Reinforcement Learning for Low-Carbon Investment Strategies

Abstract

Sustainable investing has become increasingly important in financial markets, while reinforcement learning has appeared as a useful method for solving complex decision-making problems such as portfolio optimization. This research examined whether reinforcement learning models that use CO2 emission scores, ESG metrics, and financial data can create investment strategies that are both profitable and more sustainable. A portfolio management framework based on the Proximal Policy Optimization algorithm was implemented. The model used predicted financial returns (provided by LSTM models), current financial re-turns, ESG scores, and sector-level CO2 emission scores as input features. The reward function combined financial and sustainability objectives using weighted parameters. Model hyperparameters were optimized using a genetic search algorithm. The reinforcement learning models were compared with portfolios built using Modern Portfolio Theory. The results showed that reinforcement learning models were able to construct portfolios that balance financial performance and sustainability indicators. Models focused on financial returns achieved the highest profitability but lower sustainability metrics. In contrast, sustainability-focused models produced higher ESG and emission scores but lower financial returns. Balanced models achieved stable financial performance while significantly improving sustainability indicators. The results suggest that combining financial and sustainability data into a reinforcement learning model can produce investment strategies that are both profitable and more sustainable than strategies based only on financial data.