Deep Learning for Decision Making and the Optimization of Socially Responsible Investments and Portfolio

How can we leverage machine learning to build responsible portfolios?

Nhi N.Y. Voa, Xuezhong Heb, Shaowu Liua, and Guandong Xu address this question in their article « Deep learning for decision making and the optimization of socially responsible investments and portfolio ».

The study introduces a Deep Responsible Investment Portfolio (DRIP) model that outperforms traditional portfolio models, sustainable indexes, and funds in both financial performance and social impact.

The paper's main conclusions include:

• DRIP portfolios consistently outperform the S&P 500 index, with an average annual return of 16.8% compared to 13.2% for the S&P 500 over the study period.
• The reinforcement learning approach in DRIP allows for dynamic portfolio rebalancing, resulting in a 24.68% increase in cumulative returns compared to static models over a 5-year period.
• The proposed Mean-Variance-ESG portfolio optimisation model demonstrates better performance than traditional Mean-Variance models, achieving a 13.95% higher Sharpe ratio and a 17.65% improvement in ESG scores.
• The model shows resilience during market downturns, with DRIP portfolios experiencing smaller drawdowns (-18.2%) compared to the S&P 500 (-33.9%) during the 2020 market crash.
• The DRIP model's neural network achieves superior prediction accuracy compared to other LSTM variants, with a Mean Absolute Percentage Error of 0.0148 for quarterly predictions and 0.0382 for yearly predictions.

The study responds to the limitations of traditional investment theories in incorporating ESG factors into portfolio construction with an automated, data-driven approach that balances financial returns with social impact.

The complexity of the DRIP model may however raise concerns about its interpretability, especially when using ESG ratings, whose consistency and reliability across different providers is debatable.

As mentioned by the authors, the model's performance during specific market conditions or across different asset classes may need further validation to ensure its robustness in various scenarios.