Model parameter-based transfer learning for ESG score prediction in developing markets

While ESG (Environmental, Social, and Governance) assessment plays a key role in sustainable finance, data scarcity and noise in emerging economies hinder robust model development. To address this, we propose a model parameter based transfer learning with random forest (MPBTL-RF) approach for domain adaptation situations where source data are not available. The proposed model is evaluated using three traditional learning approaches: Random Forest (RF), eXtreme Gradient Boosting (XGB), and Feed forward Neural Networks (FNN). Cross-validation is used to assess model generalizability, and domain adaptation is tested through in domain and out-of-domain settings. The proposed MPBTL-RF approach achieves competitive performance compared to traditional baselines in scenarios with limited training data, offering time advantages with predictive efficiency and stability. This work demonstrates how machine learning pipelines can adapt to data-constrained, real-world domains, fostering the synergy between AI (Artificial Intelligence) and business.