PaMIR: Open Benchmark of Public Credit-Default Datasets

TL;DR

PaMIR is an open benchmark for credit-default prediction featuring 19 public datasets from nine countries, totaling 1.24 million loan, firm, and card account records with binary default labels. It supports evaluation under realistic conditions including delayed labels and repeated i.i.d. splits, with a synthetic data harness to prevent data leakage.

Why It Matters

The benchmark addresses critical challenges in credit risk modeling, such as label scarcity and delayed labeling, by providing a standardized, reproducible evaluation framework. It enables fair comparison of models across diverse real-world financial data while minimizing data leakage risks.

Key Findings

• PaMIR consolidates 19 public credit-default datasets into a unified, audited benchmark with consistent preprocessing and no redistribution.
• Evaluation includes both standard i.i.d. splits and label-delayed streaming scenarios, reflecting real-world deployment constraints.
• A synthetic data harness allows testing model generalization without exposing the generator to held-out data.