确性（例如，Tan、Wang 和 Lee 2002；Bekkerman 和 Allan 2004）。

3.2 Measuring climate risk
3.2 衡量气候风险

Next, we construct our firm-level climate risk measures using these dictionaries. Specifically, we first decompose each of the earnings call transcripts into a list of unigrams/bigrams. Because acute or chronic physical risks are often brought up when short-term climate or weather events are reported in news headlines (e.g., hurricane, wildfire, and warm winter), we require their respective keywords to appear in the vicinity (±1 sentence) of at least one risk synonym to ensure that firms are indeed exposed to climate risks (similar to Hassan et al. 2019). Simply mentioning a well-publicized weather/climate event without explicitly mapping to a firm’s risk profile could reflect a desire for attention or shifting of blame, which does not contribute to our physical climate risk measures. We divide the frequency of these occurrences by the length of the transcript, and then multiply the quotient by 10⁴ to reduce the number of decimals. In essence, these measures capture the proportion of a conversation in which acute or chronic weather/climate events as well as a firm’s risk exposure are jointly discussed.
接下来，我们利用这些词典构建公司层面的气候风险度量。具体来说，我们首先将每份财报电话会议记录分解为单字/双字列表。因为在新闻标题中报道短期气候或天气事件时，经常会提到急性或慢性物理风险（例如：）、飓风、野火和暖冬），我们要求它们各自的关键词至少出现在一个风险同义词的附近（±1 句），以确保公司确实面临气候风险（类似于Hassan et al.2019).仅仅提及一个广为宣传的天气/气候事件，而不明确与企业的风险状况相对应，可能反映出企业希望获得关注或推卸责任，这无助于我们的实际气候风险测量。我们将这些事件发生的频率除以记录长度，然后将商乘以 10⁴ 以减少小数点的数量。从本质上讲，这些指标反映了在谈话中共同讨论急性或慢性天气/气候事件以及公司风险敞口的比例。

Transition risk differs from physical climate risk in that it relates to policies and regulations, technological improvements, and evolving climate patterns. Unlike physical risks, transition risk may not materialize in the short run and thus does not pose immediate threats or introduce any uncertainty to a firm’s business operations. As a result, we measure transition risk exposure based on discussions of the keywords in our transition risk dictionary only, without requiring these discussions to appear near a risk keyword. Moreover, firms exhibit varying perceptions of and attitudes toward climate risk, with some discussing and addressing transition risk more proactively than others. With this in mind, we develop an additional measure that captures a firm’s proactiveness when discussing transition risk. To achieve this, we analyze verbs that appear near (within ±1 sentences of) discussions of transition risk keywords in earnings calls, and manually identify a list of 30 verbs that suggest more proactive attitudes when discussing climate issues.¹⁷ Using proactive verbs, we separately identify our transition risk measures with and without proactiveness.
过渡风险与自然气候风险不同，它与政策法规、技术改进和不断变化的气候模式有关。与有形风险不同，过渡风险可能不会在短期内显现，因此不会对公司的业务运营构成直接威胁或带来任何不确定性。因此，我们仅根据过渡风险词典中关键词的讨论来衡量过渡风险敞口，而不要求这些讨论出现在风险关键词附近。此外，企业对气候风险的认识和态度各不相同，有些企业比其他企业更积极地讨论和应对过渡风险。有鉴于此，我们开发了一种额外的测量方法，用于捕捉企业在讨论过渡风险时的主动性。为此，我们分析了财报电话会议中讨论过渡风险关键词附近（±1 句之内）出现的动词，并手动确定了 30 个在讨论气候问题时表明更积极主动态度的动词列表¹⁷ 使用积极主动的动词，我们分别确定了有积极主动和没有积极主动的过渡风险衡量指标。

Applying the above-mentioned procedures, we construct three separate firm-level climate risk measures: (1) acute physical climate risk, (2) chronic physical climate risk, and (3) transition risk. We decompose the transition risk measure into proactive and nonproactive components. All are available at the firm-quarter level.
应用上述程序，我们构建了三个独立的企业级气候风险度量：(1）急性自然气候风险；（2）慢性自然气候风险；（3）过渡风险。我们将过渡风险指标分解为主动和非主动两个部分。所有数据均可在公司季度层面获得。

4 Properties of Firm-Level Climate Risk Measures
4 公司层面气候风险度量的特性

In this section, we provide some preliminary validation using the underlying keywords, present our climate risk measures, and examine their time-series and cross-sectional properties.
在本节中，我们将利用基础关键词进行一些初步验证，介绍我们的气候风险度量，并研究其时间序列和横截面特性。

4.1 Top keywords 4.1 热门关键词

In our first validation exercise, we examine the top keywords—unigrams or bigrams—used to construct the climate risk measures, rank-ordered by the frequency of mentions and frequency weight at the transcript level and report the results in Table 2.¹⁸ The results, reported in columns 1–3, show that hurricanes and hurricane are the most frequently mentioned acute climate unigrams in the proximity of risk synonyms. The keywords storms, drought, flooding, and wildfire(s) are also frequently discussed in earnings calls, trending up in the later few years of our sample period. Columns 4–6 report that weather is the single-most commonly discussed chronic climate keyword appearing near risk synonyms. It is followed by words referencing specific weather conditions, such as temperatures or snow. These keywords clearly confirm that our measures accurately capture acute and chronic climate risks.
在第一项验证工作中，我们检查了用于构建气候风险度量的顶级关键字--单字符串或双字符串，这些关键字按提及频率和频率权重在记录誊本层面进行了排序，并在 Table 2 中报告了结果。¹⁸ 第 1-3 栏报告的结果显示，hurricanes 和 hurricane 是在风险同义词附近最常被提及的急性气候单词。风暴、干旱、洪水和野火等关键词也经常在盈利电话中被讨论，在样本期的后几年呈上升趋势。第 4-6 列显示，天气是在风险同义词附近最常被讨论的慢性气候关键词。紧随其后的是提及具体天气条件的词语，如温度或雪。这些关键词清楚地证实，我们的措施准确地捕捉到了急性和慢性气候风险。

Unlike physical climate keywords, words that indicate transition risk are more evenly distributed across many keywords. Among the most frequently appearing are energy efficiency, renewable energy, solar, clean energy, and alternative energy. In addition to these words, superior energy, higher energy, new energy, the renewable, and the ecosystem are also discussed frequently. Clearly, these keywords accurately signify discussions of transition climate risk. The calculation of fweight in the case of transition climate risk is similar, but we do not require the key unigrams and bigrams to appear in proximity to risk synonyms, which leads to higher average frequencies and fweights. Table IA.7 compares the frequency of climate-related bigrams and unigrams with political-risk-related bigrams from a previous study Hassan et al. (2019) and top climate keywords from another study Sautner et al. (2023). It includes the number of earnings calls and the number of firms that mentioned each of the climate-related words besides their frequency and fweight. Our results show that the frequency of top climate-related bigrams is much higher (about 1,600 times) than that of the top political-risk-related bigrams (e.g., the constitution) in Hassan et al. (2019), and similar to that of top climate keywords in Sautner et al. (2023). Internet Appendix B provides further details.

4.2 Summary statistics

The newly constructed climate risk measures are summarized in Table 1, in which we cap them at the 99th percentile to limit outlier values. Among all 4,719 firms in our sample, 18.0%, 27.2%, and 61.8% show at least one quarter with a positive value for the acute, chronic, and transition climate risk measures, respectively.¹⁹ When we divide these measures by the respective standard deviations (SDs), the three standardized climate risk measures have average values of 0.098, 0.159, and 0.256, respectively. The correlation between the two physical risk measures is about 0.100, suggesting that the two are somewhat related. In contrast, their correlations with the transition risk measure are 0.021 and 0.033, respectively, clearly indicating the distinction between physical and transition risk measures. Conditional only on the presence of firms with at least one positive transition risk value, 23.9% of the firm-quarters are identified as being associated with some proactive keywords when transition risk is discussed.

4.3 Time-series patterns

We now shift to examining the properties of the constructed measures to provide face validation based on time-series and cross-sectional variations. Figure 1 plots the averages of the climate risk measures over time. In panel A, the acute risk series spikes six times over the past 17 years. We identify the corresponding topics discussed in the conference calls that contribute to the increases in climate risk and label each spike. For example, the spike that occurs in 2005 reflects the catastrophic and long-lasting effect of Hurricane Katrina, which flooded the New Orleans area. In contrast, the chronic risk series has remained flat over the past two decades with spikes only between 2012 and 2014. The most commonly discussed keywords during the period was abnormal weather.

Panel B plots the time series for the transition climate risk measure, which shows a steady increase from the start of the sample period through 2008Q3 with a gradual retreat to its 2005 level since then. The downtrend in the recent decade has matched well with that of U.S. greenhouse gas emissions. We observe several local spikes, in 2006, 2008, 2011, and 2015, all of which are driven by more frequent discussion of energy efficiency and renewable energy. Panel C plots the average transition risk measures with and without proactive keywords, divided by their corresponding SDs. The two time series have diverged increasingly since 2008, with firms with proactive responses displaying much lower transition risk than their 2008 levels.

4.4 Industry variations

Industries differ inherently in their exposure to climate risk, so we examine industry variations in our climate risk measures. We regress different climate risk measures on industry dummies, while controlling for time and state fixed effects. Figure 2 plots the coefficients for the NAICS two-digit dummies. The reference industry is other services (NAICS 81).

Panel A shows that utilities face the highest acute physical climate risk among all industries, followed by agriculture, mining, transportation, and construction. A significant portion of the business activities in these industries take place outdoors and thus are subject to disruptions caused by natural disasters. Panel B displays similar patterns, but with a few exceptions. While utilities continue to exhibit high chronic physical climate risk (the second-highest across industries), arts and recreation faces the highest chronic climate risk with agriculture facing the third highest. The industry variations we observe mostly conform to the industry-level exposure to both acute and chronic climate risk.

Panel C shows even wider variations in transition risk than with the physical climate risk measures. Utilities and transportation are subject to significantly higher transition risk than other industries, while service industries face significantly lower transition risk. Panel D displays the industry variations in the proactive transition risk measures. Utilities firms are more likely than other firms to use proactive keywords when their management teams discuss transition risk topics. In contrast, firms that operate in mining, information, and real estate are less likely to use proactive keywords on such occasions. The observed patterns match well with the broader industry-level exposure to climate regulatory risk.

4.5 Firm-level variations

In Table 3, we report excerpts of the transcripts with the highest $ClimateRis{k}_{i,t}$⁠. The transcripts indicating the highest acute climate risk are those of the two largest utility companies in California: Edison International and PG&E Corporation, which have been linked to some of California’s deadliest wildfires. Relatedly, the chronic risk measure captures discussions of both abnormal weather and variability in weather conditions. The transcript indicating the highest chronic climate risk comes from Suburban Propane Partners, a utility company that offers propane primarily for heating.

The transcript indicating the highest transition climate risk is that of CDTI Advanced Materials, a company that provides solutions to automotive emissions control markets in the United States. On August 11, 2011, the company discussed “states such as California continue to demonstrate their commitment for on-road diesel emission reductions through innovative programs to drive early adoption by truck operators.” The other transcripts indicating the highest transition climate risk come from New Jersey Resources Corp, Magnetek Inc., and Lime Energy Co, all of which provide clean or renewable energy services.

5 External Validation

In this section, we conduct a variety of validation tests using external benchmarks to show that our climate risk measures indeed quantify firm-level variations in exposure to climate risks.

5.1 Validating the physical risk measure

Panel A of Table 4 reports the results. The results in columns 1 and 2 indicate that natural disasters in quarter t motivate executives to discuss physical climate risk in quarter t+1. The presence of local natural disasters is associated with a significant 0.085-SD increase in the within-industry-time acute climate risk measure in the subsequent quarter. The effect is statistically significant only in quarter t, not in previous quarters. Similarly, columns 3 and 4 suggest that natural disasters in the preceding quarter are associated with a 0.036-SD increase in the within-industry-time chronic climate risk in the current quarter. Overall, our physical climate risk measures capture variations in a firm’s exposure to local natural disasters, a key driver of physical climate risks.
表 4 的面板 A 报告了结果。第 1 列和第 2 列的结果表明，t 季度的自然灾害促使高管在 t+1 季度讨论实际气候风险。在随后的季度中，当地自然灾害的存在与行业内时间急性气候风险测量值的 0.085-SD 显著增加相关。该效应仅在t季度具有统计意义，而在之前的季度则不具有统计意义。同样，第 3 列和第 4 列显示，上一季度的自然灾害与当前季度行业内时间慢性气候风险增加 0.036-SD 有关。总体而言，我们的自然气候风险度量指标捕捉到了企业在当地自然灾害中的风险变化，这是自然气候风险的一个关键驱动因素。

5.2 Validating the transition risk measure

5.2.1 Correlations with ESG scores

We start our validation of the transition risk measure with the MSCI CCI. We use MSCI rather than other ESG databases for two reasons. First, it is arguably one of the best-accepted ESG data vendors among practitioners and academia (e.g., Engle et al. 2020; Serafeim and Yoon 2023). As the leading global provider of financial indexes, MSCI has successfully incorporated its ESG ratings into a wide range of investment products. Second, CCI is a climate change theme score, which is more closely related to our transition climate risk exposure measures.²¹

To compare the two measures, we first compare the coverage of the two measures. It’s worth noting that the CCI is only available after 2013 and maintains the same value if not updated, while our earnings-call based measures have been available since 2002 and are only applied to the quarter of earnings calls. Figure IA.2 in Internet Appendix B plots the number of unique public firms for each year of our transition risk measure and the MSCI CCI measure. We can see that even during the years when the two data sets overlap, our measure adds substantial coverage beyond the MSCI data, as demonstrated by the green bars. Specifically, for each year from 2013 to 2018, our measure on average provides coverage of transition risk to an additional 952 firms with nonmissing values and 480 firms with positive values. Over the same period, on average, about 225 firms each year in the MSCI CCI data set do not have earnings conference calls and are thus not covered in our sample.

We then match the CCI data with our sample, resulting in a small panel of 15,995 firm-quarters. Panel A of Figure 3 displays the scatterplot between our transition climate risk measure and the CCI, showing a positive and significant correlation between the two series. We formalize the correlation test by regressing $\mathit{ClimateRis}{k}_{i,t}$ on the CCI following a specification that is similar to Equation (1). We report the results in panel B of Table 4. The results in column 1indicate a positive correlation between the two series, which is significant at 1%, suggesting that a one-SD increase in the CCI is associated with a 0.051-SD contemporaneous increase in the transition climate risk. In columns 2 and 3, we document similar results using proactive and nonproactive components of the transition risk measure as the dependent variables, with both coefficients being statistically significant at the 1% level. This set of results provides evidence that our transition risk measure is positively correlated with the CCI within the same industry and time.

Overall, we believe that our transition risk measure is highly complementary to these ESG scores, with several additional benefits. First, our measure is available for a large sample of public firms in the United States over a long sample period starting in 2002, while ESG scores in the CCI are available after 2012. Second, for the same reason, our measure is less subject to selection bias. Third, our measure is more timely and thus can be better used to inform real-time decisions.

5.2.2 Transition risk and CO₂ intensity

In our final validation, we examine how well our transition risk measures correlate with a firm’s carbon intensity. Recent studies use carbon intensity (carbon emissions scaled by total assets) to estimate the effects of a firm’s exposure to climate risks, especially policy and regulatory risks (e.g., Bolton and Kacperczyk 2021a,b). We also examine whether and how firms that are identified with proactive keywords in earnings calls manage their emissions in reality compared with how others do when facing similar transition risks.

We report the results in panel C of Table 4. In specification (1), we find a positive and significant correlation between the transition risk measure and the firm’s CO₂ intensity from year t + 1 onward, with the magnitude increasing over time. A one-SD increase in the transition risk measure is associated with an increase in CO₂ intensity of 0.4531 basis points (which is significant at the 5% level) in year t + 1 and of 0.6939 basis points (which is significant at the 1% level) in year t + 5. In Specification (2), where we separate the transition risk measure into proactive and nonproactive components, we find a positive and significant coefficient for the nonproactive component from year t + 2 onward, not on the proactive component, and the differences are significant at the 10% or lower level. The contrast suggests that, while firms that face higher transition risk and adopt nonproactive responses are associated with higher future CO₂ emissions, those that face higher transition risk but adopt proactive responses are not. In essence, our transition climate risk measures are predictive of the firm’s future carbon emissions.²³

6 Explaining Climate Risk Measures

In this section, we analyze the relative contributions of aggregate, sectoral, and firm-level variations as well as firm-level characteristics to the new climate risk measures.

6.1 Variance decomposition

We first conduct a variance decomposition analysis—calculating how much of the variation in each of the three climate risk measures is accounted for by firm-level characteristics and various sets of fixed effects. In panel A of Table 5, we report R² values from a variety of specifications that explain the climate risk measures. These results indicate that time + state + industries, together, can explain only 2%, 3.4% and 12.4% of the variations in the acute, chronic, and transition risk measures, respectively. Adding interactions between state, industry, and time all help increase the explanatory power of the model, but to a limited extent. Nevertheless, even with the strictest specification, where we control for county-by-time and industry-by-time fixed effects, the model explains less than 12.5% of the variations in any of the climate risk measures, leaving more than 87% attributable to firm-level or other idiosyncratic factors. This result suggests that, unlike natural disaster data or marketwide news about long-run climate risk used by Engle et al. (2020), the majority of variations in our three climate risk measures occur at the firm level.

When we add firm and time fixed effects, the model captures 9.7%, 20.9%, and 65.7% of the variations in the three climate risk measures, respectively. Further adding firm-level attributes and interaction between industry and time or state and time offers some additional power in predicting the two physical risk measures, but not the transition risk measure. This result suggests that our climate risk measures capture both cross-firm differences and within-firm variations in climate risk exposure. For example, the transition risk measure for Sempra Texas Holdings increases to 184.97 in Q3 2013 from 11.10 in Q1 2006.

6.2 Correlations with firm characteristics

Panel B of Table 5 presents the results of regressions relating climate risk measures to a list of important firm-level attributes, all lagged by one quarter, to better understand what types of firms tend to have higher values in the climate risk measures that we constructed. We control for industry-by-time fixed effects to account for time-varying heterogeneity across industries. Among all the variables, the first set is related to a firm’s physical exposure to climate risk. We find an overall positive relationship between the firm’s physical assets and the climate risk measures: the coefficients for PPE and total assets are positive and significant in most regressions. The results suggest that firms that hold more physical assets tend to face higher climate risk exposure.

A second set measures the firm’s financial leverage. We find it to be negatively correlated with the transition risk, but not with the two physical risk measures, suggesting that highly leveraged firms tend to be associated with lower transition risk exposure. This evidence is consistent with the evidence documented by Ginglinger and Moreau (2023), who find that firms with greater climate risk have lower leverage even after controlling for firm characteristics known to determine leverage.²⁴

The final set of measures included in our regressions capture external characteristics of firms, such as the number of analysts covering the firm and institutional ownership. These measures could be correlated with how climate issues are discussed in earnings calls. We find a negative relationship between the number of analysts and our climate risk measures, with one measure being statistically significant. This suggests that firms are less likely to discuss climate-related topics during earnings conference calls when a large number of analysts cover the firm. This could be because with higher analyst coverage, ample information already may be available regarding the firm’s climate exposure, leading to less need for discussion during earnings calls. We do not find a significant correlation between institutional ownership and our climate risk measures.

Lastly, we analyze the correlations between the proactive component of the transition risk measure and firm-level attributes, controlling for transition risk itself. Our results show that firms that carry low leverage, hold more physical assets, and are followed by fewer analysts tend to respond more proactively to rising climate risk.

7 Do Capital Markets Price Climate Risks?

7.1 Baseline results

In panel A of Table 6, we present the baseline results based on the entire sample, where in each column we report the results of a regression of Tobin’s q over various lead times k (1, 3 and 5). For columns 1–3 we use $\mathit{TransitionRis}{k}_{i,t}$ as the main explanatory variable. All coefficients for $\mathit{TransitionRis}{k}_{i,t}$ are negative and significant at the 1% level. For instance, the results in column 1 suggest that a one-SD increase in the transition risk measure is associated with about a 0.0389—1.9% of the mean—decrease in Tobin’s q in the next quarter.²⁵ Also, the magnitude of the coefficient increases slightly when we use Tobin’s q as the dependent variable over a longer horizon (k = 3, 5), suggesting that there is no reversal in the estimated pricing effect. Therefore, our results in this table suggest that transition risk has been priced in equity markets.

For columns 4–6 we include proactive and nonproactive components of our transition risk measures as the main explanatory variables. We also include the firm-level $\mathit{Action}\hspace{0.17em}\mathit{Index}$ as additional control, which captures the overall proactiveness of firms that do not face high transition risk. This measure is calculated as the total frequency of mentions of proactive verbs in an entire transcript (except those that fall within ±1 sentences of climate-related discussions), divided by the length of the transcript. Interestingly, we find that, while the coefficient for nonproactive transition risk is negative and significant, that on proactive measure is nonsignificant. The difference between the two coefficients is statistically different from zero at the 1% level. This result suggests that equity markets appear to discount firms that do not actively manage their transition risk, but not those that are proactive in addressing the risk. This finding is also consistent with our earlier evidence that the nonproactive transition risk measure is associated with higher CO₂ emissions intensity, while the proactive transition risk measure is not.²⁶

7.2 Subsample analysis: Before and after 2010

In this section, we investigate whether there are any time-series variations in the pricing effects of climate risk. The pricing of climate risks is likely to change substantially over time, as noted by Giglio, Kelly, and Stroebel (2021), and the rise in investor attention to climate risk is a relatively recent phenomenon. Some global events play a crucial role in shaping societal expectations and perceptions of climate change, as several studies have shown. For instance, Engle et al. (2020) report that the intensity of climate news coverage peaked in December 2009 when the UN Climate Change Conference in Copenhagen announced a U.S.-backed climate deal with pledges to meet certain emissions reduction targets. Moreover, in January 2010, the SEC issued its first guidance to public firms on existing SEC disclosure requirements as they apply to climate change issues.²⁷ To examine how the pricing of climate risk evolves over time, we conduct the analysis again after splitting the sample into observations made before and after 2010.

In panel B of Table 6, we present the results of this analysis, in which we focus on Tobin’s q in t + 1 as the dependent variable. Based on the results in column 1, the coefficient for $\mathit{TransitionRis}{k}_{i,t}$ is close to zero and not significant in the early period (⁠$\le$ 2009), but turns negative and significant in the late period (⁠$\ge$ 2010) with a much larger magnitude, suggesting that a firm’s climate risk is priced by the capital market with a significant discount in recent years. The contrast between the results in columns 1 and 2 underscores the importance of rising investor attention as conjectured by Giglio, Kelly, and Stroebel (2021) as well as various climate-related initiatives and regulations that were implemented around that time.²⁸ In columns 3 and 4, we report the results obtained when we decompose transition risk into proactive and nonproactive components. We find that it is the nonproactive component that primarily drives the negative relationship between transition risk and market valuation in the late period. The coefficient for the proactive transition climate risk measure is not statistically significant in the early or late periods. Consistent with the evidence reported in panel A, there is a significant difference in the pricing effects of proactive and nonproactive transition risk components.

7.3 Horse-race analysis

We perform additional analyses to assess the robustness of our results regarding the pricing effects of climate risk. First, we carry out a horse-race analysis between our transition risk measure and various alternative measures. These competing measures include: (1) a transition risk measure constructed using SEC filings data; (2) a transition risk measure constructed using firm-related news data; (3) external ESG scores; and (4) climate exposure measures from Sautner et al. (2023). In addition, we also perform sensitivity analysis regarding regression specifications and strategic disclosure considerations.

7.3.1 Transition risk measures constructed using SEC filings data

We construct the first set of alternative measures using Management Discussion and Analysis (MD&A) and Risk Factors (RF) sections in the 10-K/10-Q filings, respectively. We apply the same climate dictionaries to the filings data to construct $\mathit{TransitionRiskMD}{A}_{i,t}$ and $\mathit{TransitionRiskR}{F}_{i,t}$⁠. In panel A of Table 7, we present the results of a horse-race analysis in which we regress Tobin’s q, in t + 1 or t + 5, on both our transition risk measure and one of the two alternative transition risk measures in each regression.²⁹ The results in columns 1–4 show that the coefficients for our transition risk measure remain negative and significant, while those on the alternative transition climate risk measures are not statistically significantly different from zero except for in column 3, where the coefficient for $\mathit{TransitionriskR}{F}_{i,t}$ is less than half of that on our transition risk measure. We note that, compared with the earnings call data, one major drawback of using the Risk Factors section is that it contains only information about the risk factors themselves, with no discussion of how a company addresses or responds to those risks. In columns 5–8, we report the results of an analysis where we decompose transition risk into proactive and nonproactive components. We continue to find that the discount on our transition risk measure is driven primarily by its nonproactive component, which is also negative and significant at the 1% level in all columns, after controlling for competing measures.

7.3.2 Transition risk measure constructed using firm-related news data

The second alternative measure is constructed using firm-related news data. $\mathit{TransitionRisk} \mathit{New}{s}_{i,t}$ is the ratio between the number of news articles related to a firm’s transition climate risk exposure and the number of all news articles related to the company. We construct this measure by applying the same transition risk dictionary to the firm-related news data. Panel B of Table 7 reports the horse-race results. The results in columns 1 and 2 show that the coefficient of our transition risk measure remains negative and significant at the 1% level in all specifications, while the coefficient for $\mathit{Transitionrisknew}{s}_{i,t}$ is not significant, suggesting that there is no relationship between the fraction of firm-specific news that involves climate issues and Tobin’s q. The results in columns 3 and 4 are very similar when we replace the transition risk measure by its proactive and nonproactive components. The significant price discount associated with transition risk is driven by firms that do not undertake proactive responses, while the coefficient for $\mathit{Transitionrisknew}{s}_{i,t}$ remains nonsignificant. In columns 5–8, we repeat the above analysis using 50 as the relevance score cutoff in RavenPack and find almost the same results. This set of results suggests that our transition risk measure contains valuable information not already available in other public sources.

7.3.3 MSCI Climate Change index

The third alternative measure of climate risk is MSCI’s CCI. In panel C of Table 7, we report the horse-race results. In all specifications, the coefficients of our transition risk measure and its nonproactive component are negative and significant at the 5% or lower level, confirming that the estimated price discount indicated in Table 6 is robust in the horse race against the CCI. The coefficient for the CCI measure is also negative and significant at the 1% level, suggesting that firms with higher climate change scores are also priced at a significant discount in the stock market. The coexistence of the two competing measures also suggests that they complement each other in capturing firms’ climate risk exposure.³⁰

7.3.4 Climate risk measures from Sautner et al. (2023)

Our final horse-race test uses the climate change exposure measures developed by Sautner et al. (2023) based on an ML approach as the competing measure. Panel D of Table 7 reports the results. We find that the coefficients for our transition risk measure and its nonproactive component are negative and significant at the 1% level, while those on their climate exposure measures are not statistically significant from zero, as shown in columns 1–4. This pattern persists when we focus on recent years (2010 or later), as Sautner et al. (2023) show a strong correlation between their measures and Tobin’s q using only the data from more recent years. There, we find the coefficient for their regulatory climate exposure measure (CCExposure^Reg) to be marginally significant and small in magnitude compared with that on our transition climate risk measure.

7.4 Controlling for firm fixed effects

Our baseline regressions control for industry-by-time fixed effects, along with firm-level attributes that vary over time. This specification allows us to compare the differential outcomes, such as Tobin’s q, between firms that face high and low climate risk within the same industry at a given time. However, it is important to also consider within-firm variations over time to fully understand the impact of climate risk on firms’ outcomes. To address this concern, we have experimented with an alternative specification where we control for both firm and industry-by-time fixed effects, which allows us to compare within-firm changes in climate risk and firm outcomes while addressing potential endogeneity issues. The results are reported in Table 8. Panel A uses the change in Tobin’s q as the dependent variable and the change in $\mathit{TransitionRis}{k}_{i,t}$ as the main explanatory variable. Our analysis shows that a higher increase in the transition risk measure is associated with a larger decrease in Tobin’s q in the future. The effect is statistically significant at the 10% level or lower after the third quarter (including t + 4, t + 5, $t+6,\dots$⁠), indicating that the stock markets gradually price in the change in transition risk within a given firm.

Panel B focuses on changes in the proactive and nonproactive components of our transition risk measures as the main explanatory variables. The results indicate that only changes in transition risk with nonproactive responses are significantly priced at a discount, while the coefficient for changes in transition risk with proactive responses is negative, but not statistically significant. These findings are consistent with our baseline results in Section 7.1, suggesting that equity markets discount firms that do not actively manage their transition risk, but not those that proactively address the risk.

Overall, our results remain robust after controlling for firm fixed effects and further support the idea that changes in climate risk discussion correlate with changes in Tobin’s q.

7.5 Strategic disclosure in earnings calls

Like any other disclosure data, discussions during earnings calls are not immune to selection bias introduced by strategic considerations. For instance, executives may choose to speak about certain aspects of a firm’s climate risk exposure while not necessarily answering certain questions brought up by analysts. To address selection concerns regarding earnings calls, we restrict the sample in two ways, such that the particular selection concern is more constrained and repeat the pricing regression to see if our estimates remain robust. In the first exercise, we filter out earnings calls where we detect an extreme tone. The literature on qualitative disclosure has shown that management can strategically determine the tone of textual disclosures to achieve certain outcomes (e.g., Lang and Lundholm 2000; Feldman et al. 2010; Arslan-Ayaydin, Boudt, and Thewissen 2016). In the second exercise, we exclude earnings calls which rank in the top quartile based on the number of “nonanswers” from management during a call, measured using the latest linguistic analysis method proposed by Gow, Larcker, and Zakolyukina (2021).³¹ We report the results of this analysis in Table IA.5. We find that the price discount associated with high transition risk is still significant based on the restricted samples. Our results suggest that the selection issue is not a major concern for our analysis.

8 Firms’ Responses to Climate Risks

In this section, we investigate whether firm-level climate risk exposure affects a firm’s real business activities. To do so, we estimate differences in corporate responses associated with high climate risk by running regressions specified in Equation (2), where the dependent variable includes CapEx, R&D expenditures, the fraction of green patents, and employment over horizon t + k (⁠$k>$ 0). The main explanatory variables are transition risk and its proactive and nonproactive components in t. We control for a firm’s total assets as well as industry-by-time fixed effects. In essence, we compare differences in corporate responses between firms that face high and those that face low transition climate risk, as well as between firms with and without proactive responses to transition risk.

8.1 Investment

The theoretical literature has offered mixed predictions regarding investment under uncertainty. While Bernanke (1983), Pindyck (1991), Pindyck and Solimano (1993) and Dixit and Pindyck (1994) predict a decline in investment in times of high uncertainty, other studies, such as Oi (1961), Hartman (1972, 1976), Abel (1983), Roberts and Weitzman (1981), and Bar-Ilan and Strange (1996), predict a positive relationship. Ultimately, how firm-level investment varies with climate risk exposure is an empirical question.

Table 9 presents the results of an analysis using CapEx scaled by total assets as the dependent variable. The results in columns 1–3 indicate a positive, but not significant, coefficient for $\mathit{ClimateRis}{k}_{i,t}$⁠, suggesting that there is no statistically significant difference in future investment between firms that face high and those that face low transition risk. In columns 4–6, we investigate differences between the responses of firms that do and those that do not respond to climate risk proactively. To do so, we regress the same set of firm-level outcomes on transition risk with and without proactive keywords. We see that the coefficients for two of the transition risk measures are both positive, but only the coefficient for proactive transition risk is statistically significant (at the 1% level), suggesting that firms that proactively respond tend to increase their CapEx following an increase in transition risk. A one-SD increase in transition risk with proactive keywords in t is associated with a 0.046-percentage-point increase in CapEx in t + 1 and about a 0.06-percentage-point increase in CapEx in t + 3 and t + 5.³² The estimates are economically meaningful, representing approximately 1.6%–2.3% of the average investment level. In the bottom row, we report the differences between the two coefficients along with their significance levels based on F-tests, showing that the difference in CapEx between proactive and nonproactive firms, when both face high climate risk, is significant at the 10% level in t + 5.

8.2 Innovation

To reach net-zero emissions or decarbonization, firms are inevitably required to innovate or change the way they do business. Thus, innovation is a viable and important response for firms facing high transition risk. We consider two measures of innovation: one is R&D expenditure, scaled by assets, the other is the fraction of green patents. In panel A of Table 10, we report the results for R&D expenditures. We find negative and significant coefficients for $\mathit{ClimateRis}{k}_{i,t}$ in columns 1–3, suggesting high transition risk is associated with lower R&D expenditures. A one-SD increase in transition risk is associated with a 0.0529- to 0.0565-percentage-point decrease in future R&D expenditures. Again, the coefficients are fairly stable over various horizons of R&D expenditures. The results in columns 4–6 suggest that the negative relationship between transition risk and a firm’s future R&D expenditures is significant only for the firms that do not proactively respond, not for proactive firms.

In panel B of Table 10, we report the results of regressions using green patent measures as the dependent variable. The results in columns 1–4 are based on all firms and use an indicator of having at least one green patent as the dependent variable. We find a positive, but not significant, coefficient for $\mathit{ClimateRis}{k}_{i,t}$ in columns 1 and 2, suggesting that there is no statistically significant difference in future green patents between firms with high and low transition risk. For columns 3 and 4, we investigate differences between the responses of firms that do and those that do not respond to climate risk proactively. We see that the coefficients for two of the transition risk measures are both positive, but only the coefficient for proactive transition risk is statistically significant (at the 5% level), suggesting that firms that proactively respond are more likely to innovate via green patenting when facing high transition risk. A one-SD increase in transition risk with proactive keywords in t is associated with a 0.01-percentage-point increase in the likelihood that a green patent is filed in t + 1 and 0.01-percentage-point increase in t + 2. The estimates are economically meaningful, representing approximately 12.5% of the average probability that a green patent is filed.

The results in columns 5–8 are based on patenting firms only, using the ratio of green patents to the total number of patents filed by a firm as the dependent variable. We find positive and significant coefficients (at the 1% level) on $\mathit{ClimateRis}{k}_{i,t}$ as shown in columns 5 and 6, suggesting that firms that face high transition risk are associated with a higher ratio of green patents. A one-SD increase in transition risk with proactive keywords in t is associated with a 0.0321-percentage-point increase in the ratio of green patents in t + 1 and a 0.0331-percentage-point increase in t + 2. The results in columns 7 and 8 show that the coefficients for two of the transition risk measures are both positive and significant, but the coefficient for proactive transition risk is slightly higher and more significant (at the 5% or lower level). A one-SD increase in transition risk with proactive keywords in t is associated with a 0.0251-percentage-point increase in the ratio of green patents in t + 2.

Given the significant and positive relationship we find between a firm’s greenness and their proactiveness in managing transition risk, we conduct further analysis to explore the attributes of proactive firms and their potential differential impact on firm valuation in Internet Appendix C. Starting with firms that have patented green technologies and those that have not but are proactive in their responses to transition risk, we find that green patenting firms are more likely to be proactive in addressing transition risk, while nongreen patenting firms do not show a significant difference in being proactive relative to firms that do not patent. Panel A of Table IA.8 presents the results. Panel B of that table shows that while both types of proactive firms are valued positively by the equity markets, the difference between green proactive firms and those with nonproactive responses is much larger than that between nongreen proactive firms and those with nonproactive responses. Both differences are statistically significant at the 1% level, indicating that the equity markets tend to value green proactive responses to transition risk more than nongreen proactive responses.³³

8.3 Employment

Another strategy at a firm’s disposal for responding to rising climate risk is adjusting employment (e.g., through plant closings, layoffs, or hiring freezes). Layoffs and plant closings have been commonly adopted by executives at public companies to increase productivity, address ongoing risks, and appeal to capital markets. The results, reported in panel C of Table 10, indicate that there is a negative and significant relationship (at the 5% level) between transition risk and the logarithm of the employment level in the following 2 years. A one-SD increase in transition risk is associated with an approximately 0.02-percentage-point decrease in a firm’s employment stock. The negative relationship is primarily driven by firms that do not proactively respond. The relationship is not statistically significant for firms that proactively respond.

8.4 Summary

In summary, we find a significantly negative relationship between transition risk and R&D expenditure as well as employment, driven primarily by firms that face high transition risk but do not proactively respond. In contrast, firms that proactively respond increase their total CapEx investment and file more green patents following an increase in their transition risk.³⁴ These findings, while revealing divergent responses on the part of firms facing high transition risk, may not suggest any causal relationships between the two, because our constructed measures simply capture transition risk discussions during earnings calls. Instead, our evidence suggests that the new measures capture new and valuable information about business conditions and can be highly predictive of changes in these corporate outcomes.³⁵

9 Conclusion

This paper quantifies the presence and materiality of firm-level climate risk exposure. We develop a novel set of firm-level climate risk measures, covering both physical and transition risks, by applying a modified textual analysis method to earnings call transcript data. Most variations in physical climate risk appear to be idiosyncratic factors that may be unrelated to firm-level attributes, while most variations in transition risk can be explained by idiosyncratic factors at the firm level. Using external benchmarks, we find that our three risk measures capture changes in the respective types of climate risk a company faces. As a unique innovation of our study, we also measure firms’ proactiveness in addressing climate issues. One key finding of our study is that firms that face higher transition risk, especially those that do not proactively respond, are valued at a discount in the equity market. Horse-race analyses show that our measures offer unique value for studying how capital markets price climate risk, particularly transition risk.

Using several corporate outcomes as dependent variables, we find that firms that face high transition risk significantly decrease their R&D expenditures and employment. This negative relation is primarily driven by firms that do not proactively respond to rising climate risk. Firms that proactively respond to this risk tend to significantly increase their total CapEx investment and file more green patent applications. Thus, firms’ attitudes toward climate issues—whether or not proactive— matter significantly in determining how firms respond to rising climate risk.

Our key finding that firms that do not proactively respond to transition risk are valued at a discount underscores the importance of disclosing climate risks in a transparent and comprehensive manner to ensure that investors have access to accurate information and can make informed investment decisions. Our ability to identify variations in firm-level climate risk exposure and responses suggests that when such information is available, investors find it relevant. Indeed, regulators have begun to focus on how best to provide this information to investors. In March 2021, the SEC created a Climate and ESG Task Force to identify climate and ESG-related misconduct. In March 2022, the SEC proposed new rules that require public companies to report climate-related risks and emissions data in registration statements and annual reports.

Acknowledgement

We thank Itay Goldstein (the editor), two anonymous referees, William Cong, Gustavo Cortes, Kris Gerardi, Gerard Hoberg, Joel Houston, Chris James, Sehoon Kim, Nitish Kumar, Hao Liang, Tim Loughran (discussant), Xin Liu (discussant), Kevin Mullally, Veronika Penciakova, Jay Ritter, Christoph Schiller (discussant), Jenny Tucker, and Baolian Wang and conference/seminar participants at the 2021 AFA, the 2021 Second Sustainable Finance Forum, the 2021 RiskLab/BoF/ESRB Conference on Systemic Risk Analytics, the 2021 China International Risk Forum, the 2021 Rising Star Conference, the 2020 NFA, the 2020 FMA, the 2020 Shanghai Green Finance Conference, Auburn University, the Federal Reserve Bank of Atlanta, Fordham University, the University of Florida, and UT Dallas for helpful comments and suggestions. We are grateful to Söhnke Bartram, Kewei Hou, and Sehoon Kim for sharing the GHGRP-Compustat linktable and Pedro Matos for sharing the Global Corporate Patent data set. We also thank Osama Mahmood, Xiaoxiao (Ray) Sun, Da Tian, and Mingyin Zhu for excellent research assistance. Vincent Yao gratefully acknowledges financial support from Hong Kong Institute for Monetary and Financial Research. This paper represents the authors’ views, which are not necessarily the views of the Hong Kong Monetary Authority, Hong Kong Academy of Finance Limited, or Hong Kong Institute for Monetary and Financial Research. All remaining errors are our own. Supplementary data can be found on The Review of Financial Studies web site.

Appendix

Footnotes

Author Notes

Authors have furnished an Internet Appendix, which is available on the Oxford University Press Web site next to the link to the final published paper online.

References