ESG ratings evaluate firms’ sustainability across environmental, social, and governance pillars (LSEG 2023), acting as key indicators of sustainable practices (Boffo and Patalano 2020). While high ESG ratings do not always align with higher stock returns (Krueger et al. 2024), they bring financial advantages, including reduced capital costs (Bams and Van der Kroft 2022) and increased firm value (Glaum et al. 2018; Guo et al. 2018; Busch et al. 2015). On a macro scale, ESG practices contribute to national economic growth (Caldecott et al. 2020).

The ESG rating industry has seen significant growth, utilizing data from company disclosures, media, and regulatory filings (Sinclair et al. 2018). However, ratings often face “black box” transparency issues (Del Vitto et al. 2023; Kölbel et al. 2021). Researchers have thus explored predictive models using financial data (D’Amato et al. 2021; Khan et al. 2022) and sentiment analysis from media (Jatowt and Färber 2022) to address these concerns.

This study enhances the model by Chowdhury (Abdullah et al. 2023), incorporating Green Innovation as a predictive factor for ESG ratings.

Green Innovation, or eco-innovation, refers to innovations in green products and processes that offer environmental benefits (Schiederig et al. 2011). This innovation enhances financial performance and supports ESG principles (Dicuonzo et al. 2022; Hossain et al. 2023). Firms that invest in Green Innovation are seen as forward-thinking, often engaging in extensive R&D (Woon Leong et al. 2021; Lee and Min 2015). According to Signaling Theory, firms use R&D to demonstrate long-term commitment, with Green R&D particularly seen as a targeted signal of sustainability (Erdem and Swait 1998; Raschke et al. 2022).

Green R&D influences all ESG components:

While models like Chowdhury (Abdullah et al. 2023) focus on financial indicators, this study fills the gap by examining Green R&D as a proxy for Green Innovation’s impact on ESG. Given its targeted nature, Green R&D is expected to better explain ESG rating variations than General R&D, with effects that may emerge over time (Gu et al. 2023).

H1: Green R&D better explains the variation in ESG ratings than General R&D.

Advancements in ESG rating prediction are driven by machine learning algorithms like Random Forest, XGBoost, and Neural Networks, which better handle complex, nonlinear data than traditional methods (D’Ecclesia et al. 2020; D’Amato et al. 2021; Bogun et al. 2021). For instance, Aue (Jatowt et al. 2022) predicted ESG ratings for over 3,000 U.S. companies by analyzing news articles, showcasing the potential of diverse data sources. However, traditional models, such as those used by Licari (Loiseau-Aslanidi et al. 2021), demonstrate limited explanatory power, as evidenced by their modest R2 values (31.13%), highlighting the need for more refined approaches.

This study addresses these challenges by incorporating Green Innovation, a key indicator of sustainability commitment, into Chowdhury’s (Abdullah et al. 2023) model to enhance ESG prediction accuracy. Thus, the hypothesis is:

H2: Integrating Green Innovation into advanced machine learning models improves ESG rating prediction accuracy compared to traditional statistical models.

While most ESG studies focus on developed economies, recent research underscores the importance of understanding ESG in emerging markets (Lozano and Martínez-Ferrero 2022). Theory suggests that organizations are shaped by their societal context, including regulations and NGO oversight (Dal Maso et al. 2016). Developed economies, with established ESG frameworks and external audits, generally support stronger ESG performance and transparency, reducing information asymmetry (Saini et al. 2023; Singhania and Saini 2021). In contrast, emerging economies face regulatory gaps, resource constraints, and limited political support for ESG initiatives (Singhania and Saini 2021).

Given these disparities, Green Innovation investments could represent a significant departure from average ESG practices in emerging economies, signaling a stronger commitment to sustainability and potentially having a more pronounced impact on ESG ratings. Building on the model of Chowdhury (Abdullah et al. 2023), this study examines how Green Innovation’s influence on ESG ratings varies by economic development, including macroeconomic factors to assess its moderating effect.

H3: The effect of Green Innovation on ESG ratings is stronger in emerging economies than in developed economies.

Note: In H2 and H3, “Green Innovation” will use the optimal proxy identified from H1.

ESG data: ESG (Environmental, Social, and Governance) ratings are a measure of a firm’s sustainability performance across these three key dimensions. ESG data from LSEG, covering over 90% of global market cap, evaluates 15,500+ companies based on 630+ metrics since 2002 (LSEG)]. Ratings are calculated from weighted Environmental (0.44), Social (0.31), and Governance (0.26) scores, using data from company reports and news sources. As Berg et al. (2019) support the use of single-source ESG data, this study adopts a unified dataset to ensure consistency.

Firm-level indicators: fundamental financial data reflects long-term operational performance and ESG relevance (Bogun et al. 2021). Building on Chowdhury (Abdullah et al. 2023), traditional financial ratios, alongside Green Innovation, form key predictive indicators, including size (TotalAssets), debt-to-equity ratio (DER), earnings per share (EPS), times interest earned (TIE), and lagged ESG (LESG). The study will focus on two measures of Green Innovation to understand its impact on ESG ratings:

Macroeconomic variables: Key macroeconomic indicators, based on Chowdhury (Abdullah et al. 2023), include GDP, GDP growth rate (GDPG), and unemployment rate (UNEM) to contextualize ESG performance within the broader economic environment. Inflation (INF), though initially considered, was excluded due to its low predictive value.

Key transformations to ensure consistency across variables include calculating Green R&D Intensity and General R&D Intensity, lagging variables for 1–3 years, and normalizing data with log transformations (ESG, TotalAssets and GDP).

By standardizing the data and performing these calculations, the dataset is prepared for analysis. The initial dataset, sourced from LSEG and World Bank, comprised 267,778 observations. Observations were removed for countries not classified as developed or emerging by MSCI (2024), for firms with fewer than five data points, and for missing values, resulting in 1,597 observations spanning 292 companies across 26 countries. Winsorization was applied to limit extreme values at the 5^th and 95^th percentiles. Table 1 summarizes observations before and after cleaning. The cleaned dataset includes observations from five industries: Utilities, Industrials, Consumer Discretionary, Basic Materials, and Technology.

Multicollinearity, which can inflate coefficient variances in regression, was evaluated using the Variance Inflation Factor (VIF). Most variables, especially those related to Green R&D Intensity, showed VIF values below 10, indicating minimal multicollinearity. However, lagged General R&D Intensity variables exceeded this threshold. Averaging the three lagged years (Lag1, Lag2, and Lag3) reduced the VIF to 6, supporting an averaged approach for these values to enhance model interpretability.

4.1.1. H1: Impact of green innovation on ESG ratings

The study tests the impact of Green R&D Intensity on ESG ratings, comparing it to General R&D Intensity. This is evaluated using OLS regression, with the base model containing only the Lagged ESG variable, as highlighted by Chowdhury (Abdullah et al. 2023). Additional factors are incrementally added using stepwise regression, incorporating industry and year-fixed effects. The model evaluates Green Innovation’s effect on ESG ratings, including immediate and lagged influences:

ESG_it = β₀ + β₁Green Innovation_it + β₂Lag1_it + β₃Lag2_it + β₄Lag3_itβ₅Control Variables_it + ε_it

Note: Green Innovation combines Green and General R&D Intensity, analyzed separately to capture their individual impacts. Control variables refer to the firm-level and macroeconomic indicators outlined in subsection 3.1. The goal is to identify whether Green or General R&D Intensity more effectively explains variations in ESG ratings.

Using OLS regression, this analysis assesses Green R&D Intensity and General R&D Intensity as predictors of ESG ratings, with Lagged ESG as the core predictor (following Chowdhury (Abdullah et al. 2023)). Various model specifications introduce fixed effects, control variables, and the R&D Intensity variables, as shown in Table 2.

Conclusion: the analysis finds that Green R&D Intensity significantly explains ESG ratings and does so more effectively than General R&D Intensity.

This section explores how adding Green R&D Intensity as a proxy for Green Innovation influences ESG rating predictions using machine learning models. Following H1 findings, Green R&D Intensity was selected as the primary innovation metric. The Random Forest (RF) model serves as the main approach, with Artificial Neural Networks (ANN) and eXtreme Gradient Boosting (XGBoost) as robustness checks.

Conclusion: incorporating Green Innovation enhances ESG prediction accuracy, particularly when no prior ESG data is available.

Conclusion: results show no significant difference in the impact of Green R&D Intensity (Green Innovation proxy) on ESG ratings between developed and emerging economies, thus not supporting Hypothesis 3.

This study examines the impact of Green Innovation on ESG ratings, its role in enhancing machine learning predictive accuracy, and its differential effects across developed and emerging economies.

Analyzing data from 292 firms across 26 countries, Green R&D Intensity emerges as a more significant predictor of ESG ratings than General R&D Intensity, with consistent results across industries. The lagged effects of Green Innovation indicate a delayed but positive impact, emphasizing the need for a long-term perspective in sustainable investment.

Aligned with prior research (D’Ecclesia et al. 2020; Del Vitto et al. 2023; Bogun et al. 2021), machine learning models, notably Random Forest and XG- Boost, outperform traditional models in predicting ESG ratings due to their ability to capture complex relationships within ESG data. Including Green Innovation reduces prediction errors modestly by 2% overall and by 11.5% for firms without prior ESG ratings, underscoring its value for newer or transitioning firms.

Contrary to expectations, Green Innovation’s influence on ESG ratings does not differ significantly across economic contexts, with macroeconomic factors like LogGDP, GDPG, and UNEM playing a more substantial role than economic classification. This enhances the predictive power of machine learning models across various economic contexts.

The study contributes to ESG and Green Innovation literature in three ways: confirming Green R&D Intensity as a superior ESG predictor, demonstrating its value in reducing prediction errors in machine learning models, and revealing that broader economic classification has minimal impact on ESG predictions.

For firms, emphasizing Green R&D spending improves ESG ratings, aligning with studies like Bams and Van der Kroft (2022) and Yu (Guo et al. 2018), which link high ESG ratings to reduced capital costs and improved firm value. Investors benefit from including Green R&D Intensity in models, potentially lowering investment risk and fostering long-term gains (Servaes et al. 2017). The absence of significant differences between developed and emerging economies facilitates model use across diverse economic contexts, promoting broader ESG investment. Policymakers can support green innovation investments, which align with national growth and sustainability goals as suggested by Zhou (Caldecott et al. 2020).

This study has several limitations. First, the ESG data is solely from the LSEG database. Although Berg et al. (2019) note the advantages of single-source insights, significant rating differences across agencies may impact generalizability (Krueger et al. 2024). Additionally, this study’s approach of excluding observations with missing values enhances data quality but reduces dataset size, potentially omitting valuable insights. An alternative approach might involve data enrichment to expand and strengthen the dataset. The exclusion of firms with fewer than five observations further limits the sample, especially due to the limited data on Green R&D.

Future studies could address several areas to expand upon these findings. Firstly, using ESG data from multiple agencies could assess the robustness of results across different rating sources. Additionally, the significant negative impact of GDP on ESG ratings, a novel finding here, invites further exploration into how macroeconomic factors shape ESG practices. Selection bias, as discussed by Barkemeyer (Revelli et al. 2023) and Boffo and Patalano (2020), remains an underexplored aspect in ESG research and warrants focused investigation, particularly across different economic contexts. Furthermore, examining the stages and maturation periods of Green Innovation investments, as indicated by Song (Gu et al. (2023), could provide valuable insights into the long-term impacts of sustainable investments.

Potential research directions to build on these findings include analyzing ESG data from multiple rating agencies to validate these insights across datasets, investigating the GDP’s negative effect on ESG ratings to uncover broader macroeconomic influences, and exploring the specific investment stages and timeframes of Green Innovation. Such studies could enhance understanding of ESG ratings and support informed decision-making for investors and policymakers in sustainable investing.