Enhancing Climate Stress Tests with Machine Learning

As governments around the world tighten climate policies to meet their emissions reduction targets, businesses are increasingly exposed to transition risks. Higher carbon prices and stricter environmental regulations can significantly increase operating costs for many companies. These developments not only affect individual firms but may also have wider implications for banks and the stability of the financial system. 
Against this backdrop, Christian Haas, Karol Kempa, Ulf Moslener and Sebastian Rink examine how machine learning techniques can improve the assessment of climate-related risks within financial stress-testing frameworks. 
One of the key challenges in climate stress testing is the limited availability of company-level emissions data. This is particularly true for small and medium-sized enterprises (SMEs), which are often underrepresented in existing analyses. To address this issue, the authors develop a machine learning model that estimates firms’ greenhouse gas emissions based on financial and company-specific data. This approach enables climate risk assessments to cover far more firms than was previously possible.
Drawing on data from approximately 1.4 million European companies, the study examines how firms would respond to a range of carbon price shocks. The authors simulate carbon price increases ranging from €1 to €500 per tonne of CO₂ and analyse their effects on corporate profits, asset values and default probabilities.
The findings show that an additional carbon price of €100 per tonne would significantly weaken the financial position of many firms. The number of loss-making companies nearly doubles, while the volume of loan defaults also increases significantly. These adverse effects become even more pronounced as carbon prices rise further. 
The analysis also reveals substantial differences across firms. Small and medium-sized enterprises are considerably more vulnerable to carbon price shocks than large corporations, which typically benefit from stronger financial buffers and greater resilience. Particularly severe effects are observed in sectors such as mining, transport, manufacturing, construction, and retail and wholesale trade. 
A further contribution of the study lies in its comparison of firm-specific emissions estimates with conventional approaches based on sector averages. The results indicate that relying on industry averages often leads to an underestimation of climate-related financial risks, as significant differences between individual firms remain unaccounted for. This is especially relevant in emissions-intensive industries, where carbon emissions can vary considerably even within the same sector. 
Despite these advances, uncertainty remains an important challenge. As direct emissions data are unavailable for many companies-particularly SMEs-the results of climate stress tests depend heavily on the quality of the underlying estimates. The study therefore highlights the importance of more comprehensive and reliable climate-related disclosures standards. 
Overall, the study demonstrates that machine learning methods can significantly enhance both the accuracy and scope of climate stress testing. By enabling a more granular assessment of transitions risk, these approaches can help financial institutions and regulators better understand the potential consequences of climate policy measures and strengthen the resilience of the financial system.