Research report: Generative deep learning for class imbalance in credit risk modeling: comparative evaluation of synthetic oversampling methods and an elastic oversampling extension

When a bank’s model scans thousands of loan files for the handful that will actually go bad, the mathematics is stacked against it. Defaults are rare by design, so a naive classifier can score impressively on accuracy while missing nearly every true default. A new study published in Financial Innovation on 16 October 2026 tackles this imbalance problem head-on, comparing how generative deep learning models and classical synthetic oversampling techniques manufacture artificial default cases to teach classifiers what failure looks like, and it proposes a hybrid method that lets analysts dial the trade-off between catching defaults and raising false alarms up or down.

The study, authored by Tae Yeon Kwon, applies three generative architectures—Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), and Conditional Variational Autoencoders (CVAEs)—as oversampling tools for credit risk data, and pits them against three established interpolation-based methods: SMOTE, SMOTE-Tomek, and ADASYN. Oversampling works by fabricating additional minority-class examples, in this case defaults, until the two classes are balanced in the training data. The paper’s central empirical finding is that the two families behave in nearly opposite ways: synthetic methods that interpolate between existing defaults push models toward aggressive prediction, raising recall at the cost of more false positives, while generative models that learn the underlying distribution of the minority class produce more conservative predictions, favoring precision but missing more defaults.

The experiments rest on two publicly available datasets with very different characters. The first is a firm-level dataset of 4,211 firm-year observations from 571 companies listed on Poland’s NewConnect market between 2007 and 2017, containing 168 defaults—a default rate of just 3.98 percent—described by 26 financial ratios covering profitability, leverage, liquidity, and efficiency. The second is the well-known Taiwanese credit-card customer dataset collected between April and September 2005, with 29,999 customers of whom 6,635 defaulted, a rate of roughly 22.12 percent, described by demographics, credit limits, and monthly records of bills, payments, and delinquency. The authors deliberately treat the two settings as illustrative rather than strictly comparable, since they differ in imbalance severity, feature structure, national regulatory context, and time window.

One technical obstacle had to be cleared before generative models could be applied to tabular credit data at all. Unlike SMOTE, which interpolates between real observations and thereby preserves the binary nature of indicator variables, generative models sample from a continuous latent space, so a feature that should be a clean zero or one emerges as a spuriously continuous value. The authors note in a supplementary check that this procedure is a distributional correction rather than a performance booster, yielding only small and inconsistent performance differences.

On the training sets, computed as averages across the ensemble classifiers, synthetic methods produced more positive predictions and higher recall but lower precision, while generative methods did the reverse. On the test sets the pattern held: for the minority default class alone, synthetic methods achieved the highest F1-scores and generative methods the lowest, but for overall metrics that consider both classes—ROC-AUC, PR-AUC, and accuracy—methods involving generative oversampling, whether pure or in hybrid form, ranked highest within each dataset, with purely synthetic methods ranking lowest.

Visual diagnostics told a consistent story. Using UMAP projections fitted on the oversampled training sets, the authors showed that synthetic methods add samples around existing minority points, preserving mixed regions and failing to form clear boundaries, whereas generative methods can draw diverse minority samples in less-observed regions and reconstruct distinct clusters, sharpening class separation. The paper is candid about failure modes: GANs can suffer mode collapse and VAEs posterior collapse, and the UMAP embeddings showed generative samples concentrated in comparatively narrow, arc-like regions, with GAN-based methods showing larger run-to-run variability across random seeds. The authors characterize this as partial rather than catastrophic diversity loss, noting that generative oversampling still attained the highest ROC-AUC and PR-AUC across both datasets, indicating the generators retained useful structure. Interestingly, no oversampling occasionally outperformed oversampling in test performance, suggesting the benefits of these techniques depend on both the severity of imbalance and the degree of class overlap in the original data.

Unlike earlier sequential hybrids such as SMOTified-GAN, in which SMOTE-generated data feeds into GAN training and the first stage’s limitations can carry over, the parallel design lets both components operate independently on the original minority-class data. In a limited exploratory comparison on the firm dataset, a sequential hybrid attained relatively high precision but lower recall than either SMOTE or GAN alone, giving the lowest F1-score among the configurations considered, while the parallel hybrid achieved higher recall and F1-score across all mixing ratios. The authors caution that these results are exploratory and should be interpreted with caution.

Varying the mixing ratio revealed a clean mechanical relationship. As the synthetic share increases, false positives rise and false negatives fall in both datasets. In the firm dataset, where only 50 defaults appeared in the test set, generative oversampling reduced false negatives by up to six but increased false positives by up to 25 depending on the ratio. The authors frame the mixing ratio not as a globally optimized hyperparameter but as a policy-sensitive parameter reflecting a decision-maker’s tolerance for asymmetric misclassification costs: when the cost of a missed default far exceeds the cost of a false alarm, as in large-exposure lending, a higher synthetic share lowers expected cost; when the two costs are more comparable, as in small-exposure, high-volume lending, a lower share is preferred.

Feature importance, the interpretability side of credit modeling, also shifted with the oversampling method. All oversampling methods produced more concentrated feature-importance distributions than the no-oversampling scenario, underscoring that imbalance correction is needed to identify default determinants at all. But the two families diverged: in the firm dataset’s XGBoost model, synthetic methods yielded a dispersed importance pattern across features, while generative methods concentrated importance sharply and drove some features to near zero. For firms, profitability and capital-structure variables dominated; for individuals, delinquency history, repayment behavior, and credit limits emerged as the dominant predictors. The authors warn that feature importance may reflect either true default determinants or artifacts of partial posterior collapse, and should be interpreted with caution.

The paper is equally explicit about its limits. The sharp gap between training and test performance in the firm dataset indicates overfitting, a known limitation of financial-statement-only prediction, and the small absolute number of test defaults constrains how large the performance differences can be said to be. Cross-dataset comparisons are illustrative rather than strict, and the authors found no meaningful differences among the three generative architectures themselves, which they attribute to the fundamental sparsity of the minority class rather than to architectural choices. Future work, they write, will pursue broader comparisons of parallel and sequential hybrids, formal quantification of class overlap, principled guidance on choosing the mixing ratio, and systematic tuning of generative model configurations. For now, the study offers lenders something more nuanced than a single best algorithm: a menu of trade-offs, and a dial for setting them.

Subject of Research: Comparative evaluation of generative and synthetic oversampling methods for class-imbalanced credit risk modeling

Article Title: Generative deep learning for class imbalance in credit risk modeling: comparative evaluation of synthetic oversampling methods and an elastic oversampling extension

Article References: Kwon, T. Y. (2026). Generative deep learning for class imbalance in credit risk modeling: comparative evaluation of synthetic oversampling methods and an elastic oversampling extension. Financial Innovation, 12(1), Article 161. https://doi.org/10.1186/s40854-026-00970-5

Image Credits: AI Generated

DOI: 10.1186/s40854-026-00970-5

Keywords: credit risk, class imbalance, oversampling, GAN, VAE, CVAE, SMOTE, Elastic Oversampling, default prediction, machine learning, precision-recall trade-off, feature importance