A System-Aware Pipeline for Deep Neural Network Modeling in Financial Credit Risk
Abstract
Deep Neural Networks (DNNs) have demonstrated strong predictive performance in many domains, yet their adoption in credit scoring is limited due to concerns about interpretability and reliability. This study examines how a DNN behaves when trained on a rigorously preprocessed, high-dimensional financial dataset (839 features; 256k samples). Using SHAP, diagnostic rule sweeps, and a comparison with XGBoost, we show that the DNN internalizes smooth, continuous structures in the data that make post-hoc rule corrections largely unnecessary. In contrast, XGBoost benefits from rule-based diagnostics because of its discrete threshold behavior. Results also reveal that the effectiveness of interpretability techniques depends on preserving systemic feature relationships during preprocessing. The findings unify DNN interpretation with system-theoretic reasoning: the tabular dataset behaves as an interconnected financial system rather than a collection of isolated variables, and the DNN’s learned representation reflects this structure. This positions DNNs as credible, interpretable tools for credit risk modeling when supported by system-aware analysis.
References
2. L. Thomas, “A survey of credit and behavioural scoring,” Int. J. Forecast., 2000
3. H. Gorishniy et al., “Revisiting deep learning models for tabular data,” NeurIPS, 2021.
4. B. Arik and T. Pfister, “TabNet: Attentive interpretable tabular learning,” AAAI, 2021.
5.Guri Zabërgja ”Tabular Data: Is Deep Learning all you need?” arXiv:2402.03970v3 [cs.LG] 05 Oct 2025
6. S. Lundberg and S.-I. Lee, “A unified approach to interpreting model predictions,” NeurIPS, 2017.
7. M. Ribeiro et al., “Why Should I Trust You? Explaining the Predictions of Any Classifier,” KDD, 2016.
8. J. Forrester, Principles of Systems, MIT Press, 1968.
9. I.Rubachev, N.Kartashev, Yury Gorishniy, Artem Babenko, Tabred:Analyzing pitfalls and filling the gaps in tabular deep learning benchmarks arXiv:2406.19380 [cs.LG]
10. R. Little and D. Rubin, Statistical Analysis with Missing Data, Wiley, 2019.
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.
By submitting a paper for publishing the authors hereby comply with the following provisions: 1. The authors retain the copyrights and only give the journal the right for first publication while licensing the work under Creative Commons Attribution License, which grants permissions to others to share the contribution citing this journal as first publication of the text. 2. The authors may enter separate, additional contractual relations for non-exclusive distribution of the published version of the work in this journal (e.g. to upload it in an institutional depository, or to be published in a book), given that they cite the first publication in this journal. 3. The authors are allowed and are encouraged to publish their works online (e.g. to upload it in an institutional depository, personal websites, social networks, etc.) before, during, and after the submission of the paper here, because this may lead to productive exchange, as well as earlier and larger referencing of the published works (see The Effect of Open Access).
