WHSDA-AI: Interpretable Walsh-Hadamard Spectral Deviation Analysis for Structured Artificial Intelligence Applications in Banking Sector

Abstract

Artificial intelligence (AI)in banking increasingly requires models and representations that are not only predictive but also interpretable, auditable, and stable under model updates. Fraud screening,anti-money-laundering triage, and customer-risk monitoring are high-stakes structured-data tasks in which explanations must be reproducible and connected to documented feature processing. This paper proposes WHSDA-AI, a Walsh-Hadamard Spectral Deviation Analysis framework for structured and profile-based AI in the banking sector and related tabular risk-analysis settings. The proposed framework maps normalized feature profiles into a fixed orthogonal signed-contrast space. Each retained coefficient represents a deterministic comparison between two documented groups of input features. The novelty of the approach is not the classical Walsh-Hadamard transform itself but its integration into a governance oriented AI layer: semantic feature-map versioning, a fixed coefficient dictionary,reference-normalized deviation analysis, leakage-free coefficient selection, bootstrap stability assessment, and compact explanation-card generation. The full Walsh-Hadamard transform is an isometry and cannot improve Euclidean rankings by itself. Practical value arises through coefficient selection, reference deviations, template comparison, risk weighting, and integration with supervised models. Experiments on banking-style diagnostic data, credit-style fraud data, and public structured-data benchmarks compare WHSDA-AI with raw-feature models, sparse logistic regression, principal component analysis (PCA), Haar contrasts, random signed projections, random forests, gradient boosting, XGBoost, and LightGBM. Results are reported with repeated-split uncertainty and conservative robustness interpretation. The framework is positioned as an ante-hoc representation, governance, and feature-engineering layer rather than as a universal accuracy booster. 

Received: 25 May 2026 | Revised: 6 July 2026 | Accepted: 23 August 2026 

Conflicts of Interest

The authors declare that they have no conflicts of interest to this work. 

Data Availability Statement

Data available on request from the corresponding author upon reasonable request. 

Author Contribution Statement

Sergo A. Episkoposian: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Data curation, Writing – original draft, Writing – review & editing, Visualization. Suren H. Parsyan: Conceptualization, Methodology, Investigation, Resources, Writing – review & editing, Supervision, Project administration.