Estimating Corporate Greenhouse Gas Emissions

Can open models estimate the emissions companies do not report?

Mohamed Fahmaoui, Thibaud Barreau, and Peter Tankov review machine learning methods for estimating carbon emissions and propose an alternative in "Estimating Corporate Greenhouse Gas Emissions".

They combine reported emissions from CDP and Refinitiv with company financials, GICS sub-industries, and public regional data such as grid carbon intensity from Ember and carbon prices from the World Bank.

They train three models on CDP-reported emissions and test them with a design close to real-world use in poorly covered markets. Their main conclusions include:

• CatBoost performs best on every target, explaining about 86% of the variation in scope 1 emissions and 81% for scope 2, but only 63% for scope 3.
• Scope 3 errors approach an order of magnitude, and 90% prediction intervals span more than two orders of magnitude for scope 3 while containing the true value only ~80% of the time.
• A restricted model using only revenue, profits, assets, and sector, available even for private companies, still explains 80% of scope 1 variation and loses almost nothing for scope 3.
• Energy consumption is the most influential predictor for scope 1 and 2, while revenue and assets lead in the restricted model, and headquarters location matters far less than sector.
• Accuracy improves with company size, and financial and utility companies are the hardest to estimate for scope 1, reflecting wide differences in reporting methods and boundaries.
• Outlier filtering is deliberately asymmetric, removing about 5% of under-reported values but only 0.5% of over-reported ones, so estimates lean towards overestimation rather than rewarding non-disclosure.

The paper showcases an open-source model offering a transparent benchmark to challenge vendor figures, which is valuable as models underestimating emissions make silence look greener than disclosure.

Trimming implausibly low values more aggressively than high ones keeps the incentive to report intact, an asymmetry the authors note existing methodologies usually ignore.

Prediction intervals however still under-cover, so scope 3 point estimates should not drive single-name decisions until calibration and operating country data improve.