Organizations: DIMA, Polytechnic University of Marche, Italy · DII, Polytechnic University of Marche, Italy · CHIMOMO, University of Modena and Reggio Emilia, Italy
Environmental, Social and Governance (ESG) scores are increasingly used to evaluate corporate sustainability, supporting investment decisions, ensuring regulatory compliance and assessing risk. However, current ESG evaluations largely rely on proprietary methodologies that often produce inconsistent ratings. This motivates the development of data-driven prediction frameworks. In this paper, we present ESGPred, a modular framework for predicting ESG scores by combining domain-specific language modeling with graph-based learning over a network that models the relationships among the companies being evaluated. In addition to presenting the rationale behind ESGPred and detailing its ESG score prediction process, we show that ESGPred can effectively predict both overall ESG ratings and their individual Environmental, Social, and Governance components. We evaluated ESGPred using a dataset of 5,266 sustainability reports published by 2,134 publicly listed companies between 2020 and 2023. In this evaluation, we used ESG ratings provided by LSEG Data & Analytics as ground truth. Our experimental results demonstrated that integrating graph-based relational learning improves prediction performance over purely text-based baselines.
Figures & tables
Figure 1: Overview of the ESGPred’s workflow
Figure 2: Sustainability graphs obtained by applying Semantic Report Similarity Graph, Company Description Similarity Graph and Sector-Based Graph on the same nodes
Country
Count
Country
Count
USA
1,201
Luxembourg
3
England
234
Spain
2
Australia
212
Belgium
2
Canada
205
Chile
2
Sweden
52
Hong Kong
2
Finland
23
Cayman Islands
2
Table 1: Overall geographical distribution of the companies to which sustainability reports refer
Name
Value
Total number of sustainability reports
5,266
Total number of reports relating to 2020
1,133
Total number of reports relating to 2021
1,459
Total number of reports relating to 2022
1,665
Total number of reports relating to 2023
1,009
Total number of reports with score A
785
Table 2: Some statistics of the dataset R of reports used in our experiments
Category
Model / Architecture
Accuracy
Precision
Recall
F1-Score
Baselines
Vanilla FinBERT
0.72
0.69
0.60
0.63
Fine-Tuned FinBERT
0.77
0.72
0.74
0.73
Fine-Tuned Transformer
0.76
0.74
0.65
0.68
ESGPred
Description-Based Graph
Graph Transformer
0.85
0.83
0.85
0.84
Table 3: Comparison between baselines and ESGPred with different graph-based configurations
Category
Model / Architecture
Accuracy
Precision
Recall
F1-Score
Jina fine-tuned embeddings
Description-Based Graph
Graph Transformer
0.79
0.76
0.75
0.76
GraphSAGE
0.79
0.73
0.75
0 .74
GAT
0.70
0.62
0.69
0.64
GCN
0.69
0.61
0.68
0.61
Semantic-Based Graph
Graph Transformer
0.81
0.76
0.79
0.77
Table 4: Results obtained using Jina-based embeddings instead of FinBERT-based embeddings
Dimension
Model / Architecture
Accuracy
Precision
Recall
F1-Score
Environmental
ESGPred (Desc.-Based Graph+Graph Transf.)
0.87
0.87
0.88
0.88
Fine-Tuned FinBERT (baseline)
0.70
0.74
0.69
0.71
Social
ESGPred (Desc.-Based Graph+Graph Transf.)
0.85
0.82
0.84
0.83
Fine-Tuned FinBERT (baseline)
0.68
0.67
0.61
0.64
Table 5: Comparative results of ESGPred and the Fine-Tuned FinBERT for individual Environmental, Social, and Governance dimensions
Graph Type
Split Method
Accuracy
Precision
Recall
F1-Score
Description-Based Graph
Fixed (5 splits)
0.82
0.82
0.83
0.82
Fixed (10 splits)
0.83
0.81
0.84
0.82
Fixed (15 splits)
0.84
0.83
0.83
0.83
Dynamic (Used by ESGPred)
0.85
0.83
0.85
0.84
Semantic-Based Graph
Table 6: Impact of report segmentation strategies on the performance of ESGPred
Graph Type
Threshold
Accuracy
Precision
Recall
F1-Score
Description-Based Graph
τJ
0.29 (Jaccard)
0.83
0.82
0.83
0.82
0.30 (Jaccard)
0.85
0.83
0.85
0.84
0.31 (Jaccard)
0.84
0.80
0.85
0.82
0.33 (Jaccard)
0.83
0.79
0.85
0.82
Semantic-Based Graph
τcos
Table 7: Impact of graph construction thresholds on the performance of ESGPred
Eindhoven University of Technology, Eindhoven · Jožef Stefan Institute and Postgraduate School, Ljubljana · Faculty of Economics, University of Ljubljana +1