Gender as a Risk Factor: A Test of Gender-Neutral Pricing in Lithuania’s P2P Market

Abstract

European Union legislation, particularly Council Directive 2004/113/EC, mandates gender neutrality in credit scoring to prevent discrimination. However, this creates a regulatory paradox if gender is a statistically relevant predictor of default risk. This study investigates this “fairness-through-unawareness” approach by empirically testing for systematic mispricing. We employ a twofold econometric analysis on a dataset of consumer loans from a Lithuanian peer-to-peer platform. After data preparation for the regression, the sample consists of 9707 loans. First, logistic regression is used to model actual default risk, controlling for credit rating, age, loan amount, and education. Second, Ordinary Least Squares (OLS) regression is used to model the interest rate set by the platform. The Logit model finds that gender is a highly significant predictor of default (p < 0.001), with male borrowers associated with a higher probability of default. Conversely, the OLS model finds that gender is not a statistically significant factor in loan pricing (p = 0.263), confirming the platform’s compliance with EU law. The findings empirically demonstrate the regulatory paradox: the legally compliant, gender-blind pricing model fails to account for a significant risk differential. This leads to systematic risk mispricing and an implicit cross-subsidy from lower-risk female borrowers to higher-risk male counterparts, highlighting a critical tension between regulatory intent and outcome fairness. The analysis is limited to observed loan-level characteristics; it does not incorporate household composition or the internal structure of the platform’s proprietary scoring model.

1. Introduction

The financial landscape across the European Union (EU) is undergoing significant transformation driven by sophisticated algorithmic credit scoring systems. Employed across traditional banking and P2P lending, these systems promise to broaden financial inclusion Shi et al. (2022). By leveraging vast datasets, lenders aim to predict borrower default risk with greater accuracy. However, this technological advancement is not without peril; these algorithms can inadvertently embed or amplify existing societal biases, leading to discriminatory outcomes Bartlett et al. (2022).

At the heart of this challenge lies a significant conflict within the EU context. On one hand, a robust legal framework, anchored in Article 21 of the Charter of Fundamental Rights, explicitly prohibits discrimination based on sex European Union (2000). This principle is operationalized through Council Directive 2004/113/EC The Council of the European Union (2004). The interpretation of this directive by the Court of Justice of the European Union (CJEU) in the pivotal Test-Achats case cemented a mandate for gender neutrality, forbidding the use of gender as a determining factor in risk assessment and pricing for insurance and related financial services Skouris et al. (2010).

On the other hand, a growing body of empirical evidence challenges the assumption that gender is irrelevant to credit risk. Studies analyzing P2P lending platforms consistently indicate that gender can be a statistically significant predictor of creditworthiness. Specifically, multiple analyses suggest female borrowers, on average, exhibit lower default rates compared to male counterparts Aliano et al. (2023); Chen et al. (2019); Pope and Sydnor (2011).

This juxtaposition creates a regulatory paradox. The legal imperative for “fairness-through-unawareness”—that is, achieving fairness by simply omitting a variable—may inadvertently generate unintended negative consequences Hardt et al. (2016). If gender is a statistically relevant risk factor, prohibiting its use compels lenders to pool inherently different risk groups. This can lead to a systematic mispricing of risk, where the lower-risk group (e.g., women) is effectively overcharged to subsidize the higher-risk group (e.g., men) Caire and Fernandez Vidal (2024). This phenomenon raises profound questions about whether the current regulatory focus on input variables adequately addresses the complexity of achieving equitable outcomes Bartlett et al. (2022); Corbett-Davies et al. (2018).

While previous studies have descriptively highlighted these gender-based differences, many have lacked the robust econometric analysis required to separate the effect of gender from other confounding variables (like age, income, or existing credit score) and to empirically test the “paradox” itself. This paper aims to fill this gap by moving beyond descriptive statistics to econometrically model and quantify this mispricing.

Using a dataset from a Lithuanian P2P platform, we test two formal hypotheses:

H1: 

Gender is a statistically significant predictor of actual default risk, with male borrowers exhibiting a higher probability of default even when controlling for standardized risk factors (e.g., credit rating, age, and loan amount).

H2: 

The legally mandated exclusion of gender from the platform’s credit scoring model leads to a systematic mispricing of risk, where the platform’s final interest rates do not reflect the actual, higher risk associated with male borrowers.

By testing H1, we establish the statistical reality of gendered risk. By testing H2, we confirm the platform’s compliance with the law. Demonstrating both simultaneously provides a clear, empirical validation of the regulatory paradox and its economic consequences.

2. Literature Review

2.1. The EU Legal Framework Prohibiting Gender Discrimination

The European Union’s commitment to equality forms a cornerstone of its legal identity. This is firmly rooted in EU primary law, particularly Article 21 of the Charter of Fundamental Rights of the European Union, which explicitly states that “Any discrimination based on any ground such as sex… shall be prohibited” European Union (2000). This article establishes broad protection and is binding on EU institutions and Member States.

Recognizing that gender equality needed to extend beyond the workplace, the EU adopted Council Directive 2004/113/EC (the “Goods and Services Directive”) The Council of the European Union (2004). Article 5(1) of this Directive established the core “unisex rule,” stating that “the use of sex as a factor in the calculation of premiums and benefits for the purposes of insurance and related financial services shall not result in differences in individuals’ premiums and benefits”. While the Directive initially allowed derogations, this provision was declared invalid by the CJEU in the landmark Test-Achats ruling (Case C-236/09) Skouris et al. (2010). The Court reasoned that allowing an indefinite exemption undermined the Directive’s objective. This ruling effectively eliminated the legal basis for using gender directly in risk calculation for new contracts across the EU. Lithuania has transposed these principles into national law, notably through the Law on Equal Opportunities for Women and Men Law of the Republic of Lithuania (2017).

2.2. Empirical Evidence on Gender and Creditworthiness

Despite the clear legal prohibition, a significant body of empirical research indicates that gender often emerges as a statistically significant variable when predicting credit default risk Aliano et al. (2023); Liu and Mona (2025). This evidence is particularly prominent in the P2P lending sector. Studies utilizing data from major European P2P platforms like Bondora consistently find that female borrowers tend to exhibit a lower probability of defaulting compared to male borrowers, holding other factors constant Aliano et al. (2023); Liu and Mona (2025).

Similar findings have been observed in other P2P markets, such as China Chen et al. (2019). However, these findings are not universal; research on the German P2P platform Smava found no significant gender effect on funding success Barasinska and Schäfer (2014); Polena and Regner (2018), and a study on a Chinese platform found no significant gender effect on default probability Lingnan (2019), underscoring potential context dependency Barasinska and Schäfer (2010).

Research specifically examining Lithuanian P2P platforms found gender was not a statistically significant factor in determining interest rates charged, though this relates to pricing, not default risk itself Gaigalienė and Česnys (2018); Golovkina (2021). Beyond P2P lending, academic studies also point towards gender’s statistical relevance. A key study analyzing European car loans explicitly found gender to be statistically significant in predicting default within a standard credit scoring model Andreeva and Matuszyk (2019). Other research indicates female loan officers may achieve lower loan delinquency rates Moltalvo and Reynal-Querol (2020) and suggests potential links between female leadership/boardroom diversity and decreased firm-level credit risk Aguir et al. (2023); Kinateder et al. (2021), although evidence on individual risk aversion is mixed and context-dependent Aliano et al. (2023); Lobão (2024).

Recent scholarship on machine learning and fairness in financial models supports this concern. In Nwafor et al. (2024), it was argued that proxy variables often replicate hidden gender effects, while it was shown in He and Tao (2025) that, even when protected attributes are excluded, bias can persist unless explicitly constrained.

2.3. Limitations of Age and Income as Explanatory Variables

Credit risk modeling has long relied on demographic and financial variables such as age and income as proxies for borrower stability Crook et al. (2007). It was shown in Jammalamadaka and Itapu (2023) that machine learning credit models often mask structural biases under traditional risk factors like income. Similarly, in De Andrés et al. (2021), by analyzing Spanish bank loan data, it was found that income-related risk assessments systematically undervalue female borrowers.

Furthermore, in a systematic review Dastile and Celik (2024), it was concluded that older borrowers are generally less risky. The authors in He and Tao (2025) likewise observed that credit scoring algorithms often learn that older borrowers are less likely to default. These findings suggest that age and income do not fully explain gender-based variation in credit risk.

2.4. Challenges: Proxy Discrimination and Alternative Data

The legal mandate to exclude gender encounters significant practical challenges due to “proxy discrimination” Schwarcz and Prince (2020). Even when gender is explicitly removed, its influence can persist through other legally permissible variables correlated with it (e.g., occupation, income type, spending patterns) Morais Maceira (2017). Algorithms trained on large datasets can identify these correlations, potentially perpetuating historical biases and circumventing the intended goal of equal treatment Weerts et al. (2024).

The financial industry is increasingly exploring alternative data sources (e.g., utility payments, digital footprints) Alliance for Financial Inclusion (2025); Wang (2024). Proponents argue that this can promote financial inclusion. However, the European Banking Authority (EBA) notes challenges regarding data reliability, explainability, and governance. Furthermore, alternative data introduces ethical concerns, including privacy risks and the potential amplification of societal biases.

2.5. Algorithmic Fairness Concepts

The term “fairness” in the context of algorithms lacks a single, universally accepted definition Bono et al. (2021); Das et al. (2022); Liu (2024). Instead, a variety of mathematical formulations exist, often reflecting different ethical priorities and sometimes proving mutually incompatible Wu (2023). Common group fairness metrics include Demographic Parity, Equal Opportunity, and Predictive Parity.

Recent systematic reviews underscore the limitations of “Fairness Through Unawareness”—the assumption that excluding sensitive attributes like gender inherently ensures impartiality. A comprehensive synthesis by de Castro Vieira et al. (2025) observes that this approach is rarely used in academic fairness research, reinforcing the consensus that structural bias persists even when protected variables are omitted. This study contributes to filling a research gap by employing a real-world, proprietary dataset from a Lithuanian P2P platform to test this “unawareness” approach.

2.6. Ethical Considerations and Societal Impact

The debate surrounding gender inclusion extends into ethics and societal impact. The primary ethical argument against using gender aligns with non-discrimination law, emphasizing assessment based on individual merit, not group characteristics Andreeva and Matuszyk (2019). Incorporating gender risks reinforcing harmful stereotypes and perpetuating historical discrimination Morais Maceira (2017).

Conversely, arguments for controlled inclusion stem from a focus on actual outcomes and statistical fairness. Excluding gender when it is a statistically significant differentiator can lead to an unfair cross-subsidy, where the lower-risk group (females) effectively subsidizes the higher-risk group (males) Caire and Fernandez Vidal (2024). From an outcome-oriented perspective, this result can be framed as unfair. The decision impacts financial inclusion European Institute for Gender Equality (2023) and public trust in credit systems Bono et al. (2021).

3. Results

3.1. Descriptive Statistics

Table 1 presents the summary statistics for all variables used in the final regression analyses, based on the cleaned sample of 9707 observations. The mean default rate in this sample is 7.9%. Male borrowers (Gender_Numerical = 1) represent 55.1% of the sample. The average borrower is 37 years old with an average credit rating score of 10.7.

Table 1. Summary statistics for model variables (N = 9707).

Variable	N	Mean	Std. Dev.	Min	Max
Loan_status (Default)	9707	0.079	0.270	0	1
Loan_Interest (%)	9707	19.35	4.88	5.0	50.0
Gender_Numerical (1 = Male)	9707	0.551	0.497	0	1
Credit_Rating_Numerical	9707	10.70	2.40	1	16
Age	9707	37.15	11.31	18	75
Loan_Amount (EUR)	9707	4213	2750	50	25,000
Duration_of_Income_Num.	9707	3.21	1.80	0	6
Education_Basic	9707	0.016	0.125	0	1
Education_Professional	9707	0.045	0.208	0	1
Education_Highschool	9707	0.038	0.191	0	1

3.2. Exploratory Data Analysis (EDA)

Before proceeding to the formal econometric analysis, an exploratory data analysis (EDA) of the full dataset (N = 10,893) provides preliminary, non-parametric evidence for our hypotheses.

3.2.1. EDA for H1: Gender and Default Risk

To find preliminary support for H1 (that gender is a predictor of risk), we first examine the overall default rates, which are based on 4872 loans to females and 6018 loans to males. As shown in Table 2, a clear disparity exists.

Table 2. Default rate by gender (full dataset, N = 10,893).

Female Default Rate (%)	Male Default Rate (%)
5.79	9.64

The data indicates that male borrowers default approximately 1.7 times more frequently than female borrowers. However, this could be confounded by other variables, such as age. The average age for female borrowers was 40.2 years, while for male borrowers it was 34.7 years. Given that older clients tend to be less risky Dastile and Celik (2024); He and Tao (2025), this age difference could explain the lower default rate for women.

To test this, we stratified the data by age group, as shown in Table 3.

Table 3. Default rates by age and gender (%) (full dataset, N = 10,893).

Age Group	Female Default Rate (%)	Male Default Rate (%)
18–29	7.48	9.51
30–39	7.89	10.04
40–49	4.23	9.82
50–59	3.30	7.62
60+	1.85	9.68

Contrary to the confounding hypothesis, the female default rate was significantly lower in almost all age groups, and the gap widened in older cohorts. This supports H1, suggesting gender is a factor independent of age.

Next, we examine if the disparity persists within the platform’s own standardized credit rating categories. If the rating system were perfectly capturing all risk, default rates should be similar for men and women within the same rating band.

Table 4 reveals a consistent pattern: female borrowers default less frequently than male borrowers in nearly every rating category. The gap widens substantially in the higher-risk bands (e.g., C3–D3). This provides strong preliminary evidence for H1, indicating that the gender-neutral rating system is not fully capturing systematic, gender-related differences in risk Bartlett et al. (2022); Garcia et al. (2024).

Table 4. Default rates by credit rating and gender (%) (full dataset, N = 10,893).

Credit Rating	Female Default Rate (%)	Male Default Rate (%)
A1	0.00	1.96
A2	0.55	2.82
A3	2.05	3.57
B1	2.51	4.26
B2	3.46	6.07
B3	6.72	8.28
C1	8.62	10.31
C2	9.39	13.03
C3	8.04	15.16
D1	8.20	12.31
D2	8.86	15.08
D3	12.86	27.33
E1	33.33	9.52
E2	0.00	14.29
E3	0.00	40.00

Finally, we assess the explanatory power of income. To account for inflationary effects over the 12-year period, net salaries were adjusted to their 2025 equivalent values. Table 5 compares default rates within these normalized income brackets.

Table 5. Default rates by adjusted net salary and gender (%) (full dataset, N = 10,893).

Normalized Net Salary Bracket (EUR)	Female Default Rate (%)	Male Default Rate (%)
>1600	7.90	10.98
1300–1600	4.77	8.31
1000–1299	6.66	9.25
700–999	7.11	10.51
500–699	6.88	8.30
<500	4.07	6.28

Despite having lower average net salaries (801 EUR for females vs. 994 EUR for males), female borrowers exhibited lower default rates in every income segment. This challenges findings that income-related risk assessments systematically undervalue female borrowers De Andrés et al. (2021). This analysis confirms that neither age nor income fully explains the observed gender difference in default rates, providing strong preliminary support for H1.

3.2.2. EDA for H2: Gender and Loan Pricing

Having established preliminary evidence for H1 (gender as a risk factor), we now conduct a preliminary test for H2 (gender-neutral pricing). The most direct way to observe the paradox is to look at the risk–price relationship on the platform. Table 6 shows the actual default rates for men and women who were grouped into the same interest rate bracket.

Table 6. Default rates by interest rate and gender (%).

Count of Loans	Interest Rate (%)	Default Rate (Women) (%)	Default Rate (Men) (%)
58	8	0%	0%
50	9	4%	4%
128	10	0%	5%
190	11	3%	11%
493	12	5%	8%
339	13	8%	13%
711	14	4%	6%
538	15	5%	8%
942	16	3%	5%
196	17	11%	12%
1221	18	6%	8%
205	19	7%	7%
947	20	6%	9%
397	21	6%	8%
1094	22	7%	9%
93	23	0%	14%
601	24	10%	13%
747	25	6%	8%

Table 6 provides the clearest preliminary evidence for H2. In almost every single interest rate bracket offered by the platform, the default rate for female borrowers is lower—often substantially—than the default rate for male borrowers. For example, in the 11% bracket, men default at over 3.6 times the rate of women. In the 23% bracket, women default at 0% while men default at 14%. This demonstrates that men and women are being pooled together and offered the same price, despite exhibiting different levels of risk.

This directly suggests a systematic mispricing. To confirm that this pricing is indeed neutral, we also examine the average interest rates charged, stratified by age, as shown in Table 7.

Table 7. Interest rates by age and gender (full dataset, N = 10,893).

Age Group	Average Interest Rate of Females	Average Interest Rate of Males
18–29	21.88	22.00
30–39	18.98	18.90
40–49	17.88	18.47
50–59	17.62	17.82
60+	17.32	17.45

As Table 7 shows, the interest rates are nearly identical across all age groups for both genders. This confirms the platform is applying a gender-neutral pricing policy. When combined with Table 6, this is the crux of the regulatory paradox: the platform correctly follows the law by offering gender-neutral prices, but in doing so, it ignores a significant, observable difference in risk, leading to a cross-subsidy from lower-risk female borrowers to higher-risk male borrowers.

This EDA, while insightful, relies on stratified tables. To formally test these hypotheses while controlling for all variables simultaneously, we now turn to multivariate regression.

3.3. Econometric Analysis

The following analysis is performed on the cleaned, regression-ready sample of N = 9707 observations.

3.3.1. Econometric H1 Results: Gender as a Predictor of Actual Default

Table 8 presents the results of the Logit regression model predicting loan default (Loan_status). The model is statistically significant as a whole (LLR p-value: 4.321 $\times {10}^{-109}$), indicating it has strong predictive power.

Table 8. Logit regression results for loan default (H1) (N = 9707).

Dep. Var: Loan_status	Coef	Std Err	z	p > |z|	[0.025]	[0.975]
Gender_Numerical	0.3145	0.084	3.739	0.000	0.150	0.479
const	0.2129	0.242	0.881	0.379	−0.261	0.687
Duration_of_Income_Num.	−0.2392	0.037	−6.548	0.000	−0.311	−0.168
Loan_Amount	0.0001	8.01 × 10−6	13.896	0.000	0.000	0.000
Age	0.0023	0.004	0.584	0.559	−0.005	0.010
Credit_Rating_Numerical	−0.2886	0.019	−15.376	0.000	−0.325	−0.252
Education_Basic	0.5719	0.173	3.311	0.001	0.233	0.910
Education_Professional	0.2234	0.109	2.052	0.040	0.010	0.437
Education_Highschool	0.3998	0.100	4.004	0.000	0.204	0.595
No. Observations: 9707
Log-Likelihood: −2462.1
Pseudo R-sq.: 0.09700

The results strongly support H1. The variable Gender_Numerical has a positive coefficient (0.3145) and is highly statistically significant (p = 0.000).

• Interpretation: Since Gender_Numerical is coded as 1 for Male and 0 for Female, this positive coefficient indicates that male borrowers have a significantly higher likelihood of defaulting than female borrowers. This is true even when controlling forall other variables in the model simultaneously.
• Odds Ratio: The odds ratio for gender is $e^{0.3145}=1.37$. This means that the odds of a male borrower defaulting are approximately 37% higher than for a female borrower with an identical profile.

Connecting H1 to H2

The confirmation of H1, both in the EDA and the formal regression, is the first half of the regulatory paradox. We have established that gender is a statistically significant and economically meaningful predictor of default risk.

This finding necessitates the test of H2. If the platform’s pricing model were purely risk-based and unconstrained by law, we would expect it to account for this 37% higher default risk associated with male borrowers, likely by charging them a higher interest rate. The EDA (Table 6 and Table 7) already suggested this is not happening. H2 now formally tests this within the multivariate regression framework.

3.3.2. Econometric H2 Results: Gender as a Factor in Loan Pricing

Table 9 presents the results of the OLS regression model predicting the Loan_Interest rate. This model explains a large portion of the variance in interest rates, with an R-squared of 0.735.

Table 9. OLS regression results for loan interest rate (H2) (N = 9707).

Dep. Var: Loan_Interest	Coef	Std Err	t	p > |t|	[0.025]	[0.975]
Gender_Numerical	0.0814	0.073	1.120	0.263	−0.061	0.224
const	46.4888	0.220	210.995	0.000	46.057	46.921
Duration_of_Income_Num.	0.1179	0.031	3.751	0.000	0.056	0.180
Loan_Amount	−0.0005	8.34 × ${10}^{-6}$	−58.552	0.000	−0.001	−0.000
Age	−0.0258	0.003	−7.601	0.000	−0.033	−0.019
Credit_Rating_Numerical	−2.1573	0.016	−131.771	0.000	−2.189	−2.125
Education_Basic	−0.3951	0.174	−2.265	0.024	−0.737	−0.053
Education_Professional	−0.2855	0.092	−3.108	0.002	−0.466	−0.105
Education_Highschool	−0.4966	0.086	−5.801	0.000	−0.664	−0.329
No. Observations: 9707
R-squared: 0.735
F-statistic: 3364.

The results strongly support H2. The variable Gender_Numerical has a p-value of 0.263, which is well above the 0.05 threshold for statistical significance.

• Interpretation: This finding confirms what the EDA suggested: gender has no statistically significant effect on the interest rate charged by the platform. The platform’s pricing model is, in effect, gender-blind.
• Model Compliance: This result is the expected outcome for a lender that is compliantwith EU non-discrimination law (Directive 2004/113/EC).

4. Discussion

This study’s findings, supported first by exploratory data analysis and confirmed with a robust econometric framework, empirically validate the regulatory paradox. Our Logit model (H1) confirms a statistical reality: gender is a significant predictor of default risk in this dataset, with male borrowers being approximately 1.37 times more likely to default than their female counterparts, even after controlling for the platform’s own credit rating (Table 8).

Simultaneously, our OLS model (H2) confirms a legal reality: the platform’s pricing model is compliant with EU law, as gender is not a statistically significant factor in determining interest rates (Table 9).

The combination of these two findings is the critical contribution of this paper. The analysis of the platform’s historical loan data shows that gender is a statistically significant risk factor, whereas the platform’s pricing remains gender-neutral in line with legal requirements. Under such a regulatory constraint, any pricing model that excludes gender will, by construction, pool borrowers with different risk profiles, which in turn produces a systematic mispricing of risk. Because the platform’s legally constrained, gender-neutral pricing effectively pools higher-risk (male) and lower-risk (female) borrowers, the observed outcome is an average price applied to groups with different default behaviour. In this setting, lower-risk female borrowers appear to pay an interest rate that is high relative to their observed default probability, while higher-risk male borrowers pay a rate that is low relative to theirs. This pattern is consistent with an implicit cross-subsidy between the two groups, although we do not quantify the exact monetary magnitude of this effect in the present study. This outcome, while legally compliant, is arguably “unfair” from an outcome-oriented perspective.

Limitations and Future Research

The conclusions of this study must be viewed within the context of its limitations:

• Scope: The analysis is based on data from a single P2P platform in a single EU member state (Lithuania). The findings, while robust for this dataset, may not be generalizable to all P2P platforms or traditional banking sectors across the EU.
• Variable Omission: The model does not control for all possible risk factors.
• Household structure and shared liabilities: The dataset does not contain information on household composition (e.g., dependants, shared income, partner income, or household-level liabilities). As a result, part of the observed gender difference in default risk may be attributable to unobserved differences in household structure rather than gender alone.
• Proxies: This study did not test for proxy discrimination. While the platform does not use gender directly, its models may be using other variables (e.g., occupation, income source) that are highly correlated with gender, potentially perpetuating bias Morais Maceira (2017); Weerts et al. (2024).

This study’s findings establish a clear, static mispricing of risk. Future research will build directly on this by pursuing two advanced lines of inquiry. First, to quantify the precise economic cost of this regulatory paradox, we will develop a counterfactual risk-pricing model. This involves building separate, gender-aware risk models to estimate the “fair” interest rates that would have been assigned to female borrowers based on their actual, lower default risk. By comparing these counterfactual rates to the actual rates paid, we will calculate the aggregate overpayment in Euros, thereby quantifying the total monetary value of the implicit cross-subsidy.

Second, we will move beyond the static, binary (default/no default) analysis of the Logit model to a dynamic analysis of default timing. We will employ survival analysis, such as a Cox proportional hazards model, to investigate whether gender is a significant covariate in determining the time-to-default. This will reveal if gender not only affects if a loan defaults, but also how quickly it defaults, adding a critical temporal dimension to the risk profile.

5. Data and Methodology

5.1. Data and Variables

This study utilizes a proprietary, anonymized dataset provided by a major Lithuanian peer-to-peer (P2P) lending platform. The platform granted access to this dataset for research purposes only. The initial data comprised detailed records of 10,893 consumer loans issued between 2014 and 2025. After listwise deletion of observations with missing data for any of the model variables, the final sample used for regression analysis consists of 9707 loans.

The platform’s credit ratings (A1 to E3) follow a standardized scoring framework, as described in Taujanskaitė and Milčius (2022). In line with reviewer recommendations, Table 10 provides the mapping of these categories to their numerical and linguistic values.

Table 10. Frequency distribution and description of credit rating.

Categorical	Numerical Credit Rating	Linguistic Risk	Count of Loans
Credit Rating	(Used in Model)	Variable	(in N = 10,893 Sample)
A1	16	Lowest Risk	199
A2	15	Very Low Risk	360
A3	14	Low Risk	727
B1	13	Good	1266
B2	12	Good	1525
B3	11	Average-Good	1784
C1	10	Average	1796
C2	9	Average	1396
C3	8	Average-Risk	833
D1	7	High Risk	508
D2	6	High Risk	205
D3	5	High Risk	242
E1	4	Very High Risk	30
E2	3	Very High Risk	14
E3	2	Highest Risk	7
F	1	Highest Risk	1

All variables used in the regression models are described in Table 11. These variables were included in the original dataset provided by the lending platform.

Table 11. Description of variables used in regression models (N = 9707).

Variable Name	Description	Type
Dependent Variables
Loan_status	The final outcome of the loan.	Binary: 1 = Defaulted, 0 = Fully Repaid
Loan_Interest	The annual interest rate (%) charged.	Continuous
Key Independent Variable
Gender_Numerical	The borrower’s gender.	Binary: 1 = Male, 0 = Female
Control Variables
Credit_Rating_Num.	Platform’s internal risk rating.	Ordinal: 1 (Highest Risk) to 16 (Lowest Risk)
Age	Borrower’s age in years.	Continuous
Loan_Amount	Principal amount of the loan in EUR.	Continuous
Duration_of_Income_Num.	Duration of borrower’s stated income.	Continuous (as treated in model)
Education_Basic	Dummy for basic education.	Binary: 1 = Yes, 0 = No
Education_Professional	Dummy for professional education.	Binary: 1 = Yes, 0 = No
Education_Highschool	Dummy for secondary education.	Binary: 1 = Yes, 0 = No
const	The model intercept.	-

5.2. Software and Data Transformations

The data analysis and econometric modeling for this study were conducted using the Python programming language (version 3.12). The following open-source libraries were essential to the analysis:

• pandas: Used for data loading from an Excel file, data manipulation, filtering, and analysis.
• numpy: Employed for numerical operations, particularly for checking for infinite or missing (NaN) values during data cleaning.
• statsmodels.api: The primary library used for fitting the logistic (Logit) and linear (OLS) regression models and generating statistical summaries.

To prepare the raw data for the regression models (N = 9707), several transformations were performed:

• Data Loading and Filtering: Rows with zero or NaN values in “Individual _monthly_ income” or “Duration_of_Income_Numerical” were removed to ensure the integrity of the regression sample.
• Binary Encoding (Dependent Variable): The “Loan_status” column was converted into a binary numerical variable (0 for FULLY_PAID, 1 for DEFAULT).
• Numerical Encoding (Features): The “Gender” column was mapped to a single numerical variable (0 for Female, 1 for Male). The categorical “Duration_of_Income” column was similarly mapped to its numerical equivalent (“Duration_of_Income_Numerical”).
• One-Hot Encoding: To include education level in the model, the categorical “Education” column was converted into dummy variables. The resulting boolean columns were then converted to integers (0 and 1s) to be used as regressors.

Note that the “Equalized_Individual_monthly_income” variable, calculated using a linear regression formula based on the “Metai” (Year) column, was used only for the stratified tables in the Exploratory Data Analysis (Section 3.2) and not in the final regression models.

5.3. Econometric Framework

To empirically test the regulatory paradox, we employ a two-part econometric framework. This formal modeling approach provides the mathematical formulation for our hypotheses, addressing key reviewer feedback.

5.3.1. Testing H1: Gender and Actual Default Risk

To test H1—that gender is a statistically significant predictor of actual default risk—we use a logistic (Logit) regression model. This is a standard approach for binary dependent variables like loan default Crook et al. (2007). The model estimates the probability of default for a borrower i. This approach aligns with the methodology used in Cozarenco and Szafarz (2022) which detects structural biases.

$$P(\mathtt{Loan}\_{\mathtt{status}}_i=1|X_i)=\mathsf{\Lambda }({\beta }_0+{\beta }_1·{Gender}_i+{\beta }_k·{Controls}_i+{ϵ}_i)$$

(1)

where

• $P(\mathtt{Loan}\_{\mathtt{status}}_i=1)$ is the probability of default;
• $\mathsf{\Lambda }$ is the logistic function;
• ${Gender}_i$ is our key explanatory variable (Gender_Numerical);
• ${Controls}_i$ is a vector of control variables (Credit Rating, Age, etc.).

Our hypothesis H1 is supported if the coefficient ${\beta }_1$ is statistically significant (p < 0.05) and its sign indicates a higher default probability for males.

5.3.2. Testing H2: Gender and Loan Pricing

To test H2—that the platform’s pricing model is gender-blind—we use Ordinary Least Squares (OLS) regression to model the loan’s final interest rate.

$$\mathtt{Loan}\_{\mathtt{Interest}}_i={\beta }_0+{\beta }_1·{Gender}_i+{\beta }_k·{Controls}_i+{ϵ}_i$$

(2)

where

• $\mathtt{Loan}\_{\mathtt{Interest}}_i$ is the interest rate charged;
• ${Gender}_i$ is our key explanatory variable;
• ${Controls}_i$ is the same vector of control variables.

Our hypothesis H2 is supported if the coefficient ${\beta }_1$ is not statistically significant (p > 0.05), which would demonstrate that gender is not a factor in the platform’s pricing, even if H1 is true.

6. Conclusions

This study moves beyond a single-method analysis to provide robust, multi-stage evidence of a regulatory paradox in credit risk modeling. First, an exploratory data analysis on 10,893 loans revealed consistent, non-parametric evidence that female borrowers have lower default rates than males across nearly all age, income, and credit rating groups, yet are charged identical interest rates.

Second, a formal econometric analysis on a cleaned sample of 9,707 loans confirmed these findings. By applying a Logit model to actual defaults (H1) and an OLS model to loan pricing (H2), we demonstrated a clear and statistically significant disconnect: gender predicts risk (p < 0.001), but it does not influence price (p = 0.263). These conclusions are based solely on observed loan-level variables and do not account for unobserved factors such as household structure or the full internal specification and calibration of the platform’s proprietary scoring model.

This legally mandated, gender-blind approach, while well-intentioned, creates a quantifiable economic distortion. It leads to the systematic mispricing of risk and an implicit subsidy from lower-risk borrowers to their higher-risk counterparts. The findings highlight a critical need for policymakers and regulators to re-evaluate whether the current “fairness-through-unawareness” doctrine in EU law achieves the goal of gender equality in financial access, or whether, in some contexts, it inadvertently disadvantages the very groups it aims to protect.