Machine Learning for Socially Responsible Portfolio Optimisation

💡 Research Summary

The paper addresses a gap in socially responsible investment (SRI) by extending the classic mean‑variance (MV) portfolio optimization framework to incorporate Environmental, Social, and Governance (ESG) scores, creating an ESG‑MV model. The authors first evaluate three machine‑learning techniques—Random Forest (RF), Long Short‑Term Memory (LSTM), and Convolutional Neural Network (CNN)—for forecasting ETF closing prices. Using mean absolute scaled error (MASE) and root mean squared scaled error (RMSSE) as performance metrics, RF consistently outperforms the other models and is selected to generate two‑month ahead price forecasts for a set of U.S. ETFs. These forecasts, together with historical price data (2011‑2021) and ESG ratings sourced from VettaFi, feed into a portfolio optimization stage. Traditional MV optimization maximizes the Sharpe ratio, while ESG‑MV adds the average ESG score to the objective, still seeking to maximize a Sharpe‑adjusted metric. Optimization is performed with Sequential Least Squares Programming (SLSQP) on randomly sampled subsets of 100 ETFs across twelve runs. Results show that ESG‑MV portfolios achieve a 12.29 % lower Sharpe ratio on average but deliver a 65.72 % higher mean ESG score compared with MV portfolios. Risk (annualized volatility) drops by about 20 % and expected return falls by roughly 21 %, indicating that ESG‑focused investors accept lower financial returns for reduced risk and higher social impact. All portfolios maintain Sharpe ratios above 1, confirming acceptable risk‑adjusted performance. The study acknowledges limitations: ESG scores are treated as static, over‑fitting appears in RF and CNN models, and regularization techniques are not fully explored. Future work is suggested to model dynamic ESG trajectories, incorporate additional constraints (e.g., carbon caps), and apply reinforcement‑learning‑based optimization. Overall, the research demonstrates that integrating machine‑learning price forecasts with ESG‑aware optimization yields feasible, socially responsible portfolios that balance financial and ethical objectives.