When Does Size Protect and When Does It Concentrate Risk? Determinants of Default in Popular and Solidarity Economy Cooperatives in Ecuador (2020–2025)

1. Introduction

Loan delinquency constitutes a visible manifestation of financial risk in financial institutions, as it reflects the deterioration of asset quality, reduces profitability, limits liquidity, and restricts future credit expansion. Recent literature indicates that non-performing loans (NPLs) originate from a combination of structural, internal, and macroeconomic factors (Salas et al., 2024). In this context, a global study encompassing 1,631 financial institutions in 111 countries concludes that the NPL ratio is determined by a stable set of specific and macroeconomic factors across different regions and time periods. These findings support the suitability of panel data approaches for analyzing loan delinquency in cooperative financial systems such as Ecuador's (Xu, 2025).

The phenomenon observed in Ecuadorian Savings and Credit Cooperatives (COACs) is far from trivial. Recent empirical evidence indicates that average delinquency rates reveal structural differences between Segments 4 and 5; that a trend reversal began in 2022; and that the Chow test rejects the homogeneity of coefficients between segments, justifying a differentiated analytical approach. Furthermore, the estimated dynamic model shows that current delinquency levels depend on their own lagged value, portfolio concentration in microcredit, institutional size, the coverage of loan loss provisions, liquidity conditions, and temporary shocks such as the COVID-19 pandemic and the post-2022 period (Noriega et al., 2023).

However, the existing literature on NPLs in cooperative institutions within Latin America's Popular and Solidarity Economy (PSE), and particularly in Ecuador, remains limited. Consequently, most of the available empirical research focuses on the commercial banking sectors of Europe and Asia (Salas et al., 2024; Xu, 2025),microfinance systems in sub-Saharan Africa and South Asia (Alnabulsi et al., 2023; Ozili, 2025), or case studies of microcredit portfolios in Peru and Colombia (Durango-Gutiérrez et al., 2023; Quispe Mamani et al., 2022). While these studies provide valuable conceptual and empirical insights into the determinants of credit risk, they do not capture the distinctive characteristics of Ecuador's cooperative financial model. In fact, there is limited panel data evidence to identify whether the determinants of loan delinquency have similar effects across cooperatives of different sizes within a common regulatory framework, or whether loan portfolios respond differently to exogenous shocks as severe as the COVID-19 pandemic and the post-2022 credit cycle. This gap is relevant not only from an academic perspective, but also for the design of differentiated prudential policies within the Ecuadorian Popular and Solidarity Economy (EPS) system, which is supervised by the Superintendency of the Popular and Solidarity Economy (SEPS).

From a substantive perspective, these findings are consistent with contemporary literature on NPLs, which identifies three main groups of determinants: (i) macroeconomic variables, (ii) financial institution-specific factors, and (iii) institutional or regulatory factors. In fact, a systematic review of 76 peer-reviewed studies likewise identifies these three groups as the dominant analytical pillars of current NPL research (Mooneeapen and de Jager, 2025).

1.1. The Social and Solidarity Economy: Theoretical Foundations and Institutional Framework

1.1.1. Concept and Principles of the Social and Solidarity Economy

The Social and Solidarity Economy (SSE) is a heterodox model of economic organization based on the principles of voluntary cooperation, democratic and participatory governance, institutional autonomy, and the primacy of people over capital, especially in the distribution of surpluses (OECD, 2024). Unlike conventional financial institutions, whose main objective is to maximize shareholder value, SSE organizations—including cooperatives, associations, mutual societies, and social enterprises—develop activities aimed at satisfying collective needs and promoting the well-being of their members (Magallón et al., 2024).

In Latin America and the Caribbean, the COACs have gained increasing importance as a mechanism for promoting the financial inclusion of households and microenterprises with limited access to the formal banking system. The Organisation for Economic Co-operation and Development (OECD, 2024) documents that SSE organizations in the region play a central role in generating local employment, leveraging social capital, and providing financial services in areas with low banking density. This role is particularly significant in Ecuador, where the cooperative savings and credit system has progressively increased its share of total credit in the economy, rising from 12% of GDP in 2015 to 21.5% in 2024 (Altamirano et al., 2025).

1.1.2. Savings and Credit Cooperatives as Local Financial Intermediaries

COACs differ from commercial banks in three dimensions that have direct implications for credit risk. First, the dual status of member-owner and client-borrower creates incentives for mutual monitoring, which, in principle, reduces the risk of default and adverse selection (Salinas Vásquez et al., 2024). Second, democratic governance—one member, one vote—can promote more conservative risk management when members internalize portfolio losses, but it can also lead to pressure to expand credit beyond prudential limits (Rijn, 2023). Third, territorial roots and long-term relationships with borrowers provide soft information on the ability and willingness to repay that traditional scoring models do not capture. In microfinance, this constitutes a key mechanism for credit risk management (Berger and Udell, 2002).

1.1.3. The Ecuadorian EPS System and its Regulatory Segmentation

The legal framework for the EPS system in Ecuador is established by the Organic Law on the Popular and Solidarity Economy, and its prudential regulation has been the responsibility of the SEPS since 2011 (Altamirano et al., 2025). The regulatory segmentation divides the COACs into five groups according to their total assets, from Segment 1, which includes entities with assets exceeding USD 80 million, to Segment 5, comprised of entities with assets less than USD 1 million (Resolution No. 038-2015-F, 2015).

This tiered regulatory architecture implicitly recognizes that entities of different sizes operate with distinct business models, risk profiles, and technical capabilities. In this context, segments 4 and 5, the subject of this study, comprise the smallest cooperatives in the supervised system. They are particularly relevant because they serve informal-sector borrowers and the rural micro-entrepreneurs, often representing the only accessible financial intermediary in their local communities (Jiménez-Hernández et al., 2019).

1.1.4. Tension between Social Mission and Financial Sustainability: The Inherent Risk of the Cooperative Mode

A defining characteristic of COACs is the coexistence of social and financial objectives, which creates a structural tension with prudential risk management. The mission of serving low-income borrowers, informal-sector borrowers, and borrowers with limited credit histories implies accepting levels of risk that commercial banks typically reject through adverse selection mechanisms. This phenomenon, known in the literature as the “double bottom line,” highlights the institutional dilemma these entities face, particularly the pressure to maintain high loan approval rates even when borrowers' risk profiles deteriorate (Rijn, 2023).

1.1.5. Non-Performing Loans as an Indicator of Portfolio Quality and Credit Risk

NPLs are defined as the proportion of loans for which debt service obligations are not fulfilled according to the agreed repayment schedule (Noriega et al., 2023). Recent literature considers NPLs the main indicator for measuring credit risk and asset quality. According to Alnabulsi et al. (2023), NPLs "reflect the risk that cash flows from financial institutions' loan portfolios will not be fully recovered" and are directly linked to the deterioration of banking stability and the occurrence of episodes of financial distress. Furthermore, studies highlight that an increase in NPLs restricts the supply of credit, reduces structural profitability, and constitutes a fundamental indicator for monitoring the soundness of the financial system (Xu, 2025).

In the field of microfinance and community banking, this variable takes on an additional dimension. It not only reflects non-payment but also summarizes the interaction between borrower informality, weak collateral, income volatility, limited institutional monitoring capacity, and an adverse macroeconomic environment. Research conducted in Peru and internationally in the field of microfinance shows that delinquency and default rates respond simultaneously to microeconomic variables—such as provisions, efficiency, liquidity, and portfolio structure—and to macroeconomic variables, including GDP, unemployment, and inflation (Alnabulsi et al., 2023; Quispe Mamani et al., 2022).

1.1.6. Dynamic Persistence of Credit Risk

Evidence from Ozili (2025), based on the Indian banking system, demonstrates that current delinquency levels tend to increase future delinquency due to the inherent lag in loan portfolio recovery and resolution processes. This persistence can be explained by several mechanisms. First, portfolio deterioration is not immediately corrected, especially in small and medium-sized institutions with limited resources for collection, renegotiation, or write-offs (Noriega et al., 2023). Second, the accumulation of NPLs consumes operating resources and managerial capacity, reducing the capacity to originate higher-quality loans. Finally, distressed loans are often associated with asymmetric information problems and a higher reputational risk for the borrower, making their recovery slower and more costly (Enoch et al., 2021).

From a theoretical and empirical perspective, this evidence supports the dynamic and autoregressive nature of delinquency. Consequently, it is not enough to analyze the level observed in a given quarter; it is equally important to consider the recent trajectory of the portfolio. In this sense, prudential supervision should incorporate indicators of delinquency persistence, rather than limiting itself to static assessments of portfolio quality (Atari et al., 2026; Enoch et al., 2021).

1.1.7. Concentration Risk in Microcredit

Microcredit concentration is one of the most robust and consistent determinants of delinquency (Enoch et al., 2021). From a classical financial perspective, diversification reduces unsystematic risk; however, when an institution concentrates a high proportion of its portfolio in the same type of borrower or economic sector, the probability of correlated defaults increases. In the microfinance sector, borrowers often share common characteristics, such as informality, low capitalization, and the high vulnerability of their income to economic shocks, which increases the risk of simultaneous portfolio deterioration (Mahedi et al., 2026).

Studies based on microcredit portfolios in Bolivia and Colombia find that default risk is closely associated with borrower characteristics, payment arrears, the quality of collateral, and the credit analyst's assessment, highlighting the high risk sensitivity of microfinance portfolios (Durango-Gutiérrez et al., 2023).

This pattern is also supported by the most recent evidence on the vulnerability of microcredit portfolios in India, where a high concentration of credit amplified the effects of a liquidity crisis, suggesting that default can spread among borrowers with similar characteristics and economic conditions (Mahedi et al., 2026; Olayiwola, 2026). Consequently, a greater concentration in microcredit reduces institutions' capacity to absorb idiosyncratic and systemic shocks due to the high proportion of the portfolio allocated to informal or rural microenterprises. In this context, shocks to income, employment, prices, or demand can generate aggregate effects on portfolio quality (Barus et al., 2024; Noriega et al., 2023).

1.1.8. Institutional Size, Economies of Scale, and Relationship Banking

Institutional size, measured as the natural logarithm of total assets, is considered a factor that can contribute to reducing delinquency (Berger & Udell, 2002). From the perspective of economies of scale, larger institutions tend to have better credit assessment, monitoring, technological, and risk management systems. On the other hand, relationship banking theory, especially relevant in microfinance and cooperative banking, argues that a larger relative size within a local environment can translate into greater social capital, greater availability of soft information, and stronger community monitoring networks (Ramírez-Virviescas & Guevara-Castañeda, 2021).

Likewise, recent studies on microfinance show that institutional performance and efficiency depend not only on operational capacity but also on service organization, governance structure, and the strength of the relationship with the client base (Enoch et al., 2021). Similarly, Ramírez-Virviescas & Guevara-Castañeda (2021) found that greater stakeholder participation in cooperative banks is associated with lower levels of non-performing loans (NPLs).

In smaller cooperatives (segment 5), institutional growth not only implies an expansion of the balance sheet but also a strengthening of local reputation, diversification of the membership base, and greater borrower monitoring capacity. In contrast, in cooperatives in Segment 4, where institutions already have a higher level of equity, further increases in size might not generate sufficient advantages to produce significant reductions in delinquency (Enoch et al., 2021). Consequently, the effect of institutional size on delinquency depends on the organizational conditions and the context in which the cooperatives operate.

1.1.9. Provision Coverage and Proactive Prudential Management

From a conceptual perspective, higher provision coverage can be interpreted as reflecting existing credit problems or as a sign of proactive prudent management. Once other factors are controlled for, evidence suggests that higher provision coverage is associated with lower levels of NPLs, indicating that provisions can play a preventive and disciplinary role, in addition to their traditional function as a mechanism for recognizing expected losses (Becerra Guevara & Oblitas Otero, 2022).

Kyiu & Tawiah (2025) find that the implementation of International Financial Reporting Standard 9 (IFRS 9) reduces banking risk, an effect that is strengthened in environments with higher regulatory and supervisory quality. Similarly, Elashmawy & Kallunki (2026) show that the prospective provisioning approach established by IFRS 9 improves asset quality, reduces the occurrence of NPLs, and increases the predictive capacity of provisions with respect to future credit risk.

From a theoretical perspective, this approach represents a transition from the incurred loss model to the expected loss paradigm incorporated by IFRS 9. Thus, the timely recognition of expected losses contributes to strengthening portfolio quality, reducing banking risk, and improving the informational content of provisions as an indicator of the institution's financial health (Becerra Guevara & Oblitas Otero, 2022). This effect can be especially relevant in smaller cooperatives, where the high concentration of the portfolio and the limited capacity to absorb losses make a prudent provisioning policy a key mechanism for strengthening financial resilience against adverse shocks (Lunt, 2025).

1.1.10. Liquidity and the Deferral of Impairment Channel

Gelashvili et al. (2023) show that the relationship between liquidity and credit risk is complex and not strictly linear. Adusei (2022), in the context of hybrid institutions and Microfinance Institutions (MFIs), documents that liquidity is significantly associated with financial performance, although its effect varies depending on the level of credit risk. Similarly, Haris Muhammad & Yao (2024) find that, during the COVID-19 period, liquidity risk and credit risk had different effects on bank profitability, with the former having a negative impact, while liquidity was positively associated with certain performance indicators. Atari et al. (2026), using a dynamic model estimated with the Generalized Method of Moments (GMM), demonstrate that liquidity exerts a stabilizing effect and partially mitigates the adverse impact of credit risk on bank stability.

Taken together, this evidence suggests that liquidity does not directly reduce delinquency; however, it can provide flexibility for refinancing, loan restructuring, and the temporary absorption of financial stress. This mechanism can be interpreted as a deferral channel, through which liquidity delays the materialization of credit deterioration in the form of actual delinquency. Consequently, institutions with higher levels of liquidity tend to exhibit better portfolio performance in the short term, although variations in liquidity within the same institution do not necessarily translate into immediate changes in observed delinquency levels (Noriega et al., 2023).

1.1.11. Efficiency, Poor Management, and the Lagged Effect on Delinquency

The relationship between operational efficiency and portfolio quality is based on the bad management hypothesis, according to which institutions with high operating costs relative to their financial income indirectly reveal deficiencies in the processes of loan origination, monitoring, and recovery, which translate into higher levels of future delinquency (Berger & DeYoung, 1997). Under this theoretical framework, inefficiency does not immediately deteriorate portfolio quality, but rather does so through a deferred causal chain. In particular, a poor assessment of the borrower's repayment capacity increases the probability of default in subsequent periods, while weaknesses in collection processes delay the recovery of existing problematic portfolios.

Masanyiwa et al. (2022) document this complexity by identifying a nonlinear relationship between NPLs and efficiency in microfinance institutions. Their results indicate that an excessive emphasis on risk control can, paradoxically, increase operating costs without generating proportional improvements in portfolio quality. Complementarily, Enoch et al. (2021) show that efficiency improves asset quality and structural solvency. However, its effect is more clearly manifested on liquidity than on delinquency in a contemporaneous manner.

The main methodological implication for the present model is that, in a dynamic specification that incorporates one or two lags of the dependent variable, the cumulative effects of mismanagement, which materialize gradually in the delinquency trajectory, are captured by these autoregressive lags. Consequently, past delinquency already reflects the effect of origination and monitoring decisions made in previous periods, so contemporary operational efficiency loses marginal explanatory power, not because it is irrelevant to credit risk, but because its effect is already incorporated into the historical inertia of the portfolio.

This approach is consistent with the literature on dynamic panels applied to financial intermediaries (Louzis et al., 2012), and justifies including efficiency in the model as a lagged control variable at t−1, in order to mitigate simultaneous endogeneity, rather than as an independent structural determinant of delinquency.

1.1.12. Credit Cycle, Pandemic, and Exogenous Shocks

Recent research highlights that the COVID-19 pandemic and temporary relief measures altered the dynamics of NPLs and bank profitability. Haris Muhammad & Yao (2024) show that the pandemic modified how credit risk and liquidity were related to the profitability of financial institutions. Ozili (2025), in the context of the Indian banking system, concludes that the reduction in non-performing assets observed in the post-COVID-19 period is mainly explained by structural reforms implemented before the pandemic, rather than by the temporary relief measures adopted during the health emergency. This evidence underscores the need to distinguish between real improvements in portfolio quality and transitory effects resulting from regulatory or accounting changes. Likewise, Alnabulsi et al. (2023), using data from MENA countries, found that NPLs were more sensitive to specific factors within financial institutions and to institutional quality than to the direct effects of the pandemic.

Taken together, this evidence suggests that exogenous shocks, such as COVID-19, do not affect credit risk uniformly, but rather that their impact depends on institutional characteristics, the regulatory environment, and the structural soundness of each financial system.

1.2. Regulatory Segmentation and Structural Heterogeneity

Studies on cooperative banking show that credit risk depends on aggregate exposure, governance structure, and relationships with stakeholders, and that cooperatives do not constitute a homogeneous group in terms of risk profile. Beccalli and Viola (2026) find that greater stakeholder participation and involvement in Italian cooperative banks is associated with lower levels of NPLs. Similarly, Manta et al. (2026) show that, even in contexts of increasing digitalization, physical presence and territorial roots remain key determinants in the provision of credit services in Italian cooperative banks.

In the case of Ecuadorian COACs, this evidence suggests that regulatory segmentation not only classifies entities according to their total asset size, but also reflects differences in governance structure, technical capabilities, territorial density, composition of the membership base, and quality of relationship-based monitoring.

1.2.1. Research Hypotheses

H1: The delinquency rate in periods t-1 and t-2 is positively associated with the delinquency rate in period t, once other financial determinants are controlled for.

H2: A higher concentration of the loan portfolio in microcredit is associated with higher delinquency levels in credit unions.

H3: Larger institutional size is associated with lower delinquency levels.

H4: Higher loan loss provision coverage is negatively associated with delinquency.

H5: Greater liquidity is associated with lower delinquency levels.

H6: Entities in Segment 4 exhibit a higher delinquency rate than those in Segment 5, once financial factors are controlled for.

H7: Periods after 2022 are associated with higher delinquency levels, whereas the year 2020 is associated with lower delinquency levels, consistent with the regulatory deferral of loan repayments.

2. Materials and Methods

2.1. Data Source and Sample Construction

The data used in this research were drawn from the publicly available quarterly financial bulletins published by the Superintendency of Popular and Solidarity Economy (SEPS, 2025), corresponding to the COACs in segments 4 and 5 of the Ecuadorian Social and Solidarity Economy (EPS) system. Segment 4 includes entities with assets between USD 1 and 20 millions, while segment 5 comprises those with assets less than USD 1 million, according to the current regulatory classification (Resolution No. 038-2015-F, 2015). Segments 4 and 5 were selected because they contain the smallest cooperatives in the system, characterized by a stronger focus on microcredit and greater heterogeneity in their credit risk profile.

The database covers the period from the first quarter of 2020 to the fourth quarter of 2025, with a total of 24 quarters observed. The data panel is unbalanced due to mergers, liquidations, regulatory reclassifications, and the late incorporation of some entities into the supervised system. The final estimation sample comprises 5,056 quarterly observations corresponding to 368 unique entities, of which 220 belong to Segment 4 and 148 to Segment 5.

2.2. Data Processing

2.2.1. Missing Values

The provision coverage variable shows 16.5% missing observations (833 out of 5,056), resulting from entities with zero-risk portfolios in certain quarters or from reporting errors. Following the recommended practice for microfinance panels with non-random missing values (Wooldridge, 2020), each missing value is imputed using the median of the group. This method preserves sectoral and temporal variation without requiring extrapolation.

2.2.2. Outlier Processing

All continuous variables are winsorized at the 1st and 99th percentiles before being incorporated into the model. This procedure controls the influence of extreme values without eliminating observations, a standard technique in panel studies of financial intermediaries with high distributional variability (Verbeek, 2017). This technique is standard in econometric studies of financial institutions with a high dispersion in size, where extreme values often reflect legitimate exceptional events, which should not be eliminated but also should not distort the coefficients (Salas & Saurina, 2002).

2.2.3. Transformation of the Dependent Variable

The distribution of the delinquency rate exhibits a skewness of 2.88 and an excess kurtosis of 10.15, typical characteristics of credit risk indicators in microfinance systems (Salas & Saurina, 2002). To reduce skewness and stabilize variance, the ln(1 + delinquency) transformation is applied. Adding a value of one prior to the transformation avoids undefined values in observations with zero delinquency, without altering the relative order of the data (Chen and Roth, 2024). As a result, the skewness is reduced from 2.88 to 2.26 and the excess kurtosis from 10.15 to 6.14.

Mullahy & Norton (2024) question the use of the ln(1+y) transformation when the dependent variable has a high proportion of zeros and the additive constant is arbitrary, given that the estimates can be sensitive to the value of this constant. In the present study, the proportion of zero-valued observations is moderate (11.3%), the added unit is consistent with the natural scale of the indicator (delinquency = 0%), and the analytical interest focuses on the sign and statistical significance of the coefficients for comparison across models, rather than on the absolute magnitude of the retransformed marginal effects. In this context, the ln(1+y) transformation is a widely used methodological practice in the literature on financial intermediary panels (Louzis et al., 2012)

2.3. Econometric Specification

2.3.1. Reference Model

The baseline dynamic panel data model is specified as follows:

ln⁡(1+〖Mora〗_it )= ∝+β_1 ln⁡(1+〖Mora〗_(it-1) )+β_2 ln⁡〖(1+〖Mora〗_(it-2) )+∑_(K=1)^K▒〖γkX_kit 〗+δD_t+V_i+ϵ_it 〗

(1)

where i denotes the entity (COAC) and t the quarter; X_kit denotes the vector of internal financial determinants; D_t denotes the time dummy variables (D_Covid and D_post22); V_i denotes the unobserved entity-specific effect, which is constant over time; and ϵ_it denotes the idiosyncratic error term. Standard errors are estimated using the Driscoll and Kraay estimator, which is robust to heteroscedasticity, arbitrary-order autocorrelation, and contemporaneous cross-sectional dependence (Hoechle, 2007).

2.3.2. Model Variables

Table 1 presents the 12 model variables, their symbol, operational definition and applied treatment decisions, their role in the econometric model, and the corresponding research hypothesis.

2.4. Estimator Selection

The selection among the Pooled OLS, Fixed Effects (FE), and Random Effects (RE) estimators follows the sequential protocol recommended by Wooldridge (2020) for panels with large N and moderate T. In the first stage, the Breusch-Pagan LM test (Breusch & Pagan, 1980) assesses whether the variance of the unobserved entity-specific effect (V_i) is significantly different from zero. If H_0 (σ_v^2=0) is rejected, the Pooled OLS estimator is inconsistent, and an RE or FE model is required. In the second stage, the Hausman test (Hausman, 1978) evaluates whether the unobserved entity-specific effect (V_i) is correlated with the regressors. If H_0 is rejected, the RE estimator is inconsistent, and the FE estimator is preferred.

The RE model with Driscoll and Kraay standard errors is adopted as the primary specification for two reasons. First, the FE model removes all between-entity variation, which prevents estimating the coefficient of the time-invariant Segment 4 dummy variable, which is a key variable for this study. Second, the robustness of the results obtained with the RE model is evaluated using the Arellano–Bond (AB–2SLS) dynamic panel estimator (Arellano & Bond, 1991).

2.4.1. Arellano-Bond Estimator (AB-2SLS)

The inclusion of Y_(it-1) as a regressor introduces endogeneity by construction in the presence of unobserved individual effects (V_i), as the lagged values of Y are correlated with the error term in the model specified in levels (Arellano & Bond, 1991; Wooldridge, 2010). To address this endogeneity problem, the Arellano–Bond Difference GMM (AB-2SLS) model is estimated. This procedure (i) takes first differences of Equation [1], thereby eliminating (V_i) by construction, and (ii) instruments ∆Y_(it-1) using the lagged levels of ∆Y_(it-2), ∆Y_(it-3), together with financial efficiency at t-2 and t-3 as additional instruments. Roodman (2009) argues that internal instruments derived from lags of the endogenous variable itself are exogenous to the differenced error term by construction, provided that there is no serial correlation of order higher than one in the model specified in levels.

The validity of the AB-2SLS estimator requires verifying two conditions. The first is the validity of the instruments, evaluated using Sargan's over-identification test (Sargan, 1958; Hansen, 1982), whose H_0 states that all instruments are exogenous. A p-value greater than 0.05 indicates that there is no evidence to reject this H_0. The second condition is the absence of second-order serial autocorrelation AR(2), which tests the assumption of no serial correlation in the model specified in levels. In this context, a significant AR(1) statistic is expected as a mechanical consequence of the transformation into first differences, while the AR(2) test should not be significant, as this supports the validity of instruments based on second-order t-2 lags (Arellano & Bond, 1991; Blundell & Bond, 2023).

2.4.2. Robust Standard Errors: The Driscoll-Kraay Estimator

Panel data on financial intermediaries commonly exhibit heteroscedasticity, autocorrelation, and cross-sectional dependence. Heteroscedasticity reflects the high heterogeneity among entities, while cross-sectional dependence stems from exposure to common shocks, such as the regulatory framework, interest rate caps set by the Central Bank of Ecuador, and correlated provincial business cycles. Clustered standard errors at the entity level correct for autocorrelation within each entity, but not for contemporaneous cross-sectional dependence (Petersen, 2009).

In this context, the nonparametric standard errors poposed by Driscoll & Kraay (1998) provide consistent statistical inference in the presence of heteroscedasticity, arbitrary-order autocorrelation, and cross-sectional dependence across entities within the same period.

2.4.3. Analysis of Influential Observations

To assess the sensitivity of the estimators to outlier individual observations, three diagnostic criteria were applied to the residuals of the estimated RE model: (i) standardized residuals |ê| > 3 standard deviations as a standard threshold (Verbeek, 2017), (ii) Cook's distance greater than the conservative threshold of 4/N in at least 30% of the institution's observations, recommended for large N (Cook & Weisberg, 1986), and (iii) average leverage greater than 2k/N (from the intersection of the first two criterio), seven institutions with persistent influence, confirmed by multiple metrics, were identified.

It was decided to retain the full model (N = 5056), as the evidence from the formal tests supports this specification. First, the full model passed the RESET specification test, but failed after excluding the seven identified institutions. Second, the robustness of the AB-2SLS estimator deteriorated, as the AR(2) test became significant. Third, excluding these institutions did not reduce the kurtosis of the residuals. Fourth, several variables lost statistical significance. Finally, the F-statistic for parametric stability supports the relevance of the estimators. The individual analysis of these institutions showed that they mainly correspond to cooperatives undergoing supervised liquidation and trade cooperatives whose delinquency cycles respond to specific sectoral shocks.

2.5. Robustness Specification

The robustness of the results is assessed using seven alternative specifications: (R1) exclusion of observations with residuals greater than three standard deviations; (R2) 2020–2022 subsample; (R3) 2023–2025 subsample; (R4) Box-Cox transformation (λ = 0.249) as an alternative to ln(1+x); (R5) two-way fixed effects; (R6) separate estimates by segment; and (R7) strict Winsorization at the 2.5–97.5 percentiles. The central determinants of the model (microcredit concentration, provision coverage, and asset logarithm) maintain the expected sign and are statistically significant in all specifications, confirming the stability of the results.

2.6. Statement on the Use of Generative Artificial Intelligence

The econometric estimation, validation tests, and robustness specifications reported in this article were performed in Python 3.12. The estimation code was developed with the assistance of the Claude language model (Anthropic, 2024), specifically for structuring the data pipeline, implementing Driscoll-Kraay standard errors, and generating diagnostic statistical tests. All results were independently verified by the authors through code review, comparison with expected results according to econometric theory, and the comparison across alternative model specifications.}

3. Results

3.1. Model Selection

The Breusch-Pagan LM test rejects the null hypothesis of the absence of unobserved individual effects, thus ruling out the Pooled OLS specification. Meanwhile, the Hausman test (p < 0.001) indicates that the unobserved specific effects are correlated with the regressors, favoring the FE specification. However, this model does not allow for estimating the coefficient of the Segment 4 dummy variable, as it is time-invariant. For this reason, the RE model with Driscoll-Kraay standard errors (Model 2) is adopted as the primary specification, while the AB-2SLS estimator (Model 4) is used for robustness verification. The consistency in sign and statistical significance of the three main determinants (microcredit concentration, asset logarithm, and provision coverage) between Models 2 and 4 supports the robustness of the main findings.

Table 2. Panel Data Estimation Results (Dependent Variable: ln(1 + Delinquency)).

Dependent Variable: ln(1+Delinquency_it)	(1) Pooled OLS	(2) RE + DK	(3) FE	(4) AB- 2SLS
Constant	0.088***	0.088***	—	0.003**
Delinquency (t−1)	0.775***	0.775***	0.635***	0.627***
Delinquency (t−2)	0.092***	0.092**	0.055*	—
Microcredit Concentration	0.015***	0.015***	0.032***	0.030***
Log Total Assets	-0.006***	-0.006***	-0.013**	-0.089***
Loan Loss Provision Coverage	-0.034***	-0.034***	-0.044***	—
Liquidity	-0.008	-0.008**	0.002	0.016
Financial Efficiency (lag1)	0.003*	0.003	0.001	0.004
Leverage	-0.001	-0.001	—	0.005*
Segment 4 Dummy	0.011***	0.011***	—	—
D_COVID (2020)	-0.010***	-0.010***	-0.007***	-0.018***
D_post2022	0.007***	0.007**	0.011***	0.012***
R²	0.849	0.849	0.581	-0377
Observations	5,056	5,056	5,056	4,334
Entities	368	368	368	354
Durbin-Watson	—	1.848	1.897	2.836
Breusch-Pagan LM (p-value)	—	0.000	—	—
Hausman χ² (11)	—	67.54; p = 0,000	—	—
RESET (p-value)	—	0.791	—	—
Sargan (p-value)	—	—	—	0,828
AR(2) (p-value)	—	—	—	0,289
SE	Clustered	Driscoll-Kraay	Clustered	Robust

Note: ***p < 0.01; ** p < 0.05; *p < 0.10. Model (1) and Model (3): Clustered SE at entity level. Model (2): Driscoll-Kraay Kernel SE, robust to heteroskedasticity and cross-sectional dependence. Model (4): White heteroskedasticity-robust SE, estimated in first differences. The negative ${\mathrm{R}}^2$ reported for Modelo (4) is common in Difference GMM estimators and does not indicate model misspecification. Source: Authors' own elaboration.

Table 2. Panel Data Estimation Results (Dependent Variable: ln(1 + Delinquency)).

Dependent Variable: ln(1+Delinquency_it)	(1) Pooled OLS	(2) RE + DK	(3) FE	(4) AB- 2SLS
Constant	0.088***	0.088***	—	0.003**
Delinquency (t−1)	0.775***	0.775***	0.635***	0.627***
Delinquency (t−2)	0.092***	0.092**	0.055*	—
Microcredit Concentration	0.015***	0.015***	0.032***	0.030***
Log Total Assets	-0.006***	-0.006***	-0.013**	-0.089***
Loan Loss Provision Coverage	-0.034***	-0.034***	-0.044***	—
Liquidity	-0.008	-0.008**	0.002	0.016
Financial Efficiency (lag1)	0.003*	0.003	0.001	0.004
Leverage	-0.001	-0.001	—	0.005*
Segment 4 Dummy	0.011***	0.011***	—	—
D_COVID (2020)	-0.010***	-0.010***	-0.007***	-0.018***
D_post2022	0.007***	0.007**	0.011***	0.012***
R²	0.849	0.849	0.581	-0377
Observations	5,056	5,056	5,056	4,334
Entities	368	368	368	354
Durbin-Watson	—	1.848	1.897	2.836
Breusch-Pagan LM (p-value)	—	0.000	—	—
Hausman χ² (11)	—	67.54; p = 0,000	—	—
RESET (p-value)	—	0.791	—	—
Sargan (p-value)	—	—	—	0,828
AR(2) (p-value)	—	—	—	0,289
SE	Clustered	Driscoll-Kraay	Clustered	Robust

Note: ***p < 0.01; ** p < 0.05; *p < 0.10. Model (1) and Model (3): Clustered SE at entity level. Model (2): Driscoll-Kraay Kernel SE, robust to heteroskedasticity and cross-sectional dependence. Model (4): White heteroskedasticity-robust SE, estimated in first differences. The negative ${\mathrm{R}}^2$ reported for Modelo (4) is common in Difference GMM estimators and does not indicate model misspecification. Source: Authors' own elaboration.

3.2. Dynamic Persistence of Delinquency (H1)

Delinquency lags are the strongest predictors in the model. The first-order coefficient (β₁ = 0.775; p < 0.01) implies that 77.5% of a delinquency shock in one quarter is carried forward to the next; the second lag adds β₂ = 0.092 (p < 0.05), so the sum β₁ + β₂ = 0.867 indicates a slow adjustment dynamic in which 86.7% of any disturbance persists beyond the quarter in which it originates.

In the AB-2SLS estimator, which eliminates FE by first differences, the coefficient of the first lag is 0.627 (p < 0.01), confirming the robustness of the result. The evidence supports H1: lagged delinquency is the single most important determinant of current delinquency, indicating that interventions on problem portfolios produce slow improvements and supporting the use of preventive monitoring based on the trajectory of the delinquency indicator.

3.3. Microcredit Concentration (H2)

The microcredit concentration exhibits a positive and highly significant effect across all estimated specifications (β = 0.015; p < 0.01 in RE; 0.032 in FE; amd 0.030 in AB-2SLS), a result that remains robust across all seven robustness analyses. The evidence supports H2. The segment analysis reveals that the coefficient in Segment 5 (β = 0.024) is approximately three times higher than that of Segment 4 (β = 0.008), indicating that micro-cooperatives (with less capacity for geographic or product diversification) are significantly more vulnerable to sectoral portfolio concentration.

3.4. Institutional Size (H3)

The logarithm of total assets is negative and significant in all specifications (β = -0.006; p < 0.01 in RE; -0.013 in FE; and -0.089 in AB-2SLS), with stability in the robustness analyses, supporting H3 in the joint model. However, the protective effect of size is heterogeneous across segments; in Segment 5 the coefficient is negative and significant (β = -0.007; p < 0.01), whereas in Segment 4 it is practically null (β = -0.001; p = 0.385). Within micro-cooperatives, relative size is an indirect indicator of territorial rootedness and social capital, factors that improve informal credit monitoring. In Segment 4, asset differences within the USD 1–20 million range do not generate additional economies of scale in risk management.

3.5. Provision Coverage (H4)

The Loan Loss Provision Coverage exhibits the most consistent negative coefficient in the model, significant at the 1% level for both RE (β = -0.034) and FE (β = -0.044), and is the only determinant that remains statistically significant across all seven robustness specifications, thus supporting H4. The effect is 56% greater in Segment 5 (β = -0.046) than in Segment 4 ( β = -0.029), consistent with the greater relative exposure of micro-cooperatives to concentration risk. Proactive provisions partially compensate for the lower diversification capacity. The results are consistent with the early risk recognition hypothesis, according to which cooperatives that establish provisions early tend to have better portfolio quality in subsequent periods.

3.6. Liquidity, Efficiency, and Leverage (H5)

Liquidity is negative and significant in the RE model (β = -0.008; p < 0.05), although the effect disappears in the FE and AB-2SLS models, indicating that it operates primarily through between-entity variation rather than within-entity variation over time. Financial efficiency (lag1) and leverage do not show statistically significant effects in the RE model (p = 0.149 and p = 0.219, respectively). In the case of financial efficiency, these results suggest that lagged delinquency captures a large share of the accumulated deterioration in portfolio quality, which, according to the mismanagement hypothesis, is associated with inefficiency.

On the other hand, leverage shows heterogeneous behavior across segments: it presents a negative and statistically significant coefficient in Segment 4 (β = -0.002; p < 0.05), whereas in Segment 5 the coefficient is positive and not statistically significant. This pattern suggests that the relationship between leverage and delinquency may differ according to institutional size. Overall, the evidence provides partial support for H5.

3.7. Segment Analysis

The Chow test (p < 0.001) rejects the null hypothesis of coefficient homogeneity across segments, thus justifying the estimation of separate models. Table 3 reports the segment-specific estimation results.

The results show that the three main determinants of the joint model (microcredit concentration, loan loss provision coverage, and time effects) remain statistically significant in both segments, although with different magnitudes. In contrast, the logarithm of total assets is the only variable whose effect differs across segments: it has a statistically significant protective effect in Segment 5 (H3), but not in Segment 4. This result is consistent with the community relationship banking hypothesis, according to which the advantages associated with institutional size are concentrated in smaller cooperatives.

The time dummy variables confirm the cyclical nature of delinquency in both segments (H7). The payment deferral measures implemented in 2020 temporarily reduced recorded delinquency, whereas the tightening of conditions after 2022 was associated with an increase in this indicator. In the segment estimates, the D_COVID variable loses statistical significance due to its lower temporal variability. Finally, the trend reversal identified in the descriptive analysis (deterioration of Segment 4 and improvement of Segment 5 starting in 2022) is reflected by the positive and statistically significant coefficient of the Segment 4 dummy variable (β = 0.011; p < 0.01), thereby supporting H6.

3.8. Time Cycles and Structural Heterogeneity (H6 and H7)

The D_COVID variable shows a negative and statistically significant coefficient in the RE model (β = -0.010; p < 0.01) and AB-2SLS (β = -0.018), which is consistent with the payment deferral measures implemented by the SEPS in 2020. These measures temporarily reduced recorded delinquency, without reflecting a structural improvement in the member repayment capacity. Meanwhile, the D_post2022 variable shows a positive and statistically significant coefficient in RE (β = 0.007; p < 0.05) and AB-2SLS (β = 0.012), indicating a deterioration in portfolio quality associated with the tightening of the post-pandemic credit cycle. Taken together, these results support H7.

Moreover, the Segment 4 Dummy variable shows a positive and statistically significant coefficient in RE (β = 0.011; p < 0.01), which supports H6. Once internal financial determinants and time cycles are controlled for, cooperatives in Segment 4 show structurally higher delinquency levels than micro-cooperatives in Segment 5. This difference reflects unobserved factors related to institutional governance, the profile of borrowing members, and technical risk management capabilities—aspects that regulatory segmentation based solely on total asset size does not fully capture.

3.9. Robustness of the Results

The robustness of the results is assessed through three specifications: (R1) the exclusion of the 90 observations with residuals greater than three standard deviations; (R2) the 2020–2022 subsample; and (R4) the AB-2SLS estimator.

The three main determinants of this research (microcredit concentration, logarithm of total assets, and loan loss provision coverage) maintain the expected sign and statistical significance at the 1% level in all evaluated specifications. The only partial exception is liquidity, which loses significance in the models that rely on within-entity variation (FE and AB-2SLS), a result consistent with its interpretation as between-entity effect. This robustness pattern indicates that the main findings do not depend on a particular estimation technique or influential observations, but rather reflect stable structural relationships among the analyzed variables of the Ecuadorian EPS system.

Table 4. Stability of coefficients.

Variable.	RE-DK	1.	R2 Pre-2023	R4 AB-2SLS	Δ max
Delinquency (t−1)	0.775***	0.757***	0.748***	0.627***	0.026
Delinquency (t−2)	0.092**	0.108**	0.127**	—	0.035
Microcredit Concentration	0.015***	0.016***	0.007***	0.030***	0.009
Log Total Assets	-0.006***	-0.005***	-0.019***	-0.089***	0.005
Loan Loss Provision Coverage	-0.034***	-0.029***	-0.041***	—	0.004
Liquidity	-0.008**	-0.003	-0.003	0.016 ns	+0.006
Segment 4 Dummy	0.011***	0.011***	0.009**	—	0.002
D_COVID	-0.010***	0.010***	-0.010***	-0.018***	0.000
D_post2022	0.007**	0.006*	0.009***	0.012***	0.001

Note: ***p < 0.01; ** p < 0.05; *p < 0.10. Source: Authors' own elaboration.

Table 4. Stability of coefficients.

Variable.	RE-DK	1.	R2 Pre-2023	R4 AB-2SLS	Δ max
Delinquency (t−1)	0.775***	0.757***	0.748***	0.627***	0.026
Delinquency (t−2)	0.092**	0.108**	0.127**	—	0.035
Microcredit Concentration	0.015***	0.016***	0.007***	0.030***	0.009
Log Total Assets	-0.006***	-0.005***	-0.019***	-0.089***	0.005
Loan Loss Provision Coverage	-0.034***	-0.029***	-0.041***	—	0.004
Liquidity	-0.008**	-0.003	-0.003	0.016 ns	+0.006
Segment 4 Dummy	0.011***	0.011***	0.009**	—	0.002
D_COVID	-0.010***	0.010***	-0.010***	-0.018***	0.000
D_post2022	0.007**	0.006*	0.009***	0.012***	0.001

Note: ***p < 0.01; ** p < 0.05; *p < 0.10. Source: Authors' own elaboration.

4. Discussion

4.1. Delinquency as an Autoregressive Process: Dynamic Persistence and Supervisory Horizon

The most robust finding of this study is the high dynamic persistence of delinquency, evidenced by the coefficients of the first (β₁ = 0.775) and second (β₂ = 0.092) lags of the RE model for the COACs in the analyzed segments. This result is consistent with the literature on NPLs in panels of financial intermediaries. Ozili (2025) documents this pattern as one of the most methodologically consistent features in recent research on NPLs, and attributes it to the delays inherent in the processes of recovery, renegotiation, and loan write-off.

However, this article expands upon the existing evidence by showing that delinquency persistence is not homogeneous across segments. In Segment 4, the second lag coefficient is statistically significant (β₂ = 0.134; p < 0.01), while the second lag coefficient in Segment 5 (β₂ = 0.062; p > 0.10) does not reach statistical significance.

This result suggests that the portfolio deterioration process exhibits a longer inertia in medium-sized cooperatives, which could reflect, among other factors, greater operational complexity for loan restructuring and a greater relational distance between the institution and its borrowers—characteristics associated with larger institutions.

These results expand upon the recommendation of Enoch et al. (2021) regarding the importance of early intervention, showing that the first-order persistence in Segment 5 suggests that early interventions implemented during the first quarter may be sufficient to contain subsequent portfolio deterioration. On the other hand, in Segment 4, second-order inertia suggests that the effects of corrective actions tend to materialize more gradually. In this context, regulatory oversight could benefit from adopting longer alert horizons (two quarters or more), complementing indicators based on the historical trajectory of delinquency, in addition to the static cut-off points used in the SEPS monitoring systems (Noriega et al., 2023).

4.2. Microcredit Concentration: Endogenous and Differential Structural Risk by Segment

Microcredit concentration constitutes one of the main structural risk factors identified in the model, and its magnitude is substantially greater in Segment 5 (β = 0.024) than in Segment 4 (β = 0.008). This finding is theoretically consistent with the portfolio diversification principle, according to which a greater concentration of the portfolio in borrowers exposed to common risks (informal employment, volatile income, weak collateral) increases the probability of correlated defaults (Durango-Gutiérrez et al., 2023).

The evidence provided by Mahedi et al. (2026) on the spread of default within lending communities in India, as well as Olayiwola's (2026) findings on the vulnerability of microfinance portfolios to systemic shocks, support the interpretation that microcredit concentrations is associated not only with portfolio specialization but also with greater risk correlation. In the Ecuadorian context, this mechanism is especially relevant in Segment 5, where the territorial proximity of borrowing partners can amplify the synchronization of defaults in response to local shocks such as droughts, drops in agricultural prices, or episodes of social conflict. In this sense, the results highlight a paradox: while the specialization of the COACs in microcredit strengthens the financial inclusion of households and microenterprises operated by low-income informal borrowers, traditionally excluded from the banking system, it also increases the concentration of credit risk.

4.3. Institutional Size: The Community Relationship Banking Hypothesis

The protective effect of institutional size, measured by the logarithm of total assets, is statistically significant in Segment 5 (β = -0.007; p < 0.01), but not in Segment 4 (β = -0.001; p = 0.385), an asymmetric pattern with important theoretical implications for the debate between economies of scale in credit risk management and community relationship banking. Traditional theory states that larger institutions have more sophisticated risk assessment, monitoring, and management systems, which contributes to reducing delinquency (Salas et al., 2024).

However, this prediction assumes that an increase in institutional size translates into greater technical capacity across the entire asset distribution. The results obtained in Segment 5 suggest that a 1% increase in assets not only reflects greater operational capacity but also a stronger territorial rootedness, a broader network of members, and greater social capital. Taken together, these factors are associated with a 0.007% reduction in delinquency, a value consistent with the relationship banking hypothesis proposed by Petersen (2009) and Berger & Udell (2002), according to which informational proximity between lender and borrower reduces information asymmetries and strengthens informal credit monitoring.

In contrast, in Segment 4, differences in total asset size is concentrated within the USD 1 to 20 million range, indicating that relational advantages appear to have been largely exhausted. Consequently, variations in institutional size within the segment do not generate sufficient additional savings to reduce delinquency. This result suggests that regulatory segmentation based on total asset size partially, but not completely, captures the differences in business model and risk profile among cooperatives (Salas et al., 2024).

4.4. Provision Coverage: Early Recognition vs. Signaling Effect

Provision coverage is the only significant determinant at the 1% level in all robustness specifications, without exception. Furthermore, its negative effect on future delinquency is approximately 56% more pronounced in Segment 5 (β = -0.046) than in Segment 4 (β = -0.029). This result can be interpreted from two complementary and non-mutually exclusive frameworks. The first, consistent with the early risk recognition hypothesis, argues that cooperatives that proactively set aside provisions recognize expected losses before they formally materialize, allowing them to intervene early on the problem portfolio. This interpretation is consistent with the evidence presented by Kyiu & Tawiah (2025) in the context of IFRS 9, according to which earlier recognition of losses improves asset quality and the predictive capacity of provisions. The second framework interprets provisions as a sign of managerial discipline. Thus, entities that set aside higher provisions reflect a stronger risk management culture, which translates into better practices in loan origination and monitoring.

Both interpretations converge on a common regulatory implication: SEPS supervision could be strengthened by incorporating dynamic provision coverage indicators, complementing the point ratios used in its early warning systems. Likewise, differentiated incentives could be considered for Segment 5 cooperatives that maintain coverage levels above the minimum regulatory threshold.

4.5. Liquidity: A Factor of Institutional Resilience, Not Operational Management

Liquidity has a significant negative effect on delinquency in RE model (β = -0.008; p < 0.05). However, this relationship disappears in models based on within-entity variation (FE and AB-2SLS). This difference between the between-entity and within-entity effects indicates that structurally more liquid entities exhibit lower average delinquency levels, whereas temporal variations in liquidity within the same entity do not reduce delinquency contemporaneously.

This pattern is consistent with the interpretation of the deferral channel. A solid liquidity position provides the necessary operating margin to restructure troubled loans before they formally default. However, this mechanism operates through relatively stable institutional characteristics, such as treasury policy and funding structure, rather than quarterly fluctuations in the liquidity ratio. This is consistent with the evidence presented by Gelashvili et al. (2023), who document a non-linear relationship between liquidity and risk in entities of different sizes.

4.6. Efficiency, Leverage, and Autoregressive Dynamics

Lagged operating efficiency and leverage did not reach statistical significaance in the RE model (p = 0.149 and p = 0.219, respectively). However, the effect of leverage differs between segments: it shows a significant negative coefficient in Segment 4 (β = -0.002; p < 0.05), whereas in Segment 5 it did not reach significance.

These results do not imply that operating efficiency and leverage are irrelevant to portfolio quality. On the contrary, they suggest that their effect is absorbed by the lags of the dependent variable in a dynamic model, given that the historical trajectory of delinquency incorporates much of its cumulative influence. This is consistent with the evidence presented by Masanyiwa et al. (2022), who document nonlinear and temporally delayed relationships between efficiency and asset quality in microfinance.

The difference in the effect of leverage between segments is theoretically relevant. In Segment 4, higher leverage is associated with lower delinquency rates, which could reflect a greater capacity of the more capitalized entities to originate and manage higher-quality loans; in contrast, this relationship is not observed in Segment 5, possibly because the liability structure of micro-cooperatives responds primarily to the dynamics of member deposit mobilization, rather than to strategic decisions related to leverage.

4.7. Exogenous Shocks and Regulatory Distortion: A Critical Interpretation of Temporary Dummies

The interpretation of temporary dummies requires particular caution to avoid attributing real improvements in portfolio quality to transitory accounting or regulatory effects. The D_COVID dummy shows a significant negative coefficient in the main model specifications (β = -0.010 in RE and β = -0.018 in AB-2SLS). However, this result should not be interpreted as evidence of a reduction in credit risk during the pandemic.

The extraordinary measures implemented by the SEPS in 2020, including rescheduling, deferrals, and term extensions, allowed numerous loans to avoid formal classification as delinquent, without this implying an actual improvement in the borrower's repayment capacity. This is consistent with the evidence presented by Muhammad and Yao (2024), who document a similar pattern in the banking system, where accounting indicators of NPLs behaved divergently with respect to the underlying economic risk during the COVID-19 pandemic. Similarly, Alnabulsi et al. (2023) show, for countries in the MENA region, that the impact of the pandemic manifested itself heterogeneously across financial systems and depended, to a greater extent, on the specific characteristics of the institution and the quality of regulation, rather than on the health shock itself.

The dummy variable D_post2022, in turn, presents a positive and statistically significant coefficient (β = 0.007; p < 0.05), a result consistent with the exhaustion of extraordinary relief measures and the tightening of the post-pandemic credit cycle. Furthermore, the trend reversal identified in the descriptive analysis (deterioration in Segment 4 and improvement in Segment 5 starting in 2022) could be associated with differences in portfolio composition and the profile of borrowing partners in each segment. In Segment 4, where larger and longer-term loans predominate, the exhaustion of deferral mechanisms may have had a more significant impact. In contrast, in Segment 5, the structure of short-term microloans may have facilitated a faster adjustment to the new credit environment.

4.8. Structural Heterogeneity Among Segments: Implications for Regulatory Policy

The rejection of the hypothesis of coefficient homogeneity (Chow test p = 0.000) and the positive significance of the dummy variable Seg4 (β = 0.011; p < 0.01) confirm that cooperatives in Segments 4 and 5 constitute statistically distinct populations, not only in their average delinquency rates but also in the mechanisms that determine their risk dynamics. This finding has direct implications for the design of prudential policy for the Ecuadorian EPS system.

First, concentration limits in microcredit should be established differently for each segment, given that the same concentration level has an approximately three times greater effect on delinquency in Segment 5 than in Segment 4. Second, provisioning coverage thresholds and liquidity requirements should be calibrated considering the risk profile of each segment. Third, the heterogeneity observed in the effect of institutional size indicates that merger or acquisition policies for small cooperatives, aimed at generating economies of scale, should be evaluated with caution. In the context of community-based relationship banking, the growth of micro-cooperatives could be more beneficial than their integration into larger entities.

Finally, the differential behavior observed in the time dummy variables suggests a need to review the effectiveness of credit relief mechanisms. If the deferral measures implemented in 2020 temporarily reduced accounting delinquency without effectively improving the borrower's repayment capacity, and the tightening of the credit cycle after 2022 reversed that effect, prudential supervision could be strengthened by incorporating indicators that reflect the underlying quality of the portfolio, in addition to the accounting ratios used in its early warning systems. This approach is consistent with the recommendations of IFRS 9 on the early recognition of credit losses (Kyiu & Tawiah, 2025; Elashmawy & Kallunki, 2026).

4.9. Limitations and Future Research Agenda

This study has three main limitations that should be considered when interpreting the results. First, the RE specification, adopted to estimate the coefficient of the time-invariant segment dummy variable, is consistent under the assumption of no correlation between individual effects and regressors. While the AB-2SLS estimator confirms the main findings, future research could explore alternative identification strategies based on external instrumental variables to assess residual endogeneity. Second, the variables used are financial in nature and do not directly incorporate dimensions related to cooperative governance, local social capital, or membership structure, which, according to the cooperative banking literature, could be relevant determinants of delinquency (Beccalli & Viola, 2026; Manta et al., 2026). Third, the analyzed period (2020–2025) encompasses two exogenous shocks of considerable magnitude (the pandemic and the post-pandemic cycle), which limits the generalizability of the time-dummy coefficients to other contexts.

Based on this, future research could focus on: (i) incorporating variables related to cooperative governance (board size, board turnover, and proportion of women on the board) to analyze their interaction with the financial determinants of delinquency; (ii) extending the analysis to segments 1, 2, and 3 of the Ecuadorian EPS system, whose cooperatives operate with a higher asset range and whose risk dynamics may differ from the segments studied; and (iii) applying nonlinear threshold models to determine whether the protective effect of institutional size observed in Segment 5 only manifests itself above a certain minimum asset level.

5. Conclusions

Overall, the results confirm the seven hypotheses and provide robust evidence on the dynamic and heterogeneous nature of delinquency in COACs in Segments 4 and 5 of the Ecuadorian system of popular and solidarity economy (EPS). The main finding of the study, corroborated by the AB-2SLS estimator and seven robustness specifications, is that current delinquency is primarily determined by its own past trajectory (H1: β1 = 0.775; β2 = 0.092; p < 0.01), which confirms the autoregressive nature of portfolio deterioration and suggests that delinquency episodes are unlikely to correct themselves spontaneously in the short term. Although this pattern has been documented in other financial systems (Ozili, 2025), this finding takes on an additional dimension in the Ecuadorian cooperative context, showing that second-order inertia is statistically significant in Segment 4 but not in Segment 5, suggesting that prudential supervision could benefit from adjusting early warning horizons according to institutional size, rather than applying uniform thresholds to the entire EPS system.

The microcredit concentration is confirmed as the main structural determinant of risk in the model (H2), with an effect approximately three times greater in Segment 5 (β = 0.024) than in Segment 4 (β = 0.008). This result indicates that the productive specialization inherent in the microfinance model is not risk-neutral, but rather exposes smaller cooperatives to correlated defaults stemming from the socioeconomic homogeneity of their borrowers. This finding highlights a relevant paradox in the literature on financial inclusion: the very specialization that allows COACs to serve segments traditionally excluded from the formal banking system also constitutes the main source of their credit vulnerability.

Institutional size confirms a protective effect (H3), albeit an asymmetric one. This effect is particularly evident in Segment 5 (β = -0.007; p < 0.01), consistent with the community-based relationship banking hypothesis, which posits that asset growth is associated with greater territorial rootedness, higher social capital, and a greater capacity for informal credit monitoring. In contrast, the effect is not statistically significant in Segment 4 (β = -0.001; p = 0.385), where differences in asset levels do not appear to translate into additional economies of scale for risk management. This asymmetry modifies the conventional interpretation of the literature on institutional size and banking risk (Berger & Udell, 2002; Salas et al., 2024) and suggests that, in the case of micro-cooperatives, policies aimed at promoting mergers to achieve economies of scale could be counterproductive if they erode the relational capital that underpins portfolio quality.

Provision coverage is confirmed as the most robust prudential mechanism in the model (H4), being significant at the 1% level in all robustness specifications without exception, with an effect approximately 56% stronger in Segment 5 (β = -0.046) than in Segment 4 (β = -0.029). This result is consistent with the expected loss paradigm underlying IFRS 9 and suggests that proactive provisioning is, in the Ecuadorian cooperative context, a more effective risk management tool than simply expanding the balance sheet.

Liquidity partially confirms H5. Its protective effect is significant in the between dimension (differences between entities), but disappears in the within dimension (FE and AB-2SLS), indicating that liquidity operates as a deferral channel rather than a structural determinant of delinquency, providing room to restructure the problem portfolio without, on its own, correcting the underlying causes of deterioration.

The evidence confirms the structural heterogeneity between segments (H6). Once financial determinants and time cycles are controlled for, cooperatives in Segment 4 exhibit structurally higher levels of delinquency than those in Segment 5 (β = 0.011; p < 0.01), a result further supported by the rejection of the Chow test (F(10,5036) = 17.79; p = 0.001). Furthermore, H7 is confirmed: the regulatory payment deferral implemented by SEPS in 2020 artificially reduced accounting delinquency without reflecting a genuine improvement in members' ability to pay, while the increase observed from 2022 onward demonstrates the exhaustion of these measures and a tightening of the post-pandemic credit cycle.

The confirmation of the seven hypotheses allows us to conclude that delinquency in Ecuadorian COACs in Segments 4 and 5 is not due to a single factor, but rather to a dynamic, structural risk architecture that varies according to institutional size, in which time inertia, portfolio concentration, provisioning discipline, and exogenous cycles interact heterogeneously across segments. This evidence supports, from both a theoretical and applied perspective, the need to migrate from supervision based on static delinquency ratios to a prudential scheme differentiated by segment, which incorporates the historical trajectory of the portfolio, the limits of sectoral concentration and the incentives for proactive provisioning as central pillars of risk monitoring in the Ecuadorian system of popular and solidarity economy.