Paper: arXiv 2512.17945

Authors: Petr Koklev

Abstract

Financial institutions face a trade-off between predictive accuracy and interpretability when deploying machine learning models for credit risk. Monotonicity constraints align model behavior with domain knowledge, but their performance cost - the price of monotonicity - is not well quantified. This paper benchmarks monotone-constrained versus unconstrained gradient boosting models for credit probability of default across five public datasets and three libraries. We define the Price of Monotonicity (PoM) as the relative change in standard performance metrics when moving from unconstrained to constrained models, estimated via paired comparisons with bootstrap uncertainty. In our experiments, PoM in AUC ranges from essentially zero to about 2.9 percent: constraints are almost costless on large datasets (typically less than 0.2 percent, often indistinguishable from zero) and most costly on smaller datasets with extensive constraint coverage (around 2-3 percent). Thus, appropriately specified monotonicity constraints can often deliver interpretability with small accuracy losses, particularly in large-scale credit portfolios.

Complexity vs Empirical Score

  • Math Complexity: 3.0/10
  • Empirical Rigor: 7.0/10
  • Quadrant: Street Traders — empirical and applied, lower math complexity

Why this score: The paper’s core methodology relies on standard statistical metrics (AUC) and bootstrap uncertainty quantification, with relatively light mathematical formalism. However, it demonstrates high empirical rigor through its multi-dataset benchmark across five public credit datasets, three libraries, and a paired evaluation design with bootstrap uncertainty, making it heavily data- and implementation-focused.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Quantify Price of Monotonicity (PoM)<br>in Credit PD Models"] --> B["Methodology<br>Gradient Boosting:<br>Unconstrained vs Monotone-Constrained"]
  B --> C["Data Processing<br>5 Public Credit Datasets<br>(5-Fold Cross-Validation)"]
  C --> D["Computational Process<br>Bootstrap Paired Comparisons<br>Estimate Performance & Uncertainty"]
  D --> E["Key Findings<br>PoM in AUC: 0% to 2.9%<br>Low cost on large datasets<br>Higher cost on small datasets"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper benchmarks the Price of Monotonicity (PoM) for gradient boosting credit PD models across five public datasets and three libraries.

Source sentence
This paper benchmarks the Price of Monotonicity (PoM) for gradient boosting credit PD models across five public datasets and three libraries.

Assumes

No explicit assumptions found in the text.

Data

  • Polish Bankruptcy benchmark public
    Source sentence
    Polish Bankruptcy (2.4% AUC, 12.9% Brier) and almost no cost in the largest datasets (0.1–0.2% AUC PoM).
  • Lending Club benchmark public
    Source sentence
    large consumer credit datasets constraints can be almost costless (AUC PoM near zero, as in Lending Club).
  • Give Me Some Credit benchmark public
    Source sentence
    – monotonicity constraints impose minimal or negligible performance penalties.
  • German Credit benchmark public
    Source sentence
    Calibration responses are notably heterogeneous: Brier PoM ranges from about –0.7% (German Credit, CatBoost) to 13.0% (Polish Bankruptcy, CatBoost).

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • Mean AUC PoM across libraries is modest, ranging from 0.6% (CatBoost) to 0.9% (XGBoost and LightGBM). performance
    Source sentence
    The library-level average AUC PoM ranges from 0.6% (CatBoost) to 0.9% (XGBoost and LightGBM).
  • Mean Brier PoM is larger, ranging from 2.6% (CatBoost) to 3.2% (LightGBM), with a notable outlier of 12.9% in Polish Bankruptcy. performance
    Source sentence
    ... library-level means range from 2.6% (CatBoost) to 3.2% (LightGBM), with a notable outlier in the Polish Bankruptcy dataset (12.9% mean Brier PoM across libraries).
  • On large datasets constraints are almost costless, with AUC PoM typically below 0.2% and sometimes statistically indistinguishable from zero. predictive-accuracy
    Source sentence
    On large datasets with abundant observations – Give Me Some Credit (n ≈ 150,000) and Lending Club (n ≈ 1,000,000) – monotonicity constraints impose minimal or negligible performance penalties, with AUC PoM typically below 0.2% and often statistically indistinguishable from zero.
  • Polish Bankruptcy shows the highest costs, with AUC PoM between 2.0% and 2.9% and Brier PoM around 13%, driven by high constraint coverage and moderate sample size. negative
    Source sentence
    The highest costs emerge on Polish Bankruptcy, where AUC PoM ranges from 2.0% to 2.9% across all three libraries
  • Statistical significance of PoM varies: 67% of dataset‑library pairs show significant AUC PoM, with XGBoost and CatBoost each significant in 80% of their pairs. statistical
    Source sentence
    For AUC, 67% of dataset‑library pairs show statistically significant PoM, with XGBoost and CatBoost each achieving significance in 80% of their dataset pairs.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 35 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.