Comparative Analysis of Predictive Models for Financed Emissions

Sammanfattning

Financial institutions have a great interest in creating predictive models for their lending and investment portfolios. Detecting outliers and conducting scenario analysis using predictive models not only improves their risk management but also helps them meet their goals and maintain a high regulatory standard. The dominant source of emissions for banks is indirect greenhouse gas emissions (Scope 3), which are mainly emissions related to financing and investment activities. Usually, direct emissions (Scope 1 and 2) represent only a fraction of their total emissions. Therefore, understanding and accurately predicting their Scope 3 emissions, particularly the financed emissions, is important for their climate strategies. The goal of this study is to explore different models to find which of the selected ones are most suitable to predict financed emissions, specifically focusing on the commercial real estate sector. Some degree of imputation was necessary to avoid removing large parts of the dataset while keeping relevant variables. The iterative imputer method, using the random forest regressor, outperformed the less complex k-nearest neighbour imputation in general. Six different mathematical models were explored: linear regression, ridge regression, bagged trees, random forest, Catboost, and finally, neural network. The study found that bagged trees and random forest performed the best when it came to mean absolute error (MAE), where CatBoost also had good performance. Regarding root mean squared error (RMSE) and R-squared, the neural network performed best. Furthermore, linear regression had worse performance, and the regularisation from the ridge did not noticeably improve the results. Finally, the best model varies depending on what you are after: accuracy or interpretability.