Predictive Modeling of Corporate Air Pollutant Emissions by Global Industry

Abstract

Accelerated climate change has led to stronger impetus for regulatory action across nation states and the private sector. Companies are now pushed to increase accountability for the environmental impact of their business and operations, which require accessible tools to quantify environmental outputs. A major greenhouse gas, nitrogen oxide (NOx), faces suboptimal reporting among the private sector. In the absence of reporting, imputation methods were developed to estimate the NOx emissions of firms. This project seeks to contribute to and improve on an existing methodology that the Impact-Weighted Accounts (IWA) team at the Harvard Business School employs for determining NOx emissions of firms which uses the EXIOBASE database. A random forest model was proposed with input features that consist of financial metrics and emissions data of greenhouse gasses to estimate NOx emissions. The model had an R^2 score of 0.947 (out of a best possible of 1.0), which is a 2.2% improvement compared to a baseline k Nearest Neighbors model. Of the input features, greenhouse gas emissions and sub-industry categorization have the top two largest feature importances of 0.428 and 0.221 respectively. Since sub-industry categorization is an important feature, variations in data available across different sub-industries limited the performance of the model across sub-industries. Finally, NOx emissions as estimated by the model had a correlation of -0.056 with the Tobin’s Q, as well as -0.044 with the Market-to-Book ratio, which suggests that elevated NOx emissions pose a financially material risk for companies. Overall, the random forest model produced estimates that were more accurate than the EXIOBASE methodology currently used by the IWA team.