V4FinBench: A Million-Record Benchmark for Corporate Bankruptcy Prediction

Forum topic · 小凯 · 2026-05-13

Summary

V4FinBench is a new benchmark dataset for corporate bankruptcy prediction, introduced by Marcin Kostrzewa, Sebastian Tomczak, and Roman Furman in arXiv paper 2505.07227. Corporate bankruptcy prediction is a high-stakes financial task marked by severe class imbalance and multi-horizon forecasting requirements, yet existing public datasets are scarce and small: widely used free benchmarks contain only 6,000 to 80,000 company-year observations, while larger resources sit behind subscription paywalls. V4FinBench addresses this gap with over one million company-year records drawn from the Visegrád Group (V4) economies—Czechia, Hungary, Poland, and Slovakia—covering 2006 to 2021. The benchmark is designed to evaluate tabular foundation models, large language models, and standard machine learning baselines on realistic, imbalanced financial data at scale, enabling more rigorous comparison of modern approaches for bankruptcy forecasting.

Paper Overview

Research area: Machine Learning
Authors: Marcin Kostrzewa, Sebastian Tomczak, Roman Furman
Published: 2025-05-09
arXiv: 2505.07227

Abstract

Corporate bankruptcy prediction is a high-stakes financial task characterized by severe class imbalance and multi-horizon forecasting demands. Public datasets supporting it remain scarce and small: widely used free benchmarks contain between 6,000 and 80,000 company-year observations, while larger resources are behind subscription paywalls.

To address this gap, the authors introduce V4FinBench, a benchmark of over one million company-year records from the Visegrád Group (V4) economies (2006–2021). The dataset is built to evaluate a broad range of model families—including tabular foundation models, large language models (LLMs), and standard machine learning methods—on realistic bankruptcy prediction tasks at unprecedented scale for a free, public resource.

Why It Matters

Scale: 1M+ company-year records vs. 6K–80K in existing free benchmarks
Coverage: Four V4 economies over 16 years (2006–2021)
Task realism: Severe class imbalance and multi-horizon forecasting
Openness: Freely available, unlike larger paywalled alternatives

---

*Auto-collected on 2026-05-13*