In Palestine's tiny listed banking sector, voluntary disclosure is linked to better management efficiency and earnings but shows no uniform improvement in insolvency risk; AI methods can highlight unusual bank-year patterns and potential instability signals, offering a promising but unvalidated surveillance aid for supervisors.

Main Finding

Voluntary disclosure shows a conditional — not uniformly positive — relationship with banking stability in Palestinian listed banks: it is positively associated with management efficiency and earnings but is negatively (though insignificantly) associated with the Z-score, and its links to capital adequacy and asset quality depend on bank size, age, and leverage. Complementary machine-learning (ML) analyses (Random Forest, XGBoost, Isolation Forest) reveal nonlinear patterns and anomalous bank-year observations that traditional fixed-effects regressions miss. Overall, AI-augmented forensic accounting can meaningfully improve instability-risk screening and supervisory prioritization in institutionally fragile banking markets, but the ML results are exploratory proof-of-concept because of small sample size and lack of verified fraud labels.

Key Points

• Research question: how voluntary disclosure relates to banking stability, whether that relationship is conditional on bank characteristics, and whether ML can complement econometrics for forensic-risk screening.
• Data: whole population of seven Palestine Exchange–listed banks, 2019–2025 (49 bank-year observations).
• Voluntary disclosure index (VDI): 11-item index covering topics such as financial inclusion, cybersecurity, gender diversity, green financing, senior-management continuity.
• Traditional econometrics (fixed-effects panel):
  • VDI positively and significantly associated with management efficiency and earnings.
  • VDI negatively (insignificantly) associated with the Z-score.
  • Interaction effects (VDI × bank size, age, leverage) significant for capital adequacy and asset quality → conditional effects vary by bank characteristics.
• Machine learning and anomaly detection:
  • Supervised classifiers used: Logistic Regression, Random Forest, XGBoost. Ensemble tree methods outperformed Logistic Regression in cross-validation within this small sample.
  • Unsupervised anomaly detection (Isolation Forest) identified unusual bank-year records and instability-risk patterns not captured by linear models.
  • SHAP (SHapley Additive exPlanations) analysis improved interpretability of ML classifications and feature importance.
• Interpretation caveats: small sample (49 observations), no independent supervisory/fraud labels, interaction terms tested in separate specifications to avoid over-parameterization — ML findings are exploratory, not externally validated predictive claims.

Data & Methods

• Sample: full population of PEX-listed commercial and Islamic banks (7 banks), 2019–2025 (49 observations).
• Constructs:
  • Voluntary Disclosure Index (11 binary/indicator items).
  • Banking stability measures: Bank Z-score; CAMELS dimensions (capital adequacy, asset quality, management efficiency, earnings, liquidity, market-risk sensitivity).
• Econometric approach:
  • Fixed-effects panel regressions to estimate average conditional relationships and test interactions (VDI × size/age/leverage).
  • Interaction terms entered in separate models to reduce overfitting risk given small N.
• Machine-learning approach:
  • Supervised models: Logistic Regression baseline; Random Forest and XGBoost ensembles for nonlinear/interaction effects.
  • Unsupervised: Isolation Forest for anomaly detection (bank-year outliers).
  • Explainability: SHAP values to interpret feature contributions to model outputs.
  • Validation: cross-validation used for model comparison; emphasis on exploratory purposes due to limited sample and no ground-truth fraud labels.
• Limitations noted by authors:
  • Small cross-sectional N limits statistical power and generalizability.
  • Absence of independently verified fraud/enforcement labels prevents out-of-sample predictive validation.
  • Potential overfitting and model instability; results positioned as context-specific and hypothesis-generating.

Implications for AI Economics

• Surveillance efficiency and resource allocation:
  • AI-augmented screening (ensemble classifiers + anomaly detectors) can prioritize supervisory/investigative resources by flagging anomalous bank-years and conditional risk patterns that linear models miss — potentially lowering monitoring costs in resource-constrained regulators.
• Behavioral and incentive effects:
  • If regulators adopt AI screening, banks may alter disclosure strategies (e.g., more symbolic non-financial disclosure or strategic tailoring of reported metrics). Policy design should anticipate strategic responses and consider incentives that tie informative disclosure to market/regulatory benefits.
• Complementarity of methods:
  • Econometric inference and ML are complementary: fixed-effects models provide interpretable average effects and hypothesis tests; ML can uncover nonlinearities, interactions, and anomalies. AI economics research should combine both for robust policy insights.
• Interpretability matters for adoption:
  • Explainable AI tools (e.g., SHAP) increase transparency of ML-driven signals, which is critical for regulator trust, legal defensibility, and reducing false-positive/false-negative costs in supervision.
• Risks and fairness:
  • Small-sample or overfit ML models can produce misleading signals; deployment requires robust validation, out-of-time and cross-country testing, and careful threshold calibration to avoid undue penalties or missed risks.
• Research agenda for AI economics:
  • Scale up with larger pooled datasets across jurisdictions to obtain externally valid estimates and predictive models.
  • Integrate structured financial ratios with textual disclosure analysis (NLP) to capture linguistic concealment and tone as additional signals.
  • Pursue quasi-experimental or event-study designs to identify causal effects of disclosure changes on stability outcomes.
  • Explore welfare implications: quantify supervisory cost savings, reduction in crisis probability, and potential market-discipline effects from improved disclosure enforcement enabled by AI tools.

Summary recommendation: AI-augmented forensic accounting holds promise for improving banking-stability screening and supervisory prioritization in fragile markets, but meaningful policy adoption requires larger-sample validation, independent labels/enforcement data, careful interpretability, and attention to behavioral responses and fairness.