EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC 4.0

EXACT CONCEPT-LEVEL EXPLANATIONS FOR EARLY DETECTION OF FINANCIAL MISSTATEMENTS

JiaYing Chen, Xu Han, XuGuang Zhang, YongBin Yang

Eurasia Journal of Science and Technology · 05.04.2023 · 5 (2) · 34-48

DOI: 10.61784/ejst4162 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Eurasia Journal of Science and Technology, online erschienen am 05.04.2023. Verfasst von JiaYing Chen und 3 weiteren Autor:innen. Die Arbeit wurde bisher 17-mal zitiert (Crossref). Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Early warnings of material financial misstatements are useful only when auditors and regulators can understand why a firm is flagged. Machine learning models trained on financial statement data rank misstatement risk well, but their post-hoc explanations refer to individual line items, change when the model is retrained and often contradict established accounting evidence. This study proposes AGX, an accounting-grounded explainable model for the early detection of misstatement risk. AGX derives double-entry articulation features from consecutive financial statements, organises all predictors into ten accounting concepts, restricts every boosted tree to a single concept, imposes the monotone relations documented in the fraud-prediction literature and averages balanced ensembles on the log-odds scale. We prove that the resulting risk score decomposes exactly into concept contributions that coincide with group Shapley values and never contradict the imposed accounting priors. On the public AAER benchmark of 146,045 U.S. firm-years from 1990 to 2014, evaluated with a temporal design that uses only labels available at deployment, AGX matches the ranking accuracy of leading ensemble methods and detects the first year of misstatement episodes significantly better than LightGBM and the Explainable Boosting Machine (onset-year AUC of 0.722 versus 0.698 and 0.692). It is also the most robust model to measurement noise and missing prior-year data, and its explanations are exact, faithful, the most stable across retraining and free of theory violations.

Zitieren

APA

Chen, J., Han, X., Zhang, X., & Yang, Y. (2023). EXACT CONCEPT-LEVEL EXPLANATIONS FOR EARLY DETECTION OF FINANCIAL MISSTATEMENTS. Eurasia Journal of Science and Technology, 5(2), 34-48. https://doi.org/10.61784/ejst4162

Vancouver

Chen J, Han X, Zhang X, Yang Y. EXACT CONCEPT-LEVEL EXPLANATIONS FOR EARLY DETECTION OF FINANCIAL MISSTATEMENTS. Eurasia Journal of Science and Technology. 2023;5(2):34-48. doi:10.61784/ejst4162

BibTeX
@article{Chen2023,
  author = {Chen, JiaYing and Han, Xu and Zhang, XuGuang and Yang, YongBin},
  title = {EXACT CONCEPT-LEVEL EXPLANATIONS FOR EARLY DETECTION OF FINANCIAL MISSTATEMENTS},
  journal = {Eurasia Journal of Science and Technology},
  year = {2023},
  volume = {5},
  number = {2},
  pages = {34--48},
  doi = {10.61784/ejst4162}
}
Kostenlos starten →