Reproduce “What's the Price of Monotonicity? A Multi-Dataset Benchmark of Monotone-Constrained Gradient Boosting for Credit PD” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Petr Koklev
This study systematically quantifies the performance cost of imposing monotonicity constraints on gradient boosting models for credit Probability of Default (PD) prediction. It demonstrates that these constraints typically incur modest accuracy penalties, with negligible costs observed on large datasets (AUC PoM often below 0.2%), but can lead to substantial performance degradation (e.g., up to 21.2% AUC PoM) if constraints are extensive or mis-specified.
Key insights
- The "Price of Monotonicity" is highly conditional, with costs varying significantly based on dataset characteristics (size, dimensionality) and the extent of constraint coverage.
- For large datasets (n 100,000), monotonicity constraints often impose minimal or negligible performance penalties, sometimes acting as a form of structural regularization.
- Mis-specified (wrong-sign) constraints lead to significantly larger performance degradation (e.g., ~10-fold increase in discrimination cost) compared to correctly specified ones, highlighting the critical role of accurate domain knowledge.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.