Faculty of law blogs / UNIVERSITY OF OXFORD

Corporate Climate: A Machine Learning Assessment of Climate Risk Disclosures


Julian Nyarko
Assistant professor of law at Stanford Law School
Eric Talley
Isidor and Seville Sulzbacher Professor of Law at Columbia Law School


Time to read

4 Minutes

For more than a decade, regulators and policymakers alike struggled to articulate and implement disclosure expectations for public companies associated with climate risk. Although in 2010 the Securities and Exchange Commission (SEC) issued a guidance reminding that US-traded companies are required to disclose material climate risks in their public filings, the guideline is notably infirm in providing an organised approach for making such disclosures, instead leaving it largely in the issuers’  discretion whether, how and in what form to disclose climate-related risks. The resulting landscape is a disorienting array of disclosures scattered across dozens of types of securities filings, wherein the process of benchmarking them to other filings and standards underlying the majority of the US securities regulation is a challenging task. Although some efforts exist to identify climate risk disclosures, they have been limited in scope.

In this post, we present machine learning tools aimed both at extracting climate risk disclosures and inferring to what extent issuers’ characteristics affect the likelihood of submission.

In 2010, the SEC approved a notorious directive requiring issuers to disclose information about the potential effects of physical climate change events and related regulations on business operations and financial positions. As afore-mentioned, the SEC’s guidance did not prescribe the formal requirements to abide by for fulfilling the disclosure, nor did it provide any pertinent templates guiding issuers, thereby triggering a trade-off. On one hand, empowering market forces to shape the disclosure process relieves issuers’ regulatory burden. On the other hand, the lack of procedural consistency entails high coordination costs. And those costs are clearly apparent. In 2018, the General Accounting Office (GAO) published an analysis of firms performances, highlighting a lack of transparency in the climate risk disclosure process and consequent hurdles facing regulators when deprived of the parameters assessing to what extent disclosures are appropriate.

To address these issues, our project deploys machine learning tools identifying the set of public companies that have submitted climate risk disclosures in compliance with the SEC’s interpretive guidance. We begin by querying company disclosures from the EDGAR database. To identify parts of the documents that may contain a climate risk disclosures, we use a suitable set of keywords and -phrases, containg, for instance, climate, global warming, and greenhouse gas. The keywords are intentionally broad, leading to an overinclusive set on candidate disclosures. In a next step, we label a subset of these candidate disclosures by hand for whether they actually contain climate risk disclosures or not. Finally, using these labelled texts, we train and cross-validate a machine learning classifier. The classifier is able to determine the presence of a climate risk disclosure with exceptionally high accuracy and we use it extract climate risk disclosures from the complete set of all company filings.

 Examining the identified climate risk disclosures shows that ‘cookie cutter’ disclosures are common—whereby many issuers submit nearly identical ‘boilerplate’ disclosures that lack firm-specific nuance. At the same time, a variety of issuers appear to draft tailored types of disclosures that do not follow a standard template. Such bespoke disclosures are more likely to be highly informative for investors.

Further, the model provides empirical evidence that submitting climate risk disclosures often goes viral in industries. Once a few issuers start submitting, the practice gains momentum within a particular sector. In extraction, construction, transportation, and energy industries the overall trend is to submit disclosures more frequently than in others (eg finance), which coheres with broader intuitions and data about carbon footprints—though crypto-currency mining may eventually turn to be sensitive for the financial industry as well.

While the first part of the project sheds considerable light on the issues upon which submitters focus in climate risk disclosures, the second part aims to identify the category of issuers likely to draft them. Relying on asset pricing models, it uses a wealth of data provided by the Sabin Center at Columbia University to develop a climate risk factor. Intuitively, this factor captures the degree to which asset returns in the equities market are susceptible to climate-related events, such as severe weather or global warming. The resulting data allows us to identify whether reporting practices coincide with actual susceptibility to climate risks or not. Although the project is at preliminary stages, we find only a weak correlation, suggesting that the current disclosure regime may be insufficient to incentivize the right companies to disclose their risks. Finally, despite the fact that securities laws are not evaluated in terms of their potential for reducing carbon emissions on a global scale, the empirical results show that current disclosure practices are mainly committed to investor protection. Consequently, an ideal form of climate risk disclosure is still restrained on how climate change translates into investment risk and return. Thus, if a company’s activities involve climate risk that are unlikely to feed back onto investor returns, those activities would (by definition) not be material for purposes of conventional securities laws. Although it is conceivable that securities regulators may attempt to change the definition of materiality to have more of a stakeholder bent, they would receive significant pushback if they were to propose an opaque redefinition of the term. This is particularly true in litigation-driven jurisdictions like the US, where plaintiff and government attorneys are uniquely adept at seizing upon arcane definitions in their legal filings.

A better prospect, in our view, would be to encourage the continued use of climate-related financing, a still-novel market that continues to grow rapidly. Suppose, for instance, an issuer was to incur a significant debt position, whereby key borrowing terms (eg interest rates) are contingent on meeting a specified set of environmental performance benchmarks. Because equity investors are residual claimants on the firm, the failure of the firm to meet those specified benchmarks would have a direct effect on its debt servicing costs, thereby diminishing returns for equity holders as well. Consequently, even if climate risk would not have been material to shareholders under conventional debt contracts, green finance may be able to align global climate risk with private investment risk. Here regulatory authorities may be able to move the needle somewhat, by, among others, providing tax incentives for green financing terms. Such an initiative, when joined with better information about climate risk disclosures, may prove to be a satisfying approach for using pre-existing securities laws to address global climate externalities.

Julian Nyarko is an assistant professor of law at Stanford Law School.

Erick Talley is the Isidor and Seville Sulzbacher Professor of Law at Columbia Law School.

This post is part of the series ‘Business Law and the Transition to a Net Zero Carbon Economy’. This series consists mainly of posts summarizing papers presented and presentations made at the 5th Annual Oxford Business Law Blog conference on ‘Business Law and the Transition to a Net Zero Carbon Economy’ which took place online on 25 to 27 May 2021. The recordings are available here. This post is forthcoming in Andreas Engert, Luca Enriques, Georg Ringe, Umakanth Varottil and Thom Wetzer (eds), Business Law and the Transition to a Net Zero Carbon Economy (CH Beck - Hart Publishing 2021) (forthcoming).


With the support of