Original title: Eliciting ESG Preferences for Reinforcement Learning-Based Portfolio Optimization
Giovanni Dispoto, Marcello Restelli, Carmine Ventre · September 2026
Investors want to balance profits with environmental, social, and governance goals, known as ESG. However, computerized trading programs usually rely on only one rating provider, even though rating agencies often disagree with each other. Picking numbers by hand to balance profits and these ethics scores is also difficult. To fix this, researchers built a system that learns to trade using scores from three different rating agencies at once. Instead of entering confusing math settings, a human simply compares two sample portfolios and chooses the better one. A computer model tracks these choices to figure out what the person truly values. Testing with simulated managers showed that location changes these choices. For example, European personas favored ethics scores over returns, while Texas-based personas preferred higher financial returns over ethics.
Imagine packing a suitcase for a vacation where you have limited space and the weather forecast changes daily. You need to decide how much to pack for rain versus sunshine, but checking the forecast once won't help because it varies. In this paper, investors face a similar challenge balancing financial returns and ESG goals, where ratings from different agencies conflict. The system helps by showing them sample portfolios and learning their priorities from their choices, like adjusting your packing based on which outfits you prefer as the forecast shifts.
Investors and portfolio managers must balance profit with sustainability, but ESG ratings vary by agency, making choices confusing. This system helps tailor portfolios to individual preferences, which change based on location, leading to more personalized and effective investment decisions.
The abstract doesn't state explicit limitations, but the evaluation uses simulated personas and historical market data, which may not fully reflect real-world decisions. It relies on three specific ESG agencies, so results might vary with other providers. The framework's adaptability to other contexts isn't fully established.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Eliciting ESG Preferences for Reinforcement Learning-Based Portfolio Optimization — Giovanni Dispoto, Marcello Restelli, Carmine Ventre
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.