AI for ESG Integration: Training Machines to Predict Sustainable Alpha

Challenge to overcome

Investment managers face the challenge of integrating ESG factors with traditional ones like value, growth, and momentum. Mastering this integration is crucial for optimizing stock selection and portfolio construction.

AI application / use case

The influx in sustainability data has equipped quantitative researchers with input to develop various ESG-based investment factors. Leveraging on this, RAM-AI has developed a deep learning framework to model the complex interactions among different input features, and to integrate ESG and traditional factors for predicting stock returns.

One of the advantages of neural networks in this domain is their capacity to identify feature interactions autonomously. For instance, a well-trained network can discern how environmental policies interact with stock valuation to influence financial performance, without being explicitly programmed to seek out these interactions.

In this case, the network learns to identify intricate patterns that signify the interactions between financial metrics and ESG indicators, optimizing its parameters to predict future stock returns.

RAM-AI’s method involves creating monthly rebalanced longshort portfolios, where we go long on the top decile alpha prediction and short the bottom decile. We examine two distinct approaches:

1. Alpha Signal: This approach involves stock prediction without integrating ESG metrics. We train the model with around 450 features sourced from traditional data, including financial statements, market data, sentiment, and positioning.
2. Sustainable Alpha Signal: Here, ESG metrics are integrated into the stock prediction model. It’s trained with approximately 500 features, comprising the initial 450 traditional features plus an additional 50 derived from ESG data sources.

The training dataset consists of weekly data from 2011, focusing on an All Cap European universe, resulting in about 4 million observations. The analysis spans five years, from July 2017 to July 2022, and is based on out-of-sample simulations.

To delve into the nuanced characteristics of alpha signals, rather than outlining a concrete investment strategy, we intentionally bypass implementation constraints such as liquidity, market impact, transaction costs, financing costs, and borrowing availability. This provides a clear, unobstructed view of the alpha’s intrinsic characteristics.

When we integrate ESG factors into the alpha, an interesting pattern emerges. The decile spread return of the predicted alpha, enriched with ESG integration, aligns closely with the predicted alpha devoid of ESG elements. Yet, a distinct advantage surfaces – the ESG-integrated alpha exhibits reduced volatility and is characterized by diminished drawdown traits, enhancing its risk-adjusted performance.

Figure 2 showcases the cumulative log return accrued over the designated simulation period.

Source: FactSet, RAM AI, simulation from July 2017 to July 2022 (past performance is not a reliable indicator of future returns)

Use case key beneficiaries

This integration of sustainability factors with traditional investment metrics marks a significant stride in the investment field. ESG-integrated alpha, as revealed by this analysis, not only parallels but amplifies traditional alpha, characterized by diminished volatility and drawdowns.

☐ Relationship Managers

☒ Portfolio Managers

☒ Research teams, macroeconomists

☐ Control functions

☐ Support functions (HR, CFO, …)

☐ Other:

Supporting technology

The core supporting technology utilized by RAM-AI is a multi- layer perceptron. This architecture is particularly adept at handling the complex, diverse, and inherently non-linear interactions present in financial and ESG data. Its layered structure enables the model to process a large influx of information and autonomously identify feature interactions to predict outcomes like stock returns.