Algorithm for calculating the carbon footprint of a company’s value chain using machine learning

Large corporations and select small and medium-sized enterprises must report not only their own greenhouse gas emissions but also Scope 3 emissions from their value chain according to the new CSRD directive and ESRS standards. To determine Scope 3 emissions, companies are not required to collect data directly from entities in their value chain but can use proxy variables and models based on cross-sectional data analysis. A dataset containing data on American companies and their emissions from the GHGRP program of the EPA was obtained through data scraping. The data was then modeled using spatial regression models that can account for spatial relationships between data that are statistically significant according to Moran’s I test and Geary’s C test. The accuracy of individual models was tested on unseen spatial structures using Spatial Leave-One-Out Cross Validation (SLOOCV), with the spatial error model achieving the best results based on RMSE. The model can be further optimized by including data from Europe, Asia, and South America to improve its generalization to spatial structures and relationships different from those in the USA. The practical benefits of using the algorithm proposed in this paper is a significant simplification of estimating the Scope 3 emissions for companies which are required to publish ESG reports in accordance with CSRD. Companies can use the proposed model to create reliable predictions of GHG emissions of entities in their value chain based on their location and easily available information which was used as independent variables when constructing the model.