This paper defines a Price of Monotonicity (PoM) metric and benchmarks monotone‑constrained vs unconstrained gradient boosting (XGBoost, LightGBM, CatBoost) across five public credit PD datasets, showing that PoM in AUC is typically near zero on large portfolios (often <0.2%) but can reach ~2–2.9% on smaller, high-coverage datasets, while wrong-sign constraints inflict substantially larger costs; paired bootstrap is used to quantify uncertainty
速读内容
Key empirical finding — PoM is usually small but dataset‑dependent
Mean AUC PoM across libraries is modest (≈0.6–0.9%), with CatBoost slightly lower (≈0.6%) and XGBoost/LightGBM around 0.9% on average .
Dataset heterogeneity: Lending Club and Give Me Some Credit show near‑zero AUC PoM (~0.1–0.2%), while Polish Bankruptcy exhibits the largest costs (AUC PoM ≈2.0–2.9%, Brier PoM ≈12.9%) .
Calibration (Brier/LogLoss) can show larger PoM than discrimination; calibration effects are more variable and sometimes double‑digit for problematic datasets .
Datasets, constraint coverage and examples
Dataset
Train Size
Test Size
Features
Default Rate (%)
Taiwan Credit
21,000
9,000
23
22.1
Polish Bankruptcy
7,352
3,151
64
4.7
German Credit (Statlog)
700
300
20
30.0
Give Me Some Credit
105,000
45,000
10
6.7
Lending Club (Zenodo)
829,347
225,277
8
20.0
Constraint coverage varies widely (3 to 41 constrained features); Polish Bankruptcy has highest coverage (~41 features, predominantly decreasing), Taiwan Credit applies mainly increasing constraints on repayment indicators .
Temporal split applied to Lending Club to avoid look‑ahead; other datasets use stratified 70/30 splits to enable paired comparisons .
Methods & diagnostics — PoM definition, paired bootstrap and wrong‑sign test
PoM is defined as the relative change moving from unconstrained → monotone‑constrained (AUC/PR‑AUC: (AUC_unc - AUC_mono)/AUC_unc; LogLoss/Brier: (metric_mono - metric_unc)/metric_unc) to express cost in percent .
Uncertainty quantified via paired bootstrap (B=1,000) on the test set using identical resamples for both models; 95% percentile CIs indicate statistical detectability of PoM (CI excluding zero) .
Wrong‑sign diagnostic: constraining features in the opposite economic direction yields very large costs (mean wrong‑sign AUC PoM ≈8.6%, Brier PoM ≈11.0%), confirming constraints are binding and sign specification matters greatly .
思维导图
深度阅读
以下是对 Petr Koklev “What’s the Price of Monotonicity? A Multi‑Dataset Benchmark of Monotone‑Constrained Gradient Boosting for Credit PD” 报告的详尽、逐节解读与图表剖析,按报告结构组织并尽量覆盖所有重要数据点、假设、方法与结论(文中以 [] 引用页码以便追溯)。
一、元数据与概览(引言与报告总体概览)
报告标题与作者:What’s the Price of Monotonicity? A Multi‑Dataset Benchmark of Monotone‑Constrained Gradient Boosting for Credit PD,作者 Petr Koklev。该信息见封面摘要部分。
研究主旨:量化在梯度提升(XGBoost/LightGBM/CatBoost)中对特征施加单调性约束(monotonicity constraints)对信用违约概率(PD)预测性能的“价格”(Price of Monotonicity, PoM),并分析 PoM 随数据集规模、违约率、约束覆盖度及库实现的变化。該点在摘要与绪论明确提出。