Estimating companies’ GHG emissions using information from individual financial accounts and building data

The changing climate conditions also have implications for the future performance of companies, particularly through transition risks such as rising prices for greenhouse gas (GHG) emissions resulting from emission caps. Since companies play a vital role in the financial system, the implications of climate change indirectly increase the risk exposure of financial institutions. As a result, measuring the potential risk exposure posed by climate change has become a priority for central banks. Reliable data on individual companies’ greenhouse gas (GHG) emissions is essential for evaluating transition risks to the financial system. However, plant-level GHG emissions data is available for only a limited subset of companies. This study introduces a novel approach to address this data gap by estimating CO 2 emissions using companies’ individual balance sheet data. The items on balance sheets and income statements reflect the use of production factors and the generation of value added. Using a panel of firms that report emissions under the EU Emissions Trading System (EU ETS) and publish financial statements, we account for the selectivity of available emission data and the non-random nature of missing data (MNAR). We also incorporate an innovative geographical data source on company building characteristics. We compare different estimation methods and propose a bottom-up approach to carbon accounting as a complement to top-down methods that break down aggregate emission statistics to the company level. Our findings suggest that models incorporating additional production factors—beyond employment and sector information—enhance the predictive power for imputing GHG emissions and provide more accurate estimates than traditional top-down approaches. Building function emerges as a key predictor of emissions, while employment effects are negligible. The inclusion of additional novel data sources can help mitigate the selection bias caused by the MNAR process. These results are particularly relevant for micro-level analyses.