Technical Summary: V4FinBench

Problem Statement

Corporate bankruptcy prediction is a high-stakes financial task characterized by severe class imbalance and the need for multi-horizon forecasting. Despite its practical importance for investors, creditors, and regulators, public resources for studying this problem remain scarce. Widely used free benchmarks (e.g., UCI Polish, Taiwanese, and American stock-market datasets) typically contain between 6,000 and 80,000 company-year observations, which is insufficient for meaningful fine-tuning of modern foundation models. Larger, richer resources exist but are behind subscription paywalls, limiting reproducibility and the evaluation of methods that require substantial data for adaptation. Furthermore, existing public datasets often lack multi-horizon labels or consistent distress definitions, making it difficult to evaluate horizon-dependent behaviors or compare methods under realistic financial conditions.

Methodology and Benchmark Construction

To address these gaps, the authors introduce V4FinBench, a public benchmark comprising 1,106,879 company-year observations from the Visegrád Group (V4) economies (Czech Republic, Hungary, Poland, and Slovakia) spanning 2006–2021.

Data and Features

• Source: Data is extracted from the EMIS service, covering 203,900 unique companies.
• Features: Each observation is described by 131 features, including company metadata, financial ratios, growth measures, size measures, and sector-relative indicators.
• Distress Definition: Unlike prior studies relying on formal legal bankruptcy filings, V4FinBench uses a composite financial distress criterion. A company is labeled as distressed in its final reporting year only if it simultaneously satisfies three conditions:
  1. Solvency: Negative equity to total assets ($< 0$).
  2. Profitability: Negative EBITDA to total assets ($< 0$).
  3. Liquidity: Current ratio below 0.6 ($< 0.6$).
• Multi-Horizon Tasks: The benchmark defines six binary prediction tasks corresponding to horizons $h \in \{0, 1, \dots, 5\}$ years ahead. For each horizon, the final $h$ years of data for distressed companies are removed, and the resulting observation is labeled positive.
• Imbalance: The resulting tasks exhibit severe class imbalance, with positive rates ranging from 0.19% to 0.36%.

Evaluation Protocol

The authors define a fixed, reproducible evaluation protocol:

• Folding: 5-fold stratified cross-validation with company-level grouping within countries to prevent data leakage.
• Preprocessing: Missing values are imputed using training-set medians; features are standardized using training-set statistics.
• Metrics: Given the imbalance, F1-score and ROC-AUC are the primary metrics. Decision thresholds are calibrated on the validation fold by maximizing F1 on the precision-recall curve.

Methods Evaluated

The paper evaluates three families of methods:

1. Classical Tabular Baselines: Logistic Regression (LR), Multilayer Perceptron (MLP), Random Forest, XGBoost, CatBoost, and LightGBM.
2. Tabular Foundation Model (TabPFN): A pre-trained in-context learner fine-tuned with imbalance-aware context construction. The authors compare three strategies for constructing the 10,000-sample context:
   • No resampling: Uniform sampling (leaves minority class weakly represented).
   • Random undersampling: Randomly subsampling the majority class to a 0.3 ratio.
   • Prototype undersampling: Selecting majority representatives via K-means clustering to preserve the structure of the non-distressed population.
3. Large Language Model (LLM): Llama-3-8B fine-tuned via QLoRA on serialized financial records formatted as instruction-style conversations.

Key Results

1. TabPFN vs. Classical Baselines

• Context Construction Matters: Prototype undersampling yielded the best results for TabPFN, outperforming random undersampling and no resampling. This indicates that preserving the structural diversity of the majority class is crucial beyond simply increasing minority exposure.
• Performance: With prototype undersampling, TabPFN matched or exceeded gradient-boosted trees (the strongest classical baselines) on ROC-AUC at every horizon. On F1-score, TabPFN matched gradient boosting at the immediate horizon ($h=0$) and surpassed it from $h=2$ onward.

2. LLM Performance (Llama-3-8B)

• Comparison: QLoRA-finetuned Llama-3-8B trailed gradient-boosted trees (specifically XGBoost) on ROC-AUC at every horizon and was generally weaker on F1-score.
• Horizon Dependency: The LLM performed best at the immediate horizon but its ranking ability declined sharply beyond $h=0$, remaining close to chance at longer horizons. The gap between the LLM and XGBoost widened significantly as the prediction horizon increased.

3. Cross-Dataset Transfer

• External Validation: The authors evaluated a V4FinBench-finetuned TabPFN checkpoint on the American Bankruptcy Dataset (a different economy, period, and feature set).
• Result: The fine-tuned model improved upon the vanilla (pre-trained) TabPFN on both ROC-AUC (0.818 $\to$ 0.842) and F1-score (0.372 $\to$ 0.409). This suggests that the adaptation captured transferable financial-distress structure rather than merely overfitting to V4-specific patterns.

Significance and Contributions

The paper makes three primary contributions:

1. V4FinBench Dataset: A large-scale, public corporate-distress benchmark designed specifically to support foundation-model fine-tuning. It covers 1.1M observations across four countries, six prediction horizons, and a composite distress criterion, addressing the scarcity of large, freely accessible financial datasets.
2. Reproducible Evaluation Protocol: A standardized framework for comparing classical tabular methods, tabular foundation models, and LLM-based approaches under realistic class imbalance and multi-horizon forecasting conditions.
3. Empirical Insights:
   • Demonstrates that imbalance-aware context construction (specifically prototype undersampling) is critical for adapting tabular foundation models to severely imbalanced financial data, allowing them to compete with or exceed strong gradient-boosted baselines.
   • Shows that, in this structured prediction setup, serialized-record LLM fine-tuning is not yet competitive with strong tabular baselines, particularly for longer prediction horizons.
   • Provides preliminary evidence that benchmarks like V4FinBench can support adaptation that transfers to external distributions, capturing reusable financial structures.

The authors emphasize that V4FinBench is intended for benchmarking and research, not for direct automated decision-making. They note limitations regarding geographic scope (Central Europe) and the use of composite financial distress labels rather than formal legal bankruptcy filings, urging caution in deployment without additional validation and fairness analysis.