A Multimodal Dataset of Financial Disclosures, MD&A, and Audit Opinions with Next-Year Bankruptcy Labels

We publicly release a multimodal dataset derived from 10-K annual reports to support research on next-year bankruptcy prediction. For each report, we collected all reported financial figures, the corresponding Management Discussion & Analysis (MD&A), and the Audit Opinion text, along with a bankruptcy label indicating whether the company filed for bankruptcy in the year following the report’s release.
This dataset is designed to encourage research in this challenging area by presenting several key difficulties, including:
Extreme class imbalanceMultimodality — integration of both tabular and textual dataMultisource heterogeneity — signals from the three sources may align or even contradictNLP-related challenges — long documents with substantial portions of text that may be irrelevant to the bankruptcy outcomeThis dataset accompanies our published paper in 6th ACM International Conference on AI in Finance with title "A Multimodal Alignment-Based Anomaly Detection Method for Bankruptcy Prediction"

Authors:
Andreas Sideras
Konstantinos Bougiatiotis
Elias Zavitsanos
Georgios Paliouras
George Vouros