隨著 ESG 績效日益成為投資決策的關鍵因素，提升 ESG 分數預測的準確性愈顯重 要。現有研究多依賴結構化資料(如財務指標與公司特徵)或單一來源的文本資料，預 測效果常受限於資料範疇。本研究提出一項創新框架，整合來自 10-K 報告、永續報告 與商業新聞的多源文本資料，以提升 ESG 總分及各子構面的預測效能。我們深入探討 不同資料來源的結合如何改善模型表現，並進一步分析預測的時序特性與產業間差異。 實驗結果顯示，10-K 報告是預測 ESG 分數最穩定和具代表性的文本來源，MAE 為 5.005， R2為 0.688。其次是永續報告與 10-K 報告的組合，MAE 為 5.072，R2為 0.687。整合所 有文本來源的組合也提升了預測表現，MAE 為 5.214，R2為 0.680。這些結果突顯了雖 然整合多種文本來源能提升 ESG 分數預測準確性，但 10-K 報告仍然是最一致且可靠的 預測來源。此外，時序分析顯示最新的數據提供了最佳的預測結果。產業分析則強調， 材料產業在 ESG 分數預測中表現最佳，R2為 0.734。本研究為投資者和企業管理者提供 了實務指導，並且是首個將多來源文本數據應用於 ESG 分數預測的研究，為該領域提 供了創新的方法學和研究視角。;As ESG performance increasingly becomes a critical factor in investment decisions, improving the accuracy of ESG score predictions has become more important. Existing studies often rely on structured data (such as financial indicators and company characteristics) or single-source textual data, which limits prediction performance due to the scope of the data. This study proposes an innovative framework that integrates multi-source textual data from 10- K reports, sustainability reports, and business news to enhance the prediction performance of overall ESG scores and their sub-dimensions. We explore how the combination of different data sources improves model performance and further analyze the temporal characteristics of predictions and industry differences. The experimental results show that 10-K Reports emerge as the most stable and representative textual source for ESG score prediction, with a MAE of 5.005 and an R2 of 0.688. Following this, the combination of Sustainability Reports and 10-K Reports comes in second, with MAE of 5.072 and an R2 of 0.687. The combination of all text sources also improves performance, but with MAE of 5.214 and R2 of 0.680. These results highlight that while integrating multiple text sources can improve ESG score predictions, 10-K Reports remain the most consistent and reliable source for accurate predictions. Additionally, temporal analysis revealed that the most recent data provided the best prediction results. The industry-level analysis highlighted the Materials sector as the top performer, achieving an R2 of 0.734 for ESG score prediction. This research provides practical guidance for investors and corporate managers. Moreover, this study is the first to apply multi-source textual data to ESG score prediction, offering a novel methodology and research perspective for the field.
