Dynamic Morphing Augmentation for Algorithmic Fairness in Financial Models

Abstract

Automated classification models in high-stakes financial domains, such as credit scoring assessment and transaction processing, are increasingly susceptible to algorithmic performativity-a phenomenon where a model's deployment alters the environmental data distribution, leading to severe fairness degradation over time. This continuous distribution shift, compounded by historical predictions influencing future inputs, often manifests as robustness bias and overconfidence, creating unstable decision boundaries that usually disproportionately penalize minority demographic groups. Traditional evaluation metrics like aggregate accuracy routinely mask these subtle, localized vulnerabilities. To address these limitations, this thesis introduces a dual-purpose framework for both auditing and mitigating demographic robustness bias in tabular financial data. First, we propose FairMorph, a dynamic data augmentation paradigm designed to directly counteract robustness bias. Unlike static, a priori augmentation strategies, FairMorph constructs fairer training batches on-the-fly by resolving a discrete Optimal Transport problem. This mathematical alignment pairs vulnerable, misclassified minority instances with robust, correctly classified majority instances, interpolating the low-density data manifold via linear trajectories. By dynamically generating samples to balance group performance, this approach prevents the model from drastically segregating demographic groups during the optimization process. Second, to audit these models, we extend the time-series model sensitivity framework, tsMIST, to the tabular domain as TabMIST. TabMIST computes the switch Threshold, the exact structural deformation required to flip a classifier's prediction, in order to derive interpretable, macro-level statistical metrics (TabMIST_Avg and TabMIST_Std). We evaluate the utility of this combined framework by applying it to a highly sensitive real-world setting of financial fraud detection. Our analysis demonstrates that standard classifiers can exhibit drastic robustness disparities across demographic groups despite achieving uniform accuracy. The application of FairMorph serves to regularize these boundaries, mitigating fairness decay caused by performative shifts while striving to preserve core model utility. Concurrently, TabMIST provides vital, sample-specific sensitivity information, which is an essential safeguard in high-stakes financial deployment, alongside a comprehensive overview of model stability across different demographic groups. Ultimately, this framework advances responsible AI auditing and training practices, providing actionable mechanisms to ensure long-term algorithmic equity.