Algorithmic Fairness as a Risk-Management Problem in Banking and Insurance: Regulatory Frameworks, Model Governance, and Fairness-Aware Credit Scoring

1. Introduction

The rapid integration of artificial intelligence (AI) into financial decision-making has changed the logic, pace, and scale of credit allocation. Nowhere is this more visible than in credit scoring, where algorithmic systems increasingly determine access to loans, mortgages, insurance cover, and other forms of financial capital. Although predictive modeling in consumer credit is not new, the shift from transparent statistical scorecards to complex machine-learning (ML) architectures represents a qualitative leap in both technical capability and in the risks that supervisors and firms must manage (Hurley and Adebayo 2017; Binns 2020).

Credit and insurance scores are foundational to economic participation. They influence not only immediate lending and underwriting decisions but also downstream access to housing, education, entrepreneurship, and health-related products (Barocas et al. 2019). Yet these systems are frequently trained on historical data that reflect entrenched social and economic inequalities. As a result, automated scoring can encode, perpetuate, or amplify disparities across ethnicity, class, gender, and geography (O’Neil 2016; Eubanks 2018). For a regulated institution, those disparities are not only an ethical concern: they are a source of legal, conduct, reputational, and model risk that falls squarely within the remit of enterprise risk management and prudential supervision.

Despite this, the literature has paid far less attention to the distributive consequences of scoring systems than to their accuracy. Mainstream work has emphasized precision and discrimination while downplaying the disparate impacts that emerge across social groups (Mitchell et al. 2021), and the dominant industry mitigation—removing sensitive attributes from the data—is now widely recognized as insufficient, because correlated features act as proxies and reproduce bias through indirect pathways (Kim 2019; Binns 2020). Treating fairness as a first-order risk-management objective, rather than a post hoc compliance formality, remains the exception rather than the rule.

This study addresses that gap. It is positioned at the intersection of computational modeling, financial regulation, and institutional risk governance, and its central premise is that an AI scoring system must be evaluated not only for how well it predicts, but for whom it predicts well, at what social cost, and under which supervisory obligations. The analysis is anchored in credit scoring—the canonical high-risk use case—but is framed throughout for the shared banking-and-insurance perimeter, because the same proxy-discrimination mechanisms, and much of the same regulatory architecture, govern insurance pricing and underwriting.

The remainder of this section reviews the state of the field and the supervisory perimeter, and states the research questions. Section 2 describes the data, variables, modeling pipeline, and the performance and fairness metrics. Section 3 reports the empirical findings. Section 4 discusses the risk-management and governance implications. Section 5 concludes.

1.1. Machine Learning in Credit Scoring and the Model-Risk Problem

The diffusion of AI across financial services is a structural change in the infrastructure of contemporary finance (Arner et al. 2016). From credit-risk modeling to fraud detection and customer segmentation, AI technologies are reshaping banking and insurance at a scale and speed that often outpaces existing regulatory and ethical frameworks (Pasquale 2015; Bussmann et al. 2021). Each efficiency gain, however, introduces or amplifies a risk that must be governed: model risk from opaque estimators; conduct and legal risk from discriminatory outcomes; operational and data risk from alternative data pipelines; and reputational risk from decisions that customers cannot understand or contest.

Historically, credit scoring quantified default likelihood through logistic regression, discriminant analysis, and decision trees; Thomas et al. (2002) established foundational practices balancing predictive accuracy with interpretability—concerns that remain central to today’s AI-governance debates. Hand and Henley (1997) warned early that reliance on complex classifiers risked sacrificing transparency for marginal accuracy gains, a concern intensified by high-dimensional models that are powerful but difficult to audit (Lipton 2018; Rudin 2019). Since the mid-2000s, supervised ML has captured nonlinearities and high-dimensional interactions and has consistently outperformed traditional techniques across credit datasets (Lessmann et al. 2015; Baesens et al. 2003), with gradient boosting performing especially well under class imbalance and noise (Brown and Mues 2012). Behavioral credit scoring further incorporates transaction histories, mobile usage, and digital footprints (Bazarbash 2019; Berg et al. 2020), which advocates argue extends credit to thin-file applicants (Frost et al. 2019).

From a supervisory standpoint, these gains arrive coupled with model risk—the risk of adverse consequences from decisions based on incorrect or misused model output. Model-risk-management frameworks (Board of Governors of the Federal Reserve System 2011; Basel Committee on Banking Supervision 2005) require firms to validate models across their life cycle and to demonstrate conceptual soundness. Where a scoring model produces systematically different outcomes for structurally disadvantaged groups, that disparity is itself a model-risk and conduct-risk finding, not merely an ethical footnote.

1.2. Algorithmic Bias, Proxy Discrimination, and Fairness as a Risk Category

The promise of algorithmic scoring is tempered by the critique that it can mask systemic bias behind a veneer of objectivity. As O’Neil (2016) argues, systems trained on historical records often encode structural discrimination; seemingly neutral features such as ZIP code or education can proxy for ethnicity, gender, or class (Barocas and Selbst 2016; Eubanks 2018). The risks are not merely theoretical: ML systems in credit and insurance have disproportionately penalized minority and low-income applicants even when risk profiles are comparable (Bartlett et al. 2021; Fuster et al. 2022), while opacity complicates redress and makes adverse decisions difficult to contest (Citron and Pasquale 2014).

In this framing, financial inclusion is more than a design objective—it is a question of economic justice, because exclusion from credit constrains housing, entrepreneurship, education, and healthcare (Hurley and Adebayo 2017; Narayan et al. 2018). Discriminatory algorithms can deepen inequality under the guise of computational neutrality (Binns 2018; Gabor and Brooks 2017). The operative consequence for a regulated institution is that fairness must be measured, monitored, and controlled as an explicit risk category, with quantitative metrics, tolerances, and escalation paths—precisely the logic this paper operationalizes.

1.3. Fairness and Discrimination Risk in Insurance

Although the empirical setting is credit, the same mechanisms and much of the same regulatory logic extend to insurance. Insurance pricing and underwriting are, by construction, acts of risk differentiation, which makes the boundary between legitimate actuarial discrimination and unfair discrimination unusually delicate. The Court of Justice of the European Union’s Test-Achats ruling (C-236/09) prohibited the use of gender as a rating factor in EU insurance, establishing that even a statistically predictive attribute can be legally impermissible on equality grounds. As insurers adopt granular behavioral and telematics data and ML pricing, proxy discrimination becomes a first-order concern: a model can reconstruct protected attributes from correlated features and reproduce prohibited distinctions without ever observing them directly.

Supervisors have responded accordingly. The work of the European Insurance and Occupational Pensions Authority (EIOPA) on the governance of AI and on the ethical use of data in insurance emphasizes fairness, non-discrimination, transparency, and human oversight, and recent attention to differential and behavior-based pricing reflects concern that data-driven personalization can produce outcomes that are unfair even when actuarially defensible. The International Association of Insurance Supervisors (2020) has likewise highlighted fairness, accountability, and explainability as supervisory priorities. The fairness metrics, mitigation techniques, and governance controls examined here are therefore directly transferable from credit to insurance underwriting and pricing.

1.4. The Regulatory and Supervisory Landscape

Regulatory frameworks have tightened around automated decision-making. The European Union’s Artificial Intelligence Act classifies creditworthiness assessment and life- and health-insurance risk pricing as high-risk, imposing obligations on data governance, technical documentation, transparency, human oversight, accuracy, and risk management across the model life cycle (European Commission 2021). The General Data Protection Regulation (GDPR) constrains automated decisions with legal or similarly significant effects and underpins expectations of meaningful information about the logic involved (Voigt and Von dem Bussche 2017; Wachter et al. 2017). In the United States, the Equal Credit Opportunity Act (ECOA) and the Consumer Financial Protection Bureau (CFPB) require that algorithmic credit decisions be explainable enough to support specific adverse-action reasons.

Sector supervisors add a prudential layer. The European Banking Authority (2020) guidelines on loan origination and monitoring set expectations for data quality, documentation, and explainability in credit decisions, while model-risk doctrine (Board of Governors of the Federal Reserve System 2011) requires independent validation and effective challenge. On the insurance side, Solvency II governance and risk-management requirements, together with EIOPA and IAIS guidance, extend analogous expectations to underwriting and pricing models. International bodies have issued principles for responsible AI in finance—transparency, robustness, accountability, and human-in-the-loop control (OECD 2019; Financial Stability Board 2017, 2022).

These instruments have fueled interest in explainable AI (XAI). Tools such as SHapley Additive exPlanations (SHAP) (Lundberg and Lee 2017), Local Interpretable Model-agnostic Explanations (LIME) (Ribeiro et al. 2016), and counterfactual explanations (Wachter et al. 2017) help render complex models interpretable. Yet critics caution that explanation methods can oversimplify or misrepresent model logic (Lipton 2018; Rudin 2019) and do not resolve deeper problems when biased data or misaligned objectives drive behavior. Scholars therefore warn of an implementation gap—firms adopting ethical principles rhetorically while optimizing for efficiency, a dynamic labeled ethics-washing (Mittelstadt 2019; Morley et al. 2020).

1.5. Research Questions

This study asks how AI-driven scoring systems in banking and insurance can be developed, validated, and governed so that they support—rather than undermine—financial inclusion, while remaining defensible under the applicable regulatory frameworks. Three interrelated questions follow.

• RQ1. To what extent do AI-based scoring models reproduce or mitigate bias against historically marginalized or underrepresented groups (by ethnicity, gender, socioeconomic status, or geography)? This builds on Fuster et al. (2022) and Bartlett et al. (2021), who show that disparities can persist under ostensibly neutral algorithmic regimes.
• RQ2. What role do fairness-aware ML techniques play in mitigating algorithmic discrimination, and at what measurable cost to performance? While explanation tools provide transparency, Lipton (2018) and Rudin (2019) caution that explanation alone does not ensure regulatory compliance if the underlying data or objectives remain misaligned.
• RQ3. Which governance frameworks and regulatory mechanisms are most effective in ensuring accountability, transparency, and ethical alignment in financial AI across banking and insurance? A persistent implementation gap remains between principles and organizational practice (Mittelstadt 2019; Morley et al. 2020).

The contribution is threefold: it quantifies the fairness cost of model capacity on a single institutional dataset under a harmonized metric set; it prices the accuracy given up by three distinct mitigation families; and it translates both into a control framework aligned with the supervisory expectations that now govern banking and insurance.

2. Materials and Methods

2.1. Data Source and Description

The study uses an institutional dataset of 20,000 anonymized credit applications from a mid-sized digital lending institution operating in Southern Europe. The data combine information collected at the point of application, internal credit-decision variables, and observed repayment outcomes over a 12-month horizon, reflecting the industry-wide shift toward hybrid scoring systems that blend traditional financial attributes with alternative behavioral and psychometric features (Jagtiani and Lemieux 2019; Moro-Visconti et al. 2021). The target variable is a binary indicator, loan default, denoting whether the borrower failed to repay within 12 months of origination.

The raw file contained 20,000 rows (one per applicant) and 55 columns covering input features, metadata, and outcomes, extracted from the institution’s internal decision-support system, which integrates customer self-disclosures, a credit-bureau application programming interface, mobile-application behavior logs, and psychometric-testing modules. Prior to transfer, the institutional partner applied a privacy-preserving transformation. Personally identifiable information (PII)—names, national identification numbers, and full addresses—was permanently removed, and quasi-identifiers such as ZIP code, age, and education were generalized using k-anonymity techniques (Sweeney 2002; El Emam and Dankar 2008): ZIP codes were aggregated into socio-economic clusters and age was binned into five-year intervals before re-normalization.

Exploratory profiling assessed completeness, consistency, and integrity (Dasu and Johnson 2003). Over 96% of records were complete, with missingness concentrated in behavioral features and optional psychometric fields; no duplicates were found. Missing values were imputed with k-nearest neighbors (KNN, k = 5), which preserves local structure and handles mixed-type features (Troyanskaya et al. 2001; Batista and Monard 2002); mode imputation was used for ordinal psychometric scales and predictive mean matching for continuous scores. Imputation followed outlier mitigation but preceded normalization. Figure 1 summarizes the pipeline.

A feature-engineering stage then constructed derived variables for risk modeling and fairness auditing (Kuhn and Johnson 2013; Fuster et al. 2022): debt-to-income ratio, credit utilization, a financial-literacy score, a risk-tolerance index, and an ethno-socioeconomic group proxy combining ZIP-code clusters and a surname-origin classifier. Continuous variables were standardized via z-scores; categorical features were processed with one-hot encoding for unordered attributes and ordinal encoding for education and risk indices (Hastie et al. 2009; Pedregosa et al. 2011). The analytical dataset comprises 38 predictors and one binary target, with every transformation logged in a version-controlled Python pipeline.

The ethno-socioeconomic group proxy warrants explicit caution. Constructing a sensitive-attribute proxy from surname and location is a recognized auditing device where protected characteristics are not recorded (Rambachan et al. 2023), but it is measured with error and can itself embed bias. It is therefore used only for post hoc fairness auditing and mitigation, never as a prediction feature intended for deployment, and all subgroup findings are indicative of proxy-level disparity rather than measurement of any individual’s ethnicity.

2.2. Variables

The dataset spans five domains: financial and credit attributes; behavioral indicators; socio-demographic characteristics; psychometric measures; and operational metadata. Feature inclusion was guided by empirical relevance and the credit-scoring and fairness literature (Thomas et al. 2002; Crook et al. 2007; Hurley and Adebayo 2017). Conventional indicators—credit score, annual income, loan amount, loan term, and delinquencies—are well-established predictors (Hand and Henley 1997; Baesens et al. 2003), while debt-to-income and credit utilization proxy financial strain and repayment capacity consistent with regulatory expectations for explainable modeling (European Banking Authority 2020).

Behavioral variables—transaction frequency, mobile-application logins, and an e-commerce activity index—capture engagement and digital inclusion and can supplement or outperform static attributes for thin-file applicants (Berg et al. 2020; Björkegren and Grissen 2020). Demographic variables are retained for subgroup fairness auditing rather than prediction, consistent with anti-discrimination compliance (Barocas et al. 2019; Kamiran and Calders 2012). Psychometric measures reflect growing interest in integrating cognitive and behavioral signals into scoring (Lusardi and Mitchell 2014; Behr and Güttler 2007). Operational metadata serve as controls and as diagnostics for procedural bias. Table 1 summarizes the key variables.

2.3. Research Design and Model Architecture

The empirical strategy compares three modeling paradigms—an interpretable baseline, high-capacity classifiers, and fairness-aware architectures—evaluated both quantitatively, through predictive and fairness metrics, and qualitatively, through explainability and regulatory compliance. Table 2 summarizes the three phases.

2.3.1. Logistic Regression (Interpretable Benchmark)

The benchmark is a regularized logistic-regression (LR) classifier, chosen for its established role in credit scoring and its interpretability, stability, and regulatory defensibility (Hand and Henley 1997; Thomas et al. 2002). Its coefficient-level output maps features to odds ratios, supporting adverse-action explanations. Let xi denote the predictor vector for applicant i and yi ∈ {0,1} the outcome, with yi = 1 a default event. The model estimates the conditional probability of default as

P(yᵢ = 1 | xᵢ) = 1 / [1 + exp(−xᵢᵀβ)],

(1)

where β = (β₀, β₁, …, βₚ)ᵀ is the coefficient vector. Equivalently, the model assumes a log-odds transformation that is linear in the predictors,

log{P(yᵢ = 1 | xᵢ) / [1 − P(yᵢ = 1 | xᵢ)]} = β₀ + Σⱼ βⱼ xᵢⱼ.

(2)

Given the dimensionality of the feature space, an L2 (ridge) penalty constrains coefficient magnitudes, so that estimation minimizes

L(β) = −(1/n) Σᵢ [yᵢ log(p̂ᵢ) + (1 − yᵢ) log(1 − p̂ᵢ)] + λ Σⱼ βⱼ²,

(3)

where p̂ᵢ = P(yᵢ = 1 | xᵢ), λ ≥ 0 is the regularization hyperparameter, and n is the number of training observations. The penalty strength C = 1/λ is tuned by grid search under five-fold cross-validation using the Brier score as the objective. Predictors are z-score standardized and multicollinearity is screened with variance inflation factors (VIF), with predictors exceeding a VIF of 10 removed or transformed.

Training follows a two-tier validation design: an 80/20 stratified train–test split preserving the outcome distribution, and, within the training set, five-fold cross-validation embedded in the grid search. Beyond serving as a performance benchmark, the LR specification acts as a reference architecture for regulatory audit, with coefficients inspected for indirect discrimination by relating their sign and magnitude to known proxies for sensitive attributes. Figure 2 sets out the workflow.

2.3.2. Machine-Learning Classifiers

The second phase introduces two high-capacity classifiers that capture nonlinear interactions: a gradient-boosting machine (GBM) and a multi-layer perceptron (MLP) neural network, representing boosting ensembles and deep feedforward networks (Fuster et al. 2022; Jagtiani and Lemieux 2019). The GBM is implemented with LightGBM, a histogram-based, memory-efficient variant (Ke et al. 2017). Trees are grown sequentially, each correcting the residuals of its predecessors; grid search optimizes the learning rate, number of estimators, maximum depth, and subsampling ratio, with regularization to control overfitting.

The MLP comprises an input layer, two fully connected hidden layers (64 and 32 units) with rectified linear unit activations, dropout (rate 0.3), and a sigmoid output; it is trained with the Adam optimizer and early stopping on validation loss, with randomized search over learning rate, batch size, and dropout rate. Both models are trained on the same 80% stratified training set and evaluated on the 20% holdout, with variance assessed across five random seeds. Unlike the LR benchmark, these models lack inherent interpretability and therefore require model-agnostic explanation techniques, discussed in Section 4.3. Figure 3 summarizes the architecture.

2.3.3. Fairness-Aware Model Integration

The third phase introduces fairness-aware techniques addressing the distributional concerns raised by ML scoring (Barocas et al. 2019; O’Neil 2016; Fuster et al. 2022). A three-tiered strategy mirrors the ML pipeline—pre-processing, in-processing, and post-processing—so that each intervention targets a distinct source of bias.

Pre-processing reweighing (Kamiran and Calders 2012) adjusts training-sample weights to equalize the outcome distribution across sensitive groups without altering labels or features, applying inverse-probability weights to compensate for imbalances. In-processing adversarial debiasing (Zhang et al. 2018) couples the main predictor with an adversary that attempts to recover group membership from the learned representation; penalizing informative representations encourages group-invariant predictors, with a tunable fairness coefficient selected by grid search. Post-processing reject-option adjustment (Kamiran et al. 2012), related to the equalized-odds framework of Hardt et al. (2016), operates on calibrated probabilities within an uncertainty band around the decision threshold, reassigning marginal cases to reduce disparities while leaving the underlying model unchanged—attractive where retraining a production model is impractical.

2.4. Performance and Fairness Metrics

Performance is assessed with three complementary metrics. The Brier score, the mean squared difference between predicted probabilities and outcomes, is the primary optimization and comparison metric because calibration matters as much as discrimination when scores drive pricing and limit-setting:

Brier = (1/n) Σᵢ (p̂ᵢ − yᵢ)².

(4)

The area under the receiver operating characteristic curve (AUC-ROC) measures threshold-independent rank-ordering and is well suited to imbalanced data. The F1 score, the harmonic mean of precision and recall, captures the balance between false positives and false negatives:

F1 = 2 × (Precision × Recall) / (Precision + Recall).

(5)

Accuracy is reported as a secondary indicator. To keep the fairness analysis internally consistent, the favorable outcome is defined as loan approval (a predicted non-default), with G = 1 denoting the structurally disadvantaged group and G = 0 the reference group. Under this convention, the demographic-parity difference (DPD) is

DPD = P(Ŷ = 1 | G = 1) − P(Ŷ = 1 | G = 0),

(6)

where Ŷ = 1 denotes approval; negative values indicate lower approval for the disadvantaged group. The equal-opportunity difference (EOD) compares true-positive (correct-approval) rates for genuinely creditworthy applicants:

EOD = TPR(G = 1) − TPR(G = 0), with TPR = TP / (TP + FN).

(7)

The disparate-impact ratio (DIR) expresses the same comparison multiplicatively,

DIR = P(Ŷ = 1 | G = 1) / P(Ŷ = 1 | G = 0),

(8)

where values below the 0.80 four-fifths threshold signal potential legal or ethical concern (Feldman et al. 2015). Finally, the Theil index (TI) is a decomposable inequality measure from information theory that captures entropy in outcome allocation and attributes disparity to between- and within-group variation (Speicher et al. 2018):

TI = (1/n) Σᵢ (bᵢ / μ) ln(bᵢ / μ),

(9)

where bᵢ is the benefit assigned to individual i and μ its mean; higher values indicate greater inequality. Reporting performance and fairness jointly reflects the consensus that optimization cannot be separated from social impact and legitimacy (OECD 2019; European Commission 2021). In a risk-management framing, these metrics function as key risk indicators: measurable quantities with tolerances against which model behavior can be monitored and escalated.

2.5. Use of Generative Artificial Intelligence

A generative AI assistant (Claude, Anthropic) was used to support manuscript editing, structuring, and formatting for the target journal. It was not used to generate, simulate, or analyze data, to produce results, or to design the study. All modeling, estimation, and interpretation were performed by the author, who has reviewed and verified all content and takes full responsibility for it.

3. Results

This section reports the empirical findings across three axes: the comparative predictive performance of the baseline and high-capacity classifiers; group-fairness diagnostics across socio-demographic strata; and the effects of fairness-aware interventions. All models are evaluated on the holdout test set from the stratified 80/20 split under identical preprocessing and metrics, as specified in Section 2.

3.1. Predictive Performance Across Model Families

3.1.1. Logistic-Regression Coefficients

The LR model was trained with grid search and five-fold cross-validation; L2 regularization with C = 0.1 minimized the average Brier score. Table 3 reports the estimated coefficients, standard errors, and significance levels for the final specification.

The coefficient signs, magnitudes, and significance levels align with established credit-risk theory. Credit score is among the strongest predictors, with a negative and highly significant coefficient (β = −0.013, p < 0.001): each additional point reduces the log-odds of default by roughly 1.3%, all else equal (Hand and Henley 1997). Annual income (β = −0.034) and, in the opposite direction, the debt-to-income ratio (β = +1.43, p < 0.001) contribute intuitively, the latter signaling a steep rise in default risk as financial burden intensifies (Crook et al. 2007). Credit utilization (β = +0.87) and past delinquencies (β = +0.54) are strongly positive, consistent with overextension and the persistence of late-payment behavior.

Digital-behavior variables carry meaningful signal: higher mobile-application logins (β = −0.029) and e-commerce activity (β = −0.32) are associated with lower default probability, echoing evidence that digitally engaged customers exhibit greater financial monitoring (Berg et al. 2020). Psychometric traits matter as well—higher financial literacy reduces risk (β = −0.011) while higher risk tolerance increases it (β = +0.21). Most consequentially, the ethno-socioeconomic group indicator shows a large, significant positive association with default (β = +0.36, p < 0.001) that is not fully explained by observable financial behavior, motivating the fairness analysis in Section 3.2. Overall fit is sound for a behavioral model (McFadden’s R² = 0.278; likelihood-ratio χ² = 1,576.6, p < 0.001), with the log-likelihood improving from −2,830.6 to −2,042.3.

3.1.2. Test-Set Performance

Table 4 summarizes the holdout performance of the baseline. The Brier score of 0.162 indicates good calibration—important where the predicted probability, not merely the binary decision, drives pricing and limit-setting—with a tight cross-validation band. The AUC-ROC of 0.734 reflects moderate discrimination, above the 0.5 no-skill baseline but below the high-capacity models. The F1 score of 0.691 indicates a reasonable precision–recall balance, and accuracy of 78.5% confirms stable generalization.

3.2. Group Fairness and Distributional Disparities

An exclusive focus on accuracy can obscure how differently social groups are treated. Model behavior is evaluated across strata defined by ethno-socioeconomic group, gender, and ZIP-code cluster, with particular attention to Group A, a proxy for historically marginalized populations (Barocas et al. 2019; Mitchell et al. 2021). Table 5 reports approval rates and fairness metrics under the convention set out in Section 2.4.

The model exhibits material disparities between Group A and other applicants. The DPD of −0.057 means members of Group A are approved 5.7 percentage points less often, while the EOD of −0.165 reveals a substantial gap in correctly approving genuinely creditworthy applicants from this group. The DIR of 0.67 falls below the 0.80 four-fifths threshold, indicating that the model’s output could constitute de facto disparate impact even absent intentional bias, and the Theil index confirms concentrated inequality. Gender disparities are small and within acceptable bounds (DIR = 1.07; EOD = +0.035), while ZIP-cluster B applicants face moderate geographic disadvantage (DIR = 0.77).

Crucially, none of these group features were used as model inputs: the disparities are the product of proxy discrimination, where correlated non-sensitive variables such as income, utilization, and behavioral indices indirectly encode group membership (Kim 2019). This is the empirical refutation of fairness through unawareness and the case for post hoc auditing and correction in high-stakes domains (Friedler et al. 2019; Binns 2020).

3.3. Effects of Fairness-Aware Interventions

Recognizing the disparities documented in Section 3.2, this subsection evaluates whether fairness-aware interventions can mitigate bias without sacrificing predictive validity. Each technique operationalizes a distinct fairness criterion—disparate-impact reduction, equal-opportunity preservation, and error-rate parity (Verma and Rubin 2018; Corbett-Davies and Goel 2018). For each family, fairness-aware variants were trained and tested under identical conditions using the ethno-socioeconomic group as the protected attribute. Table 6 reports both performance and fairness.

All three mitigation strategies reduce demographic and opportunity disparities without prohibitive losses in performance. Reweighed LR improves the DIR from 0.67 to 0.85 and roughly halves DPD and EOD, with AUC declining only marginally from 0.734 to 0.723, making it well suited to compliance-sensitive applications. The adversarially debiased GBM achieves the most favorable trade-off among high-capacity models, preserving strong AUC (0.765) and F1 (0.701) while bringing DPD and EOD into acceptable ranges by preventing the protected attribute from being inferred from the latent representation (Zhang et al. 2018). The reject-option MLP reaches a DIR of 0.81 and cuts EOD by roughly 45% at a modest calibration cost, with the advantage of leaving the production architecture intact.

Taken together, these results qualify the presumed strict fairness–accuracy trade-off. Some pure-classification performance is given up, but the gains in distributional equity and regulatory defensibility are large relative to the efficiency lost—an especially favorable balance in high-stakes domains such as credit and insurance underwriting.

3.4. Summary of Findings

Three findings stand out. First, the interpretable LR benchmark performs robustly (Brier 0.162, AUC 0.734, F1 0.691), with coefficients consistent with domain knowledge and with financial, behavioral, and psychometric features all contributing. Second, fairness diagnostics reveal meaningful disparities: Group A faces markedly lower approval and true-positive rates (DIR = 0.67), and ZIP-based geographic disadvantage is present, whereas gender disparities are negligible—evidence of algorithmic redlining through proxy features rather than overt discrimination. Third, disparities are neither inevitable nor uncorrectable: pre-, in-, and post-processing interventions all materially improve fairness metrics with only modest performance costs. Table 7 consolidates the model-by-model trade-offs.

4. Discussion

The results have a direct translation into supervised risk management. Higher-capacity models buy accuracy at the price of wider group disparities, and those disparities are measurable, monitorable, and controllable. This section positions the findings within the model-risk and regulatory frameworks that govern automated decision-making in banking and insurance, and derives concrete controls.

4.1. Fairness Risk Within the Model-Risk and Three-Lines Framework

Disparate outcomes are best understood not as an ethical externality but as a compound of model, conduct, legal, and reputational risk. In a three-lines-of-defense structure, the first line—model development and business ownership—should embed fairness key risk indicators into model-development standards and pre-deployment testing, using the DIR against the four-fifths threshold and DPD and EOD against internally calibrated tolerances. The second line, independent model validation and compliance, should subject scoring models to fairness validation with the same rigor applied to discrimination and calibration testing, consistent with model-risk doctrine (Board of Governors of the Federal Reserve System 2011). The third line, internal audit, should verify that these controls operate as designed. Because the results show that fairness can be improved at modest performance cost, a firm that declines to test and mitigate is, in effect, accepting an uncompensated and avoidable risk.

4.2. Regulatory Alignment Across Banking and Insurance

The interventions studied here map onto specific obligations. Under the high-risk regime of the EU AI Act, requirements on data governance, technical documentation, transparency, human oversight, and risk management are directly served by pre-deployment fairness testing, an algorithmic impact assessment, and retention of the mitigation evidence documented in Section 3.3. The constraints of the GDPR on automated decisions, and the explainability standards of the ECOA and the CFPB, favor retaining an interpretable challenger model alongside any high-capacity champion, so that adverse-action reasoning remains defensible. The loan-origination and internal-model expectations of the European Banking Authority (2020) reinforce data quality, documentation, and explainability for models used in credit decisions.

In insurance the same logic applies with a sharper equality constraint: the Test-Achats prohibition on gender rating shows that a predictive attribute can be legally impermissible, so proxy-discrimination testing of behavioral and telematics-based pricing is essential. Solvency II governance, EIOPA guidance, and the supervisory priorities of the International Association of Insurance Supervisors (2020) point to the same control set—impact assessment, independent validation, disparate-impact monitoring, and human oversight. The disparate-impact test and the mitigation toolkit examined here are, in this sense, portable across the banking–insurance perimeter.

4.3. Interpretability and Validation of High-Capacity Models

Because the GBM and MLP lack inherent interpretability, their deployment depends on model-agnostic explanation and disciplined validation. Post hoc tools such as SHAP (Lundberg and Lee 2017), LIME (Ribeiro et al. 2016), and counterfactual explanations (Wachter et al. 2017) can support feature-attribution review, adverse-action generation, and the detection of proxy pathways, and should be produced and retained as part of technical documentation. However, explanation is a complement to, not a substitute for, fairness control: as Lipton (2018) and Rudin (2019) caution, an explanation of a biased model does not make it fair. Where interpretability requirements are binding and the accuracy gap is small—as it is here, at roughly six AUC points—an interpretable model or an interpretable challenger remains the prudent default.

4.4. Limitations

Several limitations qualify the findings. First, the data are cross-sectional, limiting inference about temporal dynamics and the downstream effects of credit decisions. Second, the protected-attribute proxies are measured with error and cannot fully capture ethnicity, class, or their intersections; subgroup results should be read as proxy-level disparities. Third, group-fairness metrics such as DIR and EOD, while decision-relevant, do not capture all forms of harm, particularly intersectional or contextual ones, and different metrics can conflict. Fourth, the mitigation strategies were evaluated under controlled conditions and may behave differently inside production workflows with feedback loops and strategic applicant behavior. Finally, the analysis is confined to a single institutional dataset from one market; external validity across markets and regulatory regimes remains to be established.

5. Conclusions

The integration of AI into credit and insurance scoring is a pivotal transformation in the architecture of financial decision-making, and this study has evaluated its performance and fairness through the lens of supervised risk management. The interpretable logistic-regression benchmark delivered reasonably strong, well-calibrated predictions (AUC 0.734, Brier 0.162) with high transparency, yet still produced significant group-level disparities—disadvantaging applicants from marginalized ethno-socioeconomic groups and certain high-risk geographies—arising not from overt discrimination but from proxy pathways and structural inequalities embedded in the data. High-capacity models delivered marginal accuracy gains (AUC above 0.78) while widening those disparities, with the DIR falling as low as 0.58.

Crucially, fairness-aware modeling showed that algorithmic bias is a tractable design and control problem rather than a fixed property: reweighing, adversarial debiasing, and reject-option adjustment improved fairness metrics, raising the DIR to as much as 0.85, while preserving acceptable accuracy and calibration. The deterministic view of a strict fairness–accuracy trade-off is not supported; when fairness is embedded proactively, predictive efficacy and normative responsibility can be reconciled. Fairness, moreover, cannot be secured merely by omitting protected attributes, because proxies carry latent signals of ethnicity, class, and gender, so responsible deployment requires active scrutiny of outcomes rather than passive avoidance of sensitive features.

For regulated institutions, the practical implication is to treat fairness as an explicit risk category with measurable indicators, tolerances, and escalation paths, controlled through algorithmic impact assessment, independent validation, and three-lines-of-defense governance, and aligned with the EU AI Act, GDPR, EBA, and fair-lending expectations in banking and with Solvency II, EIOPA, and IAIS expectations in insurance. While the empirical focus was credit scoring, the performance–fairness co-evaluation and the control framework extend naturally to insurance pricing and underwriting. Future work should incorporate longitudinal data on borrower well-being, develop intersectional fairness metrics, compare fairness operationalization across regulatory regimes, and build participatory mechanisms that give those most affected a voice in the design and audit of scoring systems.