ESG-Weighted Scoring

• ESG-weighted scoring is a quantitative method that aggregates non-financial performance metrics across environmental, social, and governance pillars.
• It employs fixed expert-defined or machine-learned weights with normalization and penalization techniques to ensure data comparability and mitigate greenwashing.
• This framework is applied in portfolio optimization, asset pricing, and risk management, enhancing predictive outcomes and aligning investments with sustainability metrics.

Environmental, Social, and Governance (ESG)-Weighted Scoring is a family of quantitative methodologies that aggregate multiple non-financial performance metrics—typically relating to a company’s environmental impact, social responsibility, and governance practices—into a single or composite set of scores. These scores serve as inputs for financial modeling, portfolio construction, risk measurement, or asset pricing. Advanced variants expand the ESG construct to accommodate issues such as missing data, machine-learned category weights, penalization of cross-pillar misalignment, or dynamic weighting in response to risk measures.

1. Foundations and Basic Mathematical Structure

The core of ESG-weighted scoring is the formulation of a composite index as a (possibly weighted) linear or nonlinear aggregation of component scores. Standard practice, as exemplified by major ESG data vendors and modeling frameworks, is to partition available information into three canonical "pillars" (Environmental: E, Social: S, Governance: G), further decomposed into subcategories (e.g., carbon emissions, labor practices, executive remuneration), and then aggregate these into firm-level scores.

A generic additive ESG score takes the form: $ESG_{\rm total} = \sum_{i=1}^{n}w_i c_i,$ where $c_i \in [0,10]$ (or normalized interval) is the score for subcategory $i$ and $w_i \geq 0$ are weights constrained such that $\sum_i w_i = 1$ (Garrido-Merchán et al., 2023). Category weights are often fixed—determined a priori by providers or domain experts—but frameworks also exist for learning or optimizing $w_i$ based on empirical or theoretical criteria (Sahin et al., 2021, Patel et al., 2023, Soares, 2024).

2. Category Weight Determination and Aggregation Schemes

Fixed and Expert-Defined Weights

Most industry-standard ESG frameworks use pre-defined, expert- or provider-set weights for subcategories. For example, the 14-category aggregation in (Garrido-Merchán et al., 2023) uses provider-locked weights $w_i$, normalized to sum to one. The optimizer in such cases operates on portfolio allocation ($x$), not on the weights themselves.

Data-Driven and Machine-Learned Weights

Recent research proposes learning the aggregation weights $w_i$ from data, leveraging statistical or machine-learning techniques:

• Random Forest Feature Importance: In (Patel et al., 2023), the subcategory weights are extracted from a RandomForestRegressor trained to match public ESG ratings, with the final $w_i$ proportional to the average reduction in prediction error contributed by each feature.
• Item Response Theory (IRT): (Soares, 2024) applies Rasch modeling to binary or ordinal observations (e.g., monthly sentiment), mapping item difficulties $b_i$ to weights $w_i$ via softmax or affine normalization.
• Risk-Linked Optimization: (Sahin et al., 2021) determines 4-pillar (E/S/G/Missing) weights by maximizing empirical dependence (Kendall’s $\tau$) between weighted ESG scores and financial risk measures, subject to convexity and regulatory constraints.
• Rule Ensemble Learning: Some nonlinear machine learning strategies for high-dimensional ESG features rely on adaptive weights over rule-based “experts,” updating confidences as data accumulate (Franco et al., 2020).

Multi-Attribute and MCDM Formulations

Multi-Criteria Decision Making (MCDM) models (Hassani et al., 2024) implement explicit weighted sums across normalized scores for each criterion, often using user- or expert-chosen starting weights that are rescaled to the simplex.

3. Normalization, Scaling, and Penalization Techniques

Normalization Conventions

To ensure comparability and proper weighting, most frameworks normalize subcategory scores to a common range (e.g., [0,1] or [0,10]) and impose convexity on the weights ($\sum w_i=1$) (Garrido-Merchán et al., 2023, Fischbach et al., 2022). In practice, this ensures scale-invariance and guards against distortion when subcategory counts differ or when subpillar importance varies regionally or temporally.

Penalization and Sensitivity Modifications

To address strategic manipulation ("greenwashing" or "crosswashing"), models can incorporate penalization mechanisms or alignment-based adjustments:

• Crosswashing Deterioration Penalty: (Hassani et al., 2024) introduces a “deterioration factor” reflecting misalignment between investment and true impact, which is subtracted from the baseline score.
• Penalty via Log Functions or Nonlinear Aggregation: (Garrido-Merchán et al., 2023) mentions the option to log-penalize portfolios with low ESG, although this is not applied in their core experiments.

4. Applications in Portfolio Optimization and Asset Pricing

Portfolio Optimization with ESG Constraints or Objectives

• Linear ESG Constraints: ESG-weighted scores frequently influence portfolio allocations by serving as constraints, e.g., requiring the portfolio average ESG score to exceed a threshold (Wu et al., 16 Feb 2026).
• Objective Function Integration: Some frameworks combine normalized ESG scores with financial measures (e.g., Sharpe ratio) in additive or convex-combination objective functions, allowing trade-offs between risk, return, and ESG (Garrido-Merchán et al., 2023, Lauria et al., 2022, Torri et al., 2023). The key parameter ($\lambda$ or similar) modulates the investor's ESG affinity or risk preference.
• Black-Litterman and Bayesian Integrations: Stein shrinkage of equilibrium return vectors toward ESG scores is used in a Black-Litterman framework (biased prior mean), with the weight $\lambda$ controlling the bias strength (Alpern et al., 26 Nov 2025).

Dynamic Asset Pricing

• ESG-Adjusted Asset Prices: The price process can be directly rescaled by ESG-weighted multipliers ($S^{\rm ESG}_t = S_t(1 + y_{\rm ESG} Z_t)$), thus modifying the payoff structure and option values (Rachev et al., 2023).

Deep Reinforcement Learning and ESG-Based Regulation

• Shaped Rewards: Regulated environments can grant bonuses or impose taxes proportional to a portfolio's weighted ESG score relative to a benchmark (Garrido-Merchán et al., 2023). This incentive mechanism causes DRL agents to shift allocations toward higher-ESG names without sacrificing expected return.

5. Extended and Specialized ESG Weighting Frameworks

Missing Data and the "M-Pillar"

The ESGM approach of Sahin et al. (Sahin et al., 2021) augments the classic E/S/G structure with a Missing (M) pillar that quantitatively accounts for the extent of non-disclosure in subcategory data. The relative weight of the M pillar is then learned via optimization, subject to constraints that it not surpass E/S/G weights. This construction penalizes incomplete reporting, rewards increased disclosure, and can empirically improve the alignment with realized risk.

Social-Network and Sentiment-Based ESG Scoring

Sentiment-driven ESG scoring processes large-scale social media and news corpora, applies NLP-based classification into ESG subcategories, and then averages sentiment labels within each domain (Fischbach et al., 2022, Patel et al., 2023). Aggregate weights may be set uniformly, user-specified, or determined from model-based feature importances.

Machine Learning on High-Dimensional ESG Features

Rule-based ensemble learning partitions the ESG-feature hypercube into statistically significant regions associated with future excess returns. Final scores for assets are derived from the sign or aggregation of predicted outcomes, and portfolio construction proceeds via screening “high-ESG” signals (S=+1), with empirically demonstrated outperformance over best-in-class benchmarks (Franco et al., 2020).

6. Empirical Implications and Performance Evaluation

Empirical evaluation of ESG-weighted scoring schemes covers fit to external benchmarks, stability under out-of-sample reweighting, and impact on risk-adjusted returns:

• Validation: Holdout mean-absolute errors and correlation coefficients quantify prediction fidelity (e.g., MAE 13.4%, r=0.26 for social-sentiment-based scores against S&P Global) (Patel et al., 2023).
• Portfolio Outcomes: Choices of weighting scheme and inclusion of ESG criteria alter the efficient frontier, capital market line, and realized Sharpe ratio. For example, adding ESG constraints generally raises portfolio ESG scores and may lower risk or mean returns depending on the sign and magnitude of the ESG-return relationship (Lauria et al., 2022, Azzone et al., 2024, Wu et al., 16 Feb 2026).
• Risk Measures: ESG-coherent risk measures allow continuous interpolation between pure risk-return and ESG-centric ranking, with empirical studies showing stability of risk-based asset rankings for small ESG weightings but increasing sensitivity as ESG weights grow (Torri et al., 2023).
• Sector and Region Sensitivity: The optimal weights for subcategories, pillars, or missingness vary in time and across sectors, emphasizing the need for re-estimation and contextualization (Sahin et al., 2021).

Scoring Family	Normalization	Weight Determination
Vendor-style linear (e.g., (Garrido-Merchán et al., 2023))	[0,10]; simplex	Provider/expert fixed
Machine-learned RF (e.g., (Patel et al., 2023))	[0,1] or [0,100]	Feature importance (data-driven)
Risk-coherent (e.g., (Torri et al., 2023))	[0,1] or variable	User-set $\lambda$, convex combination
MCDM (e.g., (Hassani et al., 2024))	Div by max per criterion	User/expert; optionally optimized
Missing-data pillar (Sahin et al., 2021)	[0,100] per pillar	Constraint-optimized (maximize risk-score relation)
Rasch/IRT-based (Soares, 2024)	[0,1]; softmax	Model-based from IRT difficulty estimates
DRL / regulated (Garrido-Merchán et al., 2023)	[0,10] portfolio ESG	Not scores per se, but manager-tuned or policy-driven

7. Future Directions and Challenges

ESG-weighted scoring remains an active research area, with several open directions:

• Optimization of Subcategory Weights: Most current frameworks use fixed weights; adaptive, fully empirical or multi-objective tuning inside portfolio optimizers is a frontier area (Garrido-Merchán et al., 2023).
• Data Completeness and Quality: Addressing missing or systematically manipulated data (crosswashing, greenwashing) through explicit penalization or the introduction of robustness pillars is gaining traction (Sahin et al., 2021, Hassani et al., 2024).
• Nonlinear Aggregation and Interaction Effects: High-dimensional ESG metrics frequently interact in nonlinear ways with financial performance; ensemble and rule-based ML models offer one route for exploitation, but their transparency and stability remain challenging (Franco et al., 2020).
• Integration with Asset Pricing and Risk Management: Embedding ESG measures directly into the machinery of asset pricing, risk quantification, and derivative valuation reveals effects such as differentiated equilibrium risk premia and new "shadow" riskless rates (Rachev et al., 2023, Lauria et al., 2022, Azzone et al., 2024).

Future research will likely emphasize robust, context-sensitive, and economically interpretable ESG-weighted scoring protocols—enabling investors and regulators to balance financial efficacy, sustainability impact, and data integrity in an increasingly complex landscape.