Understanding and predicting corporate carbon emission intensity is a critical priority amid global climate change and increasing stakeholder pressure. This study develops and evaluates machine learning models to forecast carbon intensity, focusing on identifying key predictors and analysing performance across pre- (2016–2019) and post-COVID (2020–2023) periods. We employ six models, three linear (Lasso, Ridge, Elastic Net) and three non-linear (Random Forest, XGBoost, Neural Network), on a comprehensive dataset of publicly listed firms. Our findings consistently demonstrate the superior predictive accuracy of non-linear models, with eXtreme Gradient Boosting (XGBoost) emerging as the most robust performer across all periods and predictor sets. A key insight is the significant predictive power of governance-related variables. These factors are particularly strong predictors when direct historical emissions data are excluded, a scenario simulating the common real-world challenge of incomplete reporting. The analysis also reveals shifts in feature importance between the pre- and post-COVID eras, suggesting that the drivers of corporate carbon intensity are not static. This research contributes to the application of machine learning in corporate sustainability, offering practical insights for investors, policymakers, and corporate managers. Ultimately, we highlight the power of predictive modelling to uncover fundamental drivers of environmental performance, with a notable emphasis on the often-overlooked predictive capacity of governance metrics.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号