{"paper_id":"2401.12652","title":"From Numbers to Words: Multi-Modal Bankruptcy Prediction Using the ECL Dataset","authors":["Henri Arno","Klaas Mulier","Joke Baeck","Thomas Demeester"],"paper_date":"2024-01-23","paper_type":"dataset-benchmark","paper_url":"https://arxiv.org/pdf/2401.12652v1","page_url":"https://thequant.space/flowcharts/from-numbers-to-words-multi-modal-bankruptcy-prediction-usi/","scores":{"empirical_rigor":7.5,"math_complexity":3,"quadrant":"Street Traders"},"proposes":{"status":"available","statement":"The paper introduces the ECL multi‑modal dataset of 10K filings with binary bankruptcy labels and provides baseline classical and neural prediction models.","quote":"In this paper, we present ECL, a novel multi‑modal dataset containing the textual and numerical data from corporate 10K filings and associated binary bankruptcy labels. Furthermore, we develop and critically evaluate several classical and neural bankruptcy prediction models using this dataset.","summary":"In this paper, we present ECL, a novel multi-modal dataset containing the textual and numerical data from corporate 10K filings and associated binary bankruptcy labels. Furthermore, we develop and critically evaluate several classical and neural bankruptcy prediction models using this dataset. Our…","strategy":null},"assumptions":{"status":"available","items":[{"assumption":"The binary bankruptcy label is sufficient for training prediction models.","type":"data","stated":"explicit","quote":"that the binary bankruptcy prediction target does not enable our models to distinguish next year bankruptcy from an unhealthy financial situation resulting in bankruptcy in later years."},{"assumption":"Evaluation metrics such as ROC‑AUC, AP, CAP ratio and recall@100 adequately capture model performance.","type":"statistical","stated":"explicit","quote":"We report the area under the receiver operating curve (ROC‑AUC), the average precision (AP), the cumulative accuracy profile ratio (CAP ratio) and the recall@100 for each classifier."},{"assumption":"The 10K filings contain all relevant textual and numerical information for bankruptcy prediction.","type":"data","stated":"implied","quote":"The most widely consulted SEC filing is the Form 10K, which is reported annually and contains detailed information on a company’s past fiscal year."}]},"datasets":{"status":"available","items":[{"name":"ECL","provider":"EDGAR corpus, Compustat, LoPucki Bankruptcy Research Databas","kind":"benchmark","access":"public","period":null,"universe":null,"quote":"It is a unique compilation of three existing data sources: the EDGAR‑corpus (Loukas et al., 2021), Compustat","known_id":"compustat"}],"known":[{"id":"edgar","name":"SEC EDGAR","provider":"U.S. Securities and Exchange Commission","kind":"text-news","access":"public","quote":"a is collected from (1) the EDGAR-corpus (Loukas et al., 2021), the numerical finan-cial data is gathered from (2) CompuStat2 while (3) the LoPucki BRD provides the labels for the bankruptcy prediction task.","mentions":7,"used_in":5},{"id":"compustat","name":"Compustat","provider":"S&P Global","kind":"fundamentals","access":"commercial","quote":"Next, we add the structured, financial information, reported in item 8 of a 10K, to the dataset by linking the collected 10K records from the previous step to CompuStat records.","mentions":9,"used_in":4}],"sample":{"start":null,"end":null,"market":null,"asset_class":null,"frequency":null},"data_resolution":"weekly-monthly"},"transaction_costs":{"status":"not_applicable","modeled":"n/a","headline_net_of_costs":null,"detail":null,"quote":null,"source":"front_matter","rule":"R3","candidates":{"front_matter":"n/a","evidence_card":"n/a","evidence_card_quoted":false},"meaning":"the paper reports no trading or execution result"},"results":{"status":"available","evaluation_type":"forecast-accuracy","findings":[{"finding":"Textual and numerical modalities provide complementary information, and the best performance is achieved by an ensemble of both.","kind":"predictive-accuracy","quote":"The best results are attained when we combine the predictions of the textual and numerical models in an ensemble."},{"finding":"When trained on binary labels, models cannot distinguish 10K records filed the year before bankruptcy from those of financially unhealthy firms that do not file within one year.","kind":"negative","quote":"We did observe that our models, trained on binary bankruptcy labels, cannot distinguish between 10K records filed in the year preceding bankruptcy and records filed by financially unhealthy companies that are close to bankruptcy but not within one year."},{"finding":"Zero‑shot GPT‑3.5 predictions are significantly worse than a simple keyword‑based TF‑IDF model.","kind":"performance","quote":"the zero‑shot prediction results of GPT‑3.5 are significantly worse than the results of a simple keyword‑based TF‑IDF model."},{"finding":"Numerical‑only model achieves a CAP ratio of 0.896, the highest among single‑modality models.","kind":"performance","quote":"Table 1: The results of the numerical, textual and combined models, tuned on ROC‑AUC, evaluated on the test set."}],"reported_performance":null,"out_of_sample":{"method":"holdout","headline_out_of_sample":null}},"independent_evidence":{"status":"none_found","supports":[],"contradicts":[],"qualifies":[],"compared_with":23,"citations":{"cited_by_in_corpus":1,"by_independent_authors":1,"how_cited":{"used_as_baseline":0,"uses_method_or_data":0,"extends":0,"discusses_findings":0,"mentions":1,"unclassified":0},"independent_papers":[],"scope":"Citations between papers in this corpus only; citations from outside it are not counted."}},"implementation":{"status":"available","public":true,"code_url":"https://github.com/henriarnoUG/ECL","note":null},"replication":{"status":"none_found","card":null,"note":"No replication of this paper has been run by thequant.space yet."},"extraction":{"text_source":"fulltext","model":"openai/gpt-oss-120b","version":1,"extracted_at":"2026-11-14T18:15:48Z"},"provenance":{"claims":"Everything here is what the authors state in the paper. Nothing was re-run or reproduced by the platform.","quotes":"Assumptions, datasets, cost detail and findings were extracted by a language model. Every quote is verbatim paper text. Where an item has `support`, \"full\" means every figure in it appears in the quoted sentences and \"partial\" means a figure is stated elsewhere in the paper; items with a figure found nowhere in the paper were discarded.","datasets_known":"datasets.known lists well-known vendors and public datasets by a stable id, found by a dictionary over the full text where the sentence says the data is used; it is independent of the model and covers only names in the dictionary.","flags":"transaction_costs.modeled, results.evaluation_type and results.out_of_sample.method are platform classifications (structured-decision models); see https://thequant.space/score-guide/#penalty-flags","comparisons":"supports / contradicts / qualifies are a language model's reading of two papers' quoted findings, restricted to closely related papers that share no author. They are pointers to read, not verified replications, and coverage is partial.","implementation":"code_url is a repository the paper itself presents as its own (rule-scored, ambiguous cases judged by a model), checked for liveness when it was recorded."},"docs":"https://thequant.space/agents/"}