Systematic Review Weighs Which Machine Learning Models Best Detect Bias in Mortgage Lending

A sweeping systematic review of machine learning research on mortgage lending discrimination has concluded that no single algorithm wins on every front: decision trees stand out for their interpretability and effective bias detection, while neural networks and support vector machines achieve superior predictive accuracy but suffer from low transparency. The review, published in Discover Data, synthesized empirical studies published between 2015 and May 2025, drawing primarily on United States mortgage data and fairness evaluation frameworks. Its central message is sobering for the financial technology industry: fairness metrics reveal persistent biases even in highly accurate models, and bias mitigation techniques such as reweighting and fairness constraints do reduce discrimination, but often at a measurable cost to model utility.

The review followed the PRISMA 2020 reporting guidelines across every stage of study identification, screening, and inclusion. The researchers searched two bibliographic databases, Scopus and Web of Science Core Collection, in May 2025, using structured queries that combined three concept groups: mortgage lending context, machine learning or algorithmic credit decisioning, and fairness, discrimination, or mitigation constructs. Of 379 records entering the selection process, 368 were retained as relevant to the review’s scope, and a core subset of 50 studies was classified as highly relevant for detailed comparative synthesis based on a structured relevance scoring rubric. Citation chaining on retained studies supplemented the keyword-based retrieval, and the authors note that no duplicates were identified after merging and deduplication.

The descriptive landscape that emerged is heavily weighted toward tree-based methods. Decision trees and random forests were the most frequently evaluated model families for detecting discriminatory patterns, appearing in thirty studies, followed by neural networks in eight studies and support vector machines in five. Although all three families achieved acceptable predictive accuracy, the review found their relative performance was contingent on dataset characteristics and modeling context rather than reflecting a stable ranking. Studies using publicly available Home Mortgage Disclosure Act (HMDA) data consistently reported strong performance from tree-based models, which frequently matched or outperformed logistic regression baselines. By contrast, studies employing richer feature sets or proprietary datasets with higher dimensionality tended to report accuracy advantages for neural networks and ensemble methods, whose capacity to capture nonlinear interactions yielded measurable gains.

A notable contradiction runs through this evidence. Several studies report that neural networks outperform tree-based models on predictive accuracy, while others find random forests or logistic regression more accurate on the same task type. The review attributes this divergence most plausibly to differences in dataset size, feature representation, hyperparameter optimization, and preprocessing pipelines rather than to inherent algorithmic superiority. The five studies evaluating support vector machines produced inconsistent rankings relative to decision trees, with performance appearing sensitive to kernel selection and regularization parameters. A recurring methodological gap across all three families is the absence of standardized evaluation protocols: studies varied in train-test split strategies, cross-validation procedures, and the performance metrics they reported, which constrains direct cross-study comparison and limits definitive conclusions about which algorithm family is most effective.

How fairness itself is measured turns out to be decisive for what any study concludes. The review classifies fairness metrics into four families. Parity-based group metrics, such as statistical parity difference and disparate impact ratio, assess whether outcomes are distributed equally across protected and non-protected groups and are widely used because of their regulatory familiarity, though they may obscure legitimate risk differences and are sensitive to base-rate variation. Error-rate-based metrics, including equalized odds and equal opportunity, compare predictive errors across groups and are particularly relevant to credit decisioning because they reflect differential treatment in approval and rejection errors. Calibration-based metrics require predicted risk scores to carry the same meaning across groups, while individual and counterfactual fairness approaches ask whether similarly situated individuals receive similar outcomes. Studies relying solely on parity metrics tended to report that bias mitigation was effective, but studies that supplemented them with error-rate measures frequently found conflicting results: a model satisfying statistical parity could simultaneously violate equalized odds by producing higher false-positive rates for protected groups, a pattern consistent with well-established impossibility results in the fairness literature.

The quantitative picture of metric adoption reveals significant blind spots. Twelve studies in the core set explicitly incorporated quantitative fairness metrics, with parity-based measures dominating. Calibration-based fairness assessments were rare, with only two studies examining whether predicted risk scores carried the same meaning across protected groups, a gap the authors call particularly consequential because calibration fairness can be violated even when parity and error-rate criteria are satisfied, meaning applicants from different groups may receive systematically different risk assessments despite similar underlying creditworthiness. Seven studies introduced or applied intersectional fairness measures designed to capture bias across multiple overlapping attributes, including an Accuracy-Weighted Intersectional metric and subgroup-specific debiasing approaches, but intersectional evaluation remains the exception rather than the norm, with most studies restricting analysis to one protected attribute at a time.

On mitigation, the review found that fifteen studies reported bias mitigation techniques reduced discriminatory outcomes, while six studies documented trade-offs between fairness gains and predictive accuracy. Pre-processing techniques, particularly reweighting and data repair, were the most frequently applied and generally produced measurable reductions in parity-based bias metrics with modest accuracy costs. In-processing approaches, including fairness constraints embedded directly in the learning objective, achieved competitive fairness improvements and were better suited to maintaining accuracy when calibrated carefully. Post-processing methods such as threshold adjustment were less commonly evaluated in mortgage-specific contexts, though they offer the advantage of being model-agnostic. Effectiveness varied by algorithm family: tree-based models responded well to pre-processing interventions, neural networks showed greater sensitivity to in-processing constraints, and for support vector machines the evidence base was too limited to draw reliable comparisons. Where quantitative estimates were reported, accuracy losses associated with mitigation typically ranged from one to three percentage points, though inconsistent reporting complicated direct comparison.

Two caveats temper any optimism about these mitigation gains. First, proxy discrimination emerged as a critical moderating factor: models trained without explicit protected attributes still reproduced discriminatory patterns through correlated features such as geography, employment type, and loan characteristics. Several studies that reported successful bias reduction did not test whether proxy effects persisted after mitigation, raising the possibility that reported fairness improvements may be incomplete or unstable. Second, five studies cautioned that overly aggressive fairness constraints risk undermining model utility for all stakeholders, including the protected groups they are intended to benefit; one demonstrated that off-the-shelf fair machine learning algorithms, applied without domain calibration, could amplify rather than reduce certain disparity measures. A small number of studies proposed ensemble or hybrid approaches reporting improved Pareto frontiers between accuracy and fairness, but these remain preliminary and have not been benchmarked under consistent conditions.

Interpretability completes the triangle of competing demands. Twenty studies highlighted the strong interpretability of decision trees, whose clear rule-based structures enable transparency and facilitate fairness assessments, making them a preferred option where regulatory scrutiny and accountability are essential. Ten studies reported that neural networks and ensemble models, despite strong predictive performance, suffer from low transparency that complicates fairness evaluation and reduces stakeholders’ ability to explain individual credit decisions. Several studies explored explainable AI techniques, such as post-hoc interpretability tools, and found they enhance the transparency of complex models. A growing strand of literature goes further, treating explanation outputs themselves as diagnostic evidence for disparate treatment: Shapley-based and Shapley-Lorenz methods have been used to compare feature-attribution patterns across population groups, demonstrating that models may appear fair under full-sample evaluation yet exhibit disparities under subgroup-focused assessment.

The review’s own limitations frame how far its conclusions travel. The evidence base is geographically concentrated, with dominant reliance on United States mortgage data, particularly HMDA and HMDA-linked datasets, limiting generalizability to other regulatory and market contexts. Fairness metrics and performance measures are reported heterogeneously, restricting quantitative meta-analysis, and most reviewed studies adopt associational rather than causal approaches to measuring bias, so fairness conclusions may vary with dataset composition, feature availability, and proxy relationships. The authors call for standardized reporting bundles that jointly document predictive performance, multiple fairness metrics, and proxy-sensitivity analyses; greater emphasis on intersectional fairness evaluation; and tighter integration of explainability tools into fairness auditing and model governance, with fairness evaluation treated as a continuous risk-management process rather than a one-time model check.

Subject of Research: A systematic review of machine learning methods for detecting and mitigating discriminatory patterns in mortgage lending

Article Title: Machine learning algorithms for detecting and mitigating discriminatory patterns in mortgage lending: a systematic review

Article References: Musonerwa, P. S., Uhoranishema, B., Mugisha, R. N., Iradukunda, P., Mpozenzi, P., & Nuwayo, E. (2026). Machine learning algorithms for detecting and mitigating discriminatory patterns in mortgage lending: a systematic review. Discover Data, 4(1), Article 12. https://doi.org/10.1007/s44248-026-00108-z

Image Credits: AI Generated

DOI: 10.1007/s44248-026-00108-z

Keywords: machine learning, mortgage lending, algorithmic fairness, bias mitigation, systematic review, decision trees, neural networks, support vector machines, fairness metrics, explainable AI, HMDA data, proxy discrimination