<![CDATA[Purpose: This study examines the performance and validity of tabular deep learning for audit-risk classification by evaluating the sequential effects of parameter-efficient ensembling, piecewise linear embeddings (PLE), and ordinal learning, with particular attention to dependence on target-adjacent audit variables. Design/methods/approach: A synthetic audit-risk dataset was evaluated using repeated nested stratified cross-validation. TabM configurations were compared with conventional machine-learning and neural-network baselines under controlled model selection. Proxy-feature sensitivity analysis was conducted by removing AuditScore, AuditStatus, and RiskFactor, complemented by shallow-tree diagnostics and statistical comparisons. Findings: Parameter-efficient ensembling and PLE produced substantial descriptive improvements under the complete-feature setting, while ordinal learning further reduced ordinal prediction error but was not statistically superior to cross-entropy after multiplicity correction. More importantly, predictive performance deteriorated markedly when target-adjacent audit variables were removed, with all evaluated models approaching baseline-level discrimination. AuditScore emerged as the dominant predictive dependency, followed by AuditStatus, whereas RiskFactor contributed little independent information. Research implications/limitations: The findings demonstrate that strong benchmark performance may depend on proxy variables rather than generalisable audit risk relationships. This study is limited by its synthetic dataset, the unavailability of a target-generation mechanism, and the absence of external or prospective validation. Originality/value: This study reframes high-performing audit-risk classification as a benchmark-validity problem and demonstrates the importance of examining proxy dependence, predictor availability, and target provenance before translating predictive performance into practical audit decision-making support.]]>
Copyrights © 2026
