Modeling the Factors Influencing Bank Failure Using a System Dynamics Approach and Genetic Algorithm

Introduction

Bank failure is rarely the product of a single adverse ratio. It normally emerges from a sequence in which asset losses, weak earnings, costly funding, liquidity pressure, and inadequate capital reinforce one another. Recent evidence spanning US commercial banks from 1863 to 2024 shows that failing institutions share three persistent features: increasing asset losses, deteriorating solvency, and growing dependence on expensive noncore funding (Correia et al., 2026). The 2023 banking turmoil further demonstrated how rapidly interest-rate risk and uninsured deposit concentration can convert apparently manageable balance-sheet weaknesses into acute fragility. Mark-to-market estimates for US banks indicated an aggregate asset-value decline of about $2 trillion during the monetary-tightening episode, with the most vulnerable institutions combining large valuation losses, low capital, and high uninsured leverage (Jiang et al., 2024). These episodes renewed the case for forward-looking supervisory tools that connect bank-level conditions to macrofinancial shocks.

Prudential supervision therefore treats capital, liquidity, governance, asset quality, and risk management as interdependent rather than separable safeguards. The revised Basel Core Principles require supervisors to assess these domains on both a solo and consolidated basis and to adapt supervisory intensity to the risk profile and systemic importance of each bank (Basel Committee on Banking Supervision [BCBS], 2024). The International Monetary Fund similarly argues that a high-interest-rate environment requires more demanding stress tests, closer attention to interest-rate and liquidity exposures, and credible access to emergency liquidity arrangements (International Monetary Fund [IMF], 2023). A useful early-warning framework should therefore do more than classify a bank as healthy or distressed. It should explain how the condition develops, identify where feedback amplifies the problem, and test whether a proposed response remains effective over time.

The empirical bank-failure literature provides a strong foundation for such work. Capital adequacy, asset quality, management quality, earnings, and liquidity remain central predictors, but their relevance varies by institution and macroeconomic regime. Corporate-governance indicators can extend the warning horizon beyond that obtained from financial ratios alone (Alzayed et al., 2023), while state-level and national conditions change the failure probability of already vulnerable banks (Wu & Cole, 2024). Text mining of failed-bank reviews identifies weak lending decisions, board oversight, supervisory deficiencies, and concentrated exposures as recurring causes (Le et al., 2023). Recent machine-learning studies achieve high classification accuracy, but sensitivity and stress tests show that predictive performance can deteriorate sharply when the most informative financial features are impaired (Hu et al., 2025). Prediction and explanation are thus related but distinct tasks.

This distinction is especially important in Iran, where prolonged inflation, administered rates, government claims, funding costs, and institution-specific balance-sheet structures may act simultaneously. The present study uses the accumulated-loss-to-capital ratio as an operational indicator of capital distress. Article 141 of Iran’s Commercial Code requires the board of a joint-stock company to convene an extraordinary general meeting when losses eliminate at least half of registered capital, so that continuation or dissolution can be considered (Legal Bill Amending Parts of the Commercial Code of Iran, 1968, art. 141). In the model’s sign convention, accumulated losses are negative; consequently, −0.5 functions as the policy constraint. This indicator is not treated as a complete legal or prudential definition of bank insolvency. It is a transparent, locally meaningful measure around which dynamic policy experiments can be organized.

Most bank-failure studies estimate the marginal contribution of predictors or optimize classification accuracy. Those approaches are valuable for surveillance, yet they may not show how a policy changes stocks and flows across several years. System dynamics addresses this limitation by representing feedback, delay, accumulation, and nonlinear response explicitly (Sterman, 2000). It has been used to examine bank resilience under crises (Bastan et al., 2024), systemic-risk governance in interacting asset–liability structures (Fatone & Mariani, 2020), and macrofinancial reform in a stock-flow framework (van Egmond & de Vries, 2024). A genetic algorithm can complement that framework by searching a large, nonlinear policy space without requiring a closed-form objective function (Katoch et al., 2021). The combination is therefore suited to a question that is both structural and operational: what time-varying combination of internal levers can keep a modeled distress indicator within a specified constraint while limiting unnecessary intervention?

Accordingly, this study has three objectives. First, it identifies and validates the internal and external factors associated with capital distress and accumulated losses in a large Iranian commercial bank. Second, it organizes those factors into causal-loop and stock-and-flow structures and evaluates historical behavior. Third, it uses genetic-algorithm optimization to compare a broad multi-lever intervention with a more restrictive policy domain. The contribution is not a universal bankruptcy classifier. It is an explainable case-based decision model that links expert knowledge, historical data, dynamic simulation, and policy search.

Literature Review and Conceptual Background

Bank distress has both bank-specific and environmental determinants. At the bank level, studies repeatedly identify weak earnings, inadequate capital, poor asset quality, inefficient operations, and unstable funding. In a European sample, credit risk, liquidity risk, and market power were among the strongest correlates of financial distress (Cipollini & Fiordelisi, 2012). Evidence from Finnish cooperative and savings banks likewise shows that cost inefficiency, income structure, and macroeconomic growth contribute to insolvency risk (Huhtilainen, 2020). A large historical panel reaches a compatible conclusion: public accounting information contains substantial warning content because asset losses, solvency deterioration, and noncore funding usually worsen before failure (Correia et al., 2026). Capital and liquidity should therefore be read jointly with the quality of income and the cost of obtaining funds.

Macroeconomic conditions modify those internal mechanisms. High inflation changes nominal balance-sheet values, funding costs, borrower repayment capacity, and the real value of bank capital. Country-level evidence suggests that inflation, financial inclusion, and the loan-to-deposit ratio interact in determining financial stability rather than operating independently (Ozili, 2025). Wu and Cole (2024) find that bank-level financial condition is indispensable for prediction, while local and national macroeconomic conditions matter particularly for vulnerable banks. The real-economy effects can also run in the opposite direction. Historical evidence from the Panic of 1825 shows that failures propagated through correspondent networks and increased bankruptcies among nonfinancial firms (Olmstead-Rumsey & Ravalli, 2026). These reciprocal effects are a central reason to model bank distress as a feedback system.

Methods for bank-failure prediction have evolved from discriminant and logit models toward ensembles, neural networks, and explainable machine learning. Comparative research finds that random forests and neural networks can deliver strong out-of-sample performance, although model rankings vary with the dataset, horizon, and error cost (Petropoulos et al., 2020). Cost-sensitive modeling is particularly relevant because failing to identify a truly vulnerable bank is generally more consequential than issuing a false warning; nevertheless, logistic regression remains a competitive benchmark (Ekinci & Sen, 2024). Machine-learning research also confirms the importance of capital, leverage, profitability, competition, and dynamic features (Lagasio et al., 2022; Hu et al., 2025). These models are powerful screening devices, but their predictions do not automatically reveal the policy sequence required to change a bank’s trajectory.

Interpretability has consequently become a substantive requirement rather than a cosmetic addition. Neural additive models can preserve nonlinear predictive relationships while showing how individual financial indicators contribute to a failure estimate (Son et al., 2026). Earlier cross-country evidence also shows why that explanation must remain context-sensitive: stronger capital, asset quality, management, earnings, and liquidity generally reduce failure probability, but macroeconomic growth, inflation, and real interest rates change the conditions under which those fundamentals operate (Lin & Yang, 2016). For Iranian banks, recent research similarly combines internal ratios with macroeconomic and interbank effects; Taheri Haft Asyabi et al. (2025), for example, integrate artificial intelligence with spatial econometrics to represent the possibility that vulnerability is not independent across banks. The present study does not compete with those models on classification accuracy. It addresses the complementary need to explain how an identified vulnerability may evolve after management changes several connected levers.

System dynamics begins from a different premise: observed behavior is generated by the interaction of stocks, flows, feedback loops, delays, and decision rules. Validity is therefore assessed through both structure and behavior, including dimensional consistency, extreme-condition reasoning, expert review, and comparison with reference modes (Barlas, 1996). In banking, this permits deposits, credit, income, operating costs, profit or loss, capital, and accumulated losses to be treated as a connected mechanism rather than a flat list of predictors. The Iranian commercial-bank application of Bastan et al. (2024) illustrates how liquidity coverage, stable funding, capital adequacy, and crisis policies can be tested within one simulation environment. The present study extends that logic by using the simulation itself as the evaluation environment for an evolutionary policy search.

The genetic algorithm represents candidate policy paths, evaluates their simulated consequences, and iteratively retains or recombines better-performing candidates. This is appropriate when policy levers interact, constraints must hold in every period, and the mapping from intervention to outcome is nonlinear. The algorithm does not remove the need for model validation; it magnifies it, because an optimizer can exploit unrealistic features in a weak model. The expert-validation and historical-reproduction stages are therefore integral to the research design rather than preliminary formalities.

Materials and Methods

Research design and case setting

The study used an applied sequential mixed-method design. The qualitative stage identified candidate determinants and elicited causal knowledge; the quantitative stage translated that knowledge into a system-dynamics model and optimized selected policy levers with a genetic algorithm. Bank Melli Iran was the empirical case. The archival series supplied for the study covered Iranian years 1390–1403 (approximately 2011–2024). Historical behavior plots in the source model cover 1390–1402, while the policy comparison extends to 1403. Monetary values and institution-specific parameters were used in the original model; the present manuscript reports the validated structure and outcomes without altering the supplied data or recalculating the simulation.

Literature review and expert panel

Candidate variables were assembled by searching Web of Science and Scopus for terms equivalent to “bank failure,” “determinants of bank failure,” “bank-failure models,” and “bank failure and system dynamics.” The evidence base was complemented by semi-structured interviews with 11 participants drawn from senior bank managers, researchers, and university faculty. Selection was nonprobability-based, purposive, and convenience-oriented. Eligibility required at least 15 years of banking-management experience, familiarity with bank failure, and relevant postgraduate education in economics, financial management, or accounting. The panel was asked to assess factor relevance, the direction of causal links, logical coherence, and the applicability of the resulting structure to bank capital distress.

Fuzzy-Delphi screening

A five-category questionnaire—very high, high, moderate, low, and very low importance—was administered in three fuzzy-Delphi rounds. Fuzzy Delphi is designed to retain the linguistic uncertainty in expert judgments while supporting convergence across repeated assessments (Ishikawa et al., 1993). The source analysis used its established fuzzy coding and defuzzification procedure. At the end of round three, all retained factors had reported mean scores above 8 and between-round differences below 20%, meeting the acceptance rule used in the original study (Modiri et al., 2019). The final factor table is reproduced in English with the same entries and source assignments as the Persian manuscript.

System-dynamics model development

Panel members first specified the sign and logic of relationships among the retained variables. Unsupported links were removed, and the remaining structure was encoded in Vensim as causal-loop diagrams. The stock-and-flow model was then developed from those loops and contained 17 state variables. The architecture connected deposit acquisition and capital increases to available resources; resources to lending, investment, assets, and income; and deposit costs, general and personnel expenses, and doubtful-debt expenses to profit or loss. Profit or loss accumulated over time and fed back to capital. Inflation, national liquidity, deposit and lending rates, and central-bank supervision entered as environmental or policy variables.

Model equations were formulated from the causal logic, accounting relations, and the case data. Experts reviewed the structure for coherence and relevance. Historical reproduction then compared simulated and observed behavior. The source manuscript reports an overall correlation of approximately 0.92 and a mean error below 9% for the baseline reproduction. Sensitivity analysis varied model parameters to assess their influence on the accumulated-loss-to-capital trajectory. These tests were used to establish confidence in the model as a policy-comparison device rather than to claim universal predictive validity.

The model boundary was deliberately centered on capital distress and accumulated losses. It included the deposit categories, resource and credit mechanisms, major income and cost components, assets, liabilities, capital, and the environmental variables listed in Table 1. It did not attempt to reproduce every prudential ratio or every legal condition that could define bank failure. This boundary keeps the simulation aligned with the stated Article 141-based indicator while avoiding the stronger claim that one ratio is equivalent to regulatory insolvency. The reported structural review and historical comparison are the validation evidence available in the source manuscript; no additional numerical validation tests are attributed to the original analysis in this English version.

Genetic-algorithm policy search

The genetic algorithm searched annual values for three internal policy levers: resource-mobilization efficiency, funding-composition efficiency, and credit efficiency. Each candidate policy path was passed to the Vensim simulation, and its time-dependent accumulated-loss-to-capital ratio was evaluated. The search objective was to identify the smallest intervention compatible with the distress constraint across the full simulation horizon. In the model’s sign convention, a ratio below −0.5 violates the Article 141-based constraint.

Scenario 1 permitted each of the three levers to vary jointly from 1.00 to 1.35 at 13 annual transition points. Scenario 2 imposed a more restrictive domain in which final lever effectiveness could not exceed 1.00, operationalized in the source study as avoiding reliance on the market share or performance space of other banks. Before optimization under Scenario 2, a maximum-intervention feasibility test set all levers simultaneously at their permitted ceiling. This test distinguishes an infeasible policy domain from a poor algorithmic search: if the ceiling itself violates the constraint, no alternative candidate within that domain can be fully feasible.

Results

Identification and validation of influencing factors

The literature review and expert process produced the final factor set reported by the original analysis as 36 validated determinants. These span capital and financial structure, lending activity, deposit structure, resource mobilization, liquidity, income, costs, macroeconomic conditions, and supervision. Every retained factor met the study’s reported fuzzy-Delphi thresholds after three rounds. Table 1 reproduces the Persian source table in the same paired four-column arrangement and retains every source designation.

The retained variables can be read in two broad groups. Internally manageable variables include capital injection and revaluation, assets, lending, the composition and cost of deposits, resource-mobilization efficiency, credit efficiency, fee and lending income, personnel expense, doubtful-debt expense, liquidity, and accumulated profit or loss. Environmental or institutionally constrained variables include national inflation and liquidity, the money-to-liquidity ratio, claims on government, administered or market-linked rates, and central-bank supervision. The distinction is analytical rather than absolute: external conditions alter the effectiveness of internal decisions, while bank decisions determine how strongly an external shock is transmitted to capital.

Note. The factor entries and their source attributions are translated and reproduced in the same four-column structure as the Persian source table. “Experts” denotes variables introduced or retained through the expert-interview and fuzzy-Delphi process.

Feedback structure

After the factor set was accepted, experts specified plausible positive and negative relationships and removed links judged irrelevant. Figure 1 shows the reinforcing growth-and-profit-accumulation loop (R1). An increase in assets expands lending capacity; a larger lending volume produces more income; income supports profit; profit increases accumulated profit and internal funds; and those funds support additional investment and assets. The loop is reinforcing because a change in one component can circulate through the structure and amplify the original direction of change.

Note. The original model output from the Persian source manuscript is reproduced without alteration; the numerical content, links, and trajectories are unchanged.

The wider causal architecture is presented in Figure 2. It connects internal balance-sheet and operating variables with external conditions through reinforcing and balancing feedback. The map is intentionally dense because it preserves the expert-validated network: deposit types affect total deposits and funding cost; resources support credit and investment; credit generates income but also creates exposure to doubtful accounts; income and cost determine periodic profit or loss; and accumulated profit or loss changes capital. Inflation, national liquidity, rates, government claims, and central-bank supervision condition several of these relationships. Figure 2 is therefore the structural bridge between the factor list in Table 1 and the quantified simulation.

Several feedback mechanisms are embedded in this network. A resource–credit–income mechanism can reinforce growth when additional deposits are converted into performing credit. A cost mechanism can counter that growth when higher deposit rates or a more expensive maturity mix raise funding expense. A credit-quality mechanism can reverse the benefit of lending if doubtful accounts and provisioning expenses increase. Finally, the profit–capital mechanism carries current operating performance into future lending capacity through accumulated profit or loss. These mechanisms explain why a policy that improves only deposit volume may not improve the distress ratio if the acquired funds are costly or used inefficiently.

Note. The original model output from the Persian source manuscript is reproduced without alteration; the numerical content, links, and trajectories are unchanged.

Stock-and-flow model and baseline behavior

Figure 3. Stock-and-flow model of bank failure from the perspective of capital distress and accumulated losses.

Note. The original model output from the Persian source manuscript is reproduced without alteration; the numerical content, links, and trajectories are unchanged.

Following structural confirmation, selected variables were simulated over the historical reproduction period. Figure 4 presents the original four panels. Resource-mobilization efficiency increased by about 20% relative to the base year, although the path was nonmonotonic. Funding-composition efficiency weakened, indicating that growth in the volume of resources did not by itself ensure a desirable mix by type, cost, maturity, or operational usability. Credit efficiency and the lending rate also varied across the period. The accumulated-loss-to-capital ratio reached a less favorable position in 1398 and subsequently improved through 1402, matching the reference value at the end of that reproduction interval. The baseline model achieved the reported correlation of approximately 0.92 and mean error below 9%.

Figure 4. Simulated behavior of selected variables during the historical reproduction period.

Note. Panels reproduce the four original charts exactly as provided in the Persian manuscript. From upper left to lower right: resource-mobilization efficiency, funding-composition efficiency, credit efficiency, and the lending rate.

Policy-scenario results

Two combined policies were assessed in addition to the reference trajectory. Under Scenario 1, the genetic algorithm optimized 13 annual values for each of the three levers over the common range 1.00–1.35. The reported best final fitness was approximately −0.34929; the minimum accumulated-loss-to-capital ratio was approximately −0.49996; and the maximum ratio was approximately 0.31498. Feasibility was positive. The algorithm therefore identified a policy path that respected the constraint in all 13 simulated years. The largest intervention pressure appeared in the critical years 1395–1398, and the ratio approached its minimum permitted value in 1398. Values close to 1 in the final years indicate that the stated objective avoided unnecessary intervention; they do not imply that the associated levers were economically irrelevant.

Scenario 2 used a narrower policy domain. The maximum-intervention test placed all levers at the permitted ceiling of 1.00. The resulting ratios for 1395, 1396, 1397, and 1398 were approximately −0.7144, −0.6535, −0.7129, and −1.0543, respectively. Thus, even the strongest permitted intervention failed to satisfy the constraint in those four years. Because the ceiling itself was infeasible, the absence of a fully feasible solution did not depend on the quality of the genetic-algorithm search. It reflected the structural limits of the restricted policy space.

The source comparison produced a dual result. Scenario 1 was the best feasible policy when intervention could vary across the general range of 1.00–1.35, because it respected the constraint in every simulated year. When the scenario trajectories were compared with the baseline, both policy paths changed the accumulated-loss-to-capital trajectory, and the original analysis reported Scenario 2 as having the lowest indicator value. Figure 5 reproduces the scenario chart exactly as supplied in the Persian manuscript. Sensitivity analysis identified inflation as the most influential parameter, with a mean absolute effect of 0.178. Among the controllable policy levers, credit efficiency had the strongest influence on the accumulated-loss-to-capital path.

Note. The original model output from the Persian source manuscript is reproduced without alteration; the numerical content, links, and trajectories are unchanged.

Discussion

This study shows that capital distress is generated by the interaction of internal banking decisions and the macroeconomic environment. The qualitative phase did not reduce failure risk to a single accounting ratio. It identified clusters involving capital, assets, liabilities, deposit composition, credit activity, income, operating costs, liquidity, government claims, inflation, interest rates, and supervision. That pattern is consistent with contemporary evidence that earnings, liquidity, governance, and balance-sheet quality jointly shape failure risk (Alzayed et al., 2023; Citterio, 2024). It also aligns with the historical finding that fundamental deterioration is usually visible before failure in asset losses, solvency, and funding structure (Correia et al., 2026). The added value of the present model is to place those factors in an explicit causal sequence instead of treating them only as parallel regressors.

The feedback representation clarifies why apparently favorable changes can have adverse downstream effects. Deposit growth increases the resources available for lending and investment, but it simultaneously expands liabilities and may raise interest expense. Lending can increase income, yet weak credit screening can generate future doubtful-debt costs. Asset revaluation may strengthen recorded capital, but it does not necessarily improve operating cash flow or funding quality. These interactions explain why funding volume, funding mix, and credit efficiency were modeled as separate levers. They also support the BCBS (2024) emphasis on integrated capital, liquidity, governance, and risk assessment rather than isolated compliance indicators.

The scenario results emphasize the importance of the feasible policy domain. Scenario 1 distributed adjustment across three levers and kept the modeled ratio within the constraint over all 13 transitions. Scenario 2 was more restrictive and failed the constraint in 1395–1398 even when every lever was placed at its maximum permitted value. The practical message is not simply that one numeric path is superior. It is that internal reform may become insufficient when the permissible adjustment range is too narrow relative to accumulated losses and external shocks. In such periods, recapitalization, supervisory intervention, or macroeconomic stabilization may be required alongside operating reform. This conclusion is compatible with system-dynamics research showing that resilient banking depends on a balanced combination of capital adequacy, liquidity, stable funding, profitability, and crisis management (Bastan et al., 2024).

Inflation’s mean absolute sensitivity of 0.178 is particularly important in the Iranian setting. Inflation can alter deposit behavior, nominal rates, funding expense, borrower quality, asset values, and the real purchasing power of capital at the same time. Its influence therefore passes through several model pathways. Cross-country evidence similarly finds that inflation conditions the relationship between financial inclusion, loan-to-deposit positions, and stability (Ozili, 2025). The result does not imply that internal management is unimportant; credit efficiency was the most sensitive policy lever. Rather, it indicates a shared responsibility: bank management must improve controllable processes, while monetary and supervisory authorities address the environment in which those processes operate.

For supervision, the model supports a staged early-warning process. First, the accumulated-loss-to-capital ratio should be monitored together with leading measures of funding cost, resource composition, credit efficiency, and liquidity. Second, dynamic simulation should be used before the legal threshold is reached to test whether realistic internal actions are sufficient. The original study recommends an early-warning flag when the loss ratio reaches approximately −0.30, giving decision makers a buffer before the −0.50 constraint. Third, if the maximum credible internal intervention remains infeasible, the issue should be escalated rather than masked by a mathematically optimized but operationally unattainable policy. This approach is consistent with the call for more stringent stress testing and closer review of interest-rate and liquidity risk in a high-rate environment (IMF, 2023).

For the study bank, the policy implications reported in the source analysis are specific. Management should improve the mobilization of lower-cost resources; the optimized path indicated a coefficient of approximately 1.27 in the initial years. It should also improve the mix of resources to reduce dependence on expensive long-term funding, strengthen credit-efficiency monitoring, adjust lending rates gradually in inflationary periods, and conduct asset revaluation at an appropriate time. These actions should be evaluated as a package, because isolated improvement in one lever can be offset elsewhere in the system.

Methodologically, using the system-dynamics simulation as the genetic algorithm’s evaluation function allows policy search in a nonlinear, path-dependent setting. This goes beyond asking which variable has the largest coefficient. It asks whether a sequence of feasible decisions keeps the system within a constraint at every step. The benefit comes with a corresponding obligation: model structure, parameterization, and optimization settings must be sufficiently documented for independent reproduction. Historical fit alone cannot validate every causal assumption, and a high-performing algorithm cannot compensate for an incorrectly specified model (Barlas, 1996).

The model also clarifies the difference between prediction, stress testing, and policy design. A predictive model estimates the probability or class of failure from observed features. A stress test asks how a predefined shock changes a bank’s condition. The present optimization asks a third question: among admissible time-varying actions, which path satisfies a constraint with the least stated intervention? These tasks can support one another. Machine-learning models can identify signals that deserve closer attention; a dynamic model can explain how those signals interact; and constrained optimization can test the operational sufficiency of candidate responses. Treating one method as a substitute for all three tasks would obscure important information.

Implementation should therefore be governed as a recurring decision process rather than a one-time simulation. Parameters and reference modes should be updated when accounting definitions, interest-rate rules, balance-sheet composition, or the macroeconomic regime changes. Policy paths should be screened for legal, operational, and market feasibility before management adopts them. Differences between simulated and observed behavior should trigger model review, not automatic recalibration until a desired result appears. Finally, the policy dashboard should display both the optimized path and the maximum-intervention feasibility test. The latter communicates when internal action is structurally insufficient and prevents decision makers from interpreting an infeasible optimization as a software failure.

Limitations and Future Research

The findings are conditional on a single bank, the supplied historical series, the expert-defined structure, and the policy ranges used in the simulation. They should not be generalized to all Iranian banks without re-estimation and renewed structural validation. The principal outcome is an accumulated-loss-to-capital ratio; it does not replace prudential measures of capital adequacy, liquidity, asset quality, profitability, or supervisory judgment. Expert elicitation introduces subjectivity, and the raw fuzzy-Delphi matrices were not available for independent recalculation in this language-editing stage.

The source manuscript also reports a computational limitation in running larger genetic-algorithm populations and does not provide the complete optimization hyperparameter set, including population size, crossover and mutation rates, stopping rule, and random seed. Those details should be supplied from the original computational files before journal submission. Future studies should compare several banks, use longer time series, evaluate alternative distress indicators, and test whether machine-learning forecasts or multicriteria causal methods add information beyond the present feedback structure.

Conclusion

Bank capital distress is a cumulative process created by feedback among resources, credit, income, costs, capital, and the macroeconomic environment. By combining expert elicitation, fuzzy-Delphi screening, system dynamics, and genetic-algorithm search, this study provides an explainable framework for examining that process and comparing policy paths. The broad multi-lever scenario yielded a feasible path across all 13 simulated transitions, whereas the restricted domain was infeasible during four critical years. Inflation was the most sensitive external parameter, and credit efficiency was the strongest policy lever. These results support early dynamic testing of internal reform capacity and timely escalation when that capacity is insufficient. They remain scenario-based estimates, not a guarantee that bank failure can be prevented.

Declarations

Ethics approval: The source manuscript did not report the name of an approving or waiving ethics body, an approval/waiver number, or an approval date. Because the study involved expert interviews and a Delphi process, these details must be supplied by the authors before journal submission and cannot be inferred or created during language editing.

Informed consent: The source manuscript did not document the consent procedure used for the 11 expert participants. The corresponding author should confirm whether written or verbal informed consent was obtained and whether participants consented to the use of anonymized responses.

CRediT author contributions: Author-level CRediT roles were not specified in the source manuscript. All four authors should review and approve a role-specific contribution statement before submission.

Funding: No funding statement was provided in the source manuscript. The authors should disclose all financial support or explicitly confirm that the work received no specific funding.

Conflict of interest: No conflict-of-interest statement was provided in the source manuscript. Each author should disclose relevant financial and nonfinancial interests or explicitly declare that none exist.

Data availability: The analysis used institution-specific banking data. The source manuscript did not specify public access conditions or confidentiality restrictions. No dataset is distributed with this manuscript; the corresponding author should provide the journal with an accurate access statement.

Acknowledgments: No acknowledgments were reported in the source manuscript.

Use of artificial intelligence tools: An AI-assisted language-editing tool was used to improve English expression, organization, citation consistency, and document formatting. It was not used to generate, recalculate, or alter the study data, figures, table entries, analyses, or numerical results. The authors retain full responsibility for the manuscript and must verify the final text before submission.