Researchers from Palestine Technical University – Kadoorie, Sulaiman AlRajhi University, and University College of Bahrain examined every bank listed on the Palestine Exchange between 2019 and 2025. Their panel covered seven banks and 49 bank-year observations, spanning both conventional and Islamic institutions. The team built an eleven-item voluntary disclosure index from the banks’ annual reports, measuring how much each institution chose to reveal beyond regulatory minimums about topics like cybersecurity, gender diversity, green financing, and management continuity.

The econometric results surprised the team. Voluntary disclosure was not uniformly linked to financial stability. Instead, its relationship with the Bank Z-score and CAMELS indicators depended on bank size, age, and leverage, with statistically significant interaction effects that shifted direction across the observed range of each moderator. Disclosure was positively associated with the earnings proxy alone, while its associations with capital adequacy, asset quality, management efficiency, liquidity, and market-risk sensitivity were statistically insignificant on their own.

That conditional structure is precisely where linear regression runs out of steam. Fixed-effects models require the analyst to specify in advance which interactions to test, and they estimate average relationships across the sector rather than flagging individual outliers. The researchers therefore added an exploratory machine-learning layer, deploying Logistic Regression as a benchmark alongside Random Forest, XGBoost, and Isolation Forest for anomaly detection.

The ensemble models outperformed Logistic Regression in stratified five-fold cross-validation. XGBoost achieved 89.4 percent accuracy, compared with 74.5 percent for the logistic benchmark. Random Forest landed in between at 85.7 percent. The team converted the continuous Bank Z-score into three sample-relative risk categories for the classification exercise, using thresholds of 20 and 35 drawn from the empirical distribution of the panel.

SHAP-based explainability then revealed which variables drove the model classifications. Leverage ranked highest in feature importance, followed by asset quality and the voluntary disclosure index. The interaction between disclosure and leverage received the highest SHAP interaction value, broadly aligned with the econometric finding that the disclosure-stability relationship varies with leverage. Isolation Forest flagged several anomalous bank-year records, including one institution combining low disclosure with high leverage that received a 0.91 anomaly score, the highest in the panel.

The researchers are careful about what these numbers mean. Because the dataset contains only 49 observations and no independently verified fraud labels, the machine-learning results are exploratory proof-of-concept evidence rather than externally validated predictive outcomes. The supervised target categories were derived from the contemporaneous Z-score, so the analysis maps existing patterns rather than forecasting future distress. The team notes that future validation requires grouped bank-level, leave-one-bank-out, and out-of-time evaluation.

Despite those caveats, the study demonstrates how AI-augmented forensic accounting can enhance supervisory prioritization in institutionally fragile banking markets. The framework treats disclosure gaps, instability risk, statistical anomalies, and professional forensic review as sequentially connected but conceptually distinct stages. Anomaly scores serve as triage signals that direct scarce audit and supervisory attention, not as automated determinations of wrongdoing.

The authors propose a three-stage validation roadmap. First, expand the disclosure and CAMELS data-collection protocol to a pooled panel of banks across several Levantine or Arab markets with comparable institutional fragility. Second, digitize and text-mine the full annual-report narratives underlying the binary disclosure checklist, enabling natural-language features to capture tone, specificity, and hedging language. Third, partner with banking supervisors or audit firms willing to share historical examination outcomes under confidentiality safeguards, so that ensemble classifiers can be tested against genuine supervisory or enforcement labels rather than Z-score proxies.

Until such validation is complete, the findings remain confined to the Palestinian listed banking sector and comparable small or institutionally unstable markets. But the study’s core message travels well beyond Palestine: in fragile banking environments where supervisory resources are limited and disclosure practices are uneven, explainable machine learning can complement traditional econometric analysis by surfacing the conditional, nonlinear, and interaction-dependent risk patterns that linear models cannot fully capture.

The distinction between the two Z-score traditions deserves emphasis, because it is a common source of confusion in the applied literature. Altman’s original corporate-failure model was built on discriminant analysis of manufacturing firms and weights five accounting ratios into a single composite score. The Bank Z-score used in banking research, by contrast, is constructed from the return on assets, the equity-to-asset ratio, and the standard deviation of returns, so that it expresses the distance in standard deviations between a bank’s profitability and the point at which equity is exhausted. A higher value implies a larger buffer against insolvency, which is why the Palestinian study treats low Z-score categories as elevated risk. The authors are explicit that the two measures share a name and a general intuition but not a formula, and readers comparing results across studies should check which version is in use before drawing conclusions.

The CAMELS framework, whose six dimensions anchor the econometric specification, originated as an internal supervisory rating system used by examiners during on-site inspections. Capital adequacy, asset quality, management efficiency, earnings, liquidity, and sensitivity to market risk together summarize the conditions examiners traditionally assess. Using publicly available proxies for these dimensions in a panel regression is a well-established practice in markets where confidential examination ratings are not accessible to researchers, and it is precisely this gap between public proxies and confidential supervisory information that the study’s proposed validation roadmap is designed to close.

The choice of fraud-detection screeners also situates the work within a broader methodological lineage. Deterministic composite measures such as the Beneish M-score and the Dechow F-score long predate machine-learning approaches, and each condenses a set of accounting signals into a single index of potential misstatement. Ensemble tree learners extend this tradition by letting the data determine which combinations of variables and thresholds best separate observations, rather than fixing the weights in advance. The Palestinian results are consistent with the wider finding that tree-based ensembles often outperform linear benchmarks when the underlying relationships are conditional and interaction-dependent, though the authors rightly caution that a 49-observation panel cannot support strong claims about generalization.

SHAP values themselves rest on a well-defined foundation from cooperative game theory, where they allocate the contribution of each feature by averaging its marginal effect across all possible orderings of inputs. This gives the attribution a consistency property that simpler importance measures lack, which is why SHAP has become a standard tool in explainable AI for regulated domains. In the Palestinian application, the alignment between the highest SHAP interaction value and the econometric interaction term offers a rare point of convergence between the two analytical layers, suggesting that the machine-learning model was picking up structure the regression had already identified rather than artifacts of the small sample.

Isolation Forest operates on a different principle from the supervised classifiers. Rather than learning a decision boundary from labeled examples, it isolates observations by recursively partitioning the feature space, so that points requiring few splits to isolate receive high anomaly scores. This makes it suitable when genuine fraud labels are scarce or absent, as they are here. The flagged record combining low disclosure with high leverage illustrates the intended workflow: the algorithm does not assert wrongdoing, it nominates a bank-year for human review, and the forensic accountant or supervisor must then examine the underlying filings, loan books, and governance circumstances before drawing any conclusion.

The institutional context of the Palestine Exchange amplifies both the challenge and the potential value of this approach. A market with seven listed banks offers no statistical power for large-sample inference, yet it is exactly the kind of environment where supervisory capacity is stretched thinnest and where even modest triage tools could meaningfully redirect attention. The 2019 to 2025 window also spans periods of severe economic disruption in the region, which likely injected volatility into the stability measures and may partly explain why the disclosure-stability relationship shifted direction across moderators. Panel data from calmer periods, or from larger neighboring markets, would help disentangle crisis effects from structural relationships.

The eleven-item disclosure index itself reflects an evolving international agenda. Items covering cybersecurity policy, customer-data protection, green financing, financial inclusion, and gender diversity go well beyond the traditional voluntary-disclosure checklist of segment reporting, accounting policies, and forward-looking information. Their inclusion acknowledges that stakeholders in emerging markets increasingly evaluate banks on non-financial dimensions, and that silence on these topics can itself be informative. Whether narrative text mining can eventually capture tone and hedging in Arabic-language annual reports, as the roadmap proposes, remains an open empirical question, since most large language models and text-mining benchmarks were developed and validated primarily on English-language corporate filings.

For practitioners, the most transferable element may be the sequential architecture rather than any specific number. Treating disclosure gaps as the first screen, statistical anomalies as the second, and professional forensic judgment as the final arbiter mirrors how mature supervisory early-warning systems are designed in larger jurisdictions. The Palestinian study demonstrates that this architecture can be assembled with modest data and open-source tooling, which lowers the barrier for other small markets to adopt and adapt the framework while the validation agenda proceeds.

Subject of Research: AI-augmented forensic accounting applied to Palestinian banks reveals conditional disclosure-stability relationships and flags anomalous bank-year records for supervisory review.

Article Title: Voluntary disclosure, banking stability, and AI-augmented forensic accounting: an exploratory econometric and machine-learning study of Palestinian banks

Article References: Razia, B. S., Jadallah, N. I., Zureigat, Q., Khamis, R., & Awwad, B. S. (2026). Voluntary disclosure, banking stability, and AI-augmented forensic accounting: an exploratory econometric and machine-learning study of Palestinian banks. Frontiers in Big Data, 9, Article 1914133. https://doi.org/10.3389/fdata.2026.1914133

Image Credits: AI Generated

DOI: 10.3389/fdata.2026.1914133

Keywords: machine learning, forensic accounting, banking stability, voluntary disclosure, anomaly detection, explainable AI, Palestine Exchange, financial fraud, Voluntary, disclosure, banking, stability