What's the Price of Monotonicity? A Multi-Dataset Benchmark of Monotone-Constrained Gradient Boosting for Credit PD

Battle Count: 82

The paper focuses on credit risk modeling and probability of default estimation, which is relevant to credit portfolio management and lending decisions but not directly to quantitative trading strategies. The findings on monotonicity constraints in gradient boosting could inform credit risk models used in structured finance or credit derivative pricing, but the paper does not address trading, market microstructure, or portfolio optimization directly.

Moderate complexity. The core methodology uses standard gradient boosting libraries (XGBoost, LightGBM, CatBoost) with built-in monotonicity constraint support. However, careful constraint specification requires domain expertise, economic reasoning, and train-only validation. The paired bootstrap procedure and multi-dataset benchmarking add computational overhead. The main challenge is not coding but ensuring economically justified constraint signs and appropriate feature selection.

Replication code and full experiment configuration files available on GitHub (https://github.com/petrkokl/price-of-monotonicity). Fixed global random seed (42) ensures identical splits across all model regimes. All five datasets are publicly available. Hyperparameter grids, constraint specifications, and best hyperparameters are fully documented in tables and appendices. Paired bootstrap with 1,000 replicates provides transparent uncertainty quantification.