Multimodal Early Warning of Corporate Financial Distress Using LSTM and BERT from Macroeconomic and Firm Perspectives
Mathematics · 15.11.2026 · 14 (22) · 4198
DOI: 10.3390/math14224198 ↗Kurz gefasst
Fachartikel in der Zeitschrift Mathematics, online erschienen am 15.11.2026. Verfasst von Zhenhua He und 3 weiteren Autor:innen. Die Arbeit nennt 46 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Hunan Provincial Natural Science Foundation.
Abstract (Originalsprache)
Macroeconomic uncertainty motivates financial distress models that combine firm financial indicators with numerical and textual macroeconomic information. Using nonfinancial A-share firms during 2008–2024, this study integrates long short-term memory (LSTM) and frozen bidirectional encoder representations from transformers (BERT) representations through attention within a Cox survival framework. The primary firm-disjoint test C-index is 0.9508 (95% CI 0.9050–0.9869). Adding text to numerical fusion increases mean inverse probability of censoring weighted (IPCW) average precision (AP) from 0.7091 to 0.7715 and F1 from 0.5617 to 0.6335. However, the financial-only network has higher primary-test C-index, AP and F1; Cox and XGBoost are also competitive. Across five company-partition sensitivity runs, the multimodal model has the highest mean AP and F1 among six configurations (0.8906 and 0.8166). With histories fixed to three records, its AP and F1 are 0.6559 and 0.4547, the highest values among four evaluated models. An additional calendar-landmark analysis fits earlier outcome windows ending by 2018, selects models using 2019–2021 outcomes, and tests 2022–2024 outcomes; multimodal three-year AP and F1 are 0.1205 and 0.1287, respectively. The main endpoint-dependent histories support retrospective discrimination, whereas the calendar-landmark analysis evaluates subsequent events from fixed prediction dates. The combined evidence describes design-dependent performance and potential early-warning value, with limited events and unverified historical publication dates constraining deployment claims.
Förderung
- Hunan Provincial Natural Science Foundation
Zitieren
APA
He, Z., Qi, Z., Hu, H., & Chen, L. (2026). Multimodal Early Warning of Corporate Financial Distress Using LSTM and BERT from Macroeconomic and Firm Perspectives. Mathematics, 14(22), 4198. https://doi.org/10.3390/math14224198
Vancouver
He Z, Qi Z, Hu H, Chen L. Multimodal Early Warning of Corporate Financial Distress Using LSTM and BERT from Macroeconomic and Firm Perspectives. Mathematics. 2026;14(22):4198. doi:10.3390/math14224198
BibTeX
@article{He2026,
author = {He, Zhenhua and Qi, Zijian and Hu, Huijiao and Chen, Liming},
title = {Multimodal Early Warning of Corporate Financial Distress Using LSTM and BERT from Macroeconomic and Firm Perspectives},
journal = {Mathematics},
year = {2026},
volume = {14},
number = {22},
pages = {4198},
doi = {10.3390/math14224198}
}
EUVIMED