Paper: arXiv 2311.06292
Authors: Abel Sancarlos, Edgar Bahilo, Pablo Mozo, Lukas Norman, Obaid Ur Rehma, Mihails Anufrijevs
Abstract
The European debt purchase market as measured by the total book value of purchased debt approached 25bn euros in 2020 and it was growing at double-digit rates. This is an example of how big the debt collection and debt purchase industry has grown and the important impact it has in the financial sector. However, in order to ensure an adequate return during the debt collection process, a good estimation of the propensity to pay and/or the expected cashflow is crucial. These estimations can be employed, for instance, to create different strategies during the amicable collection to maximize quality standards and revenues. And not only that, but also to prioritize the cases in which a legal process is necessary when debtors are unreachable for an amicable negotiation. This work offers a solution for these estimations. Specifically, a new machine learning modelling pipeline is presented showing how outperforms current strategies employed in the sector. The solution contains a pre-processing pipeline and a model selector based on the best model calibration. Performance is validated with real historical data of the debt industry.
Complexity vs Empirical Score
- Math Complexity: 4.0/10
- Empirical Rigor: 6.5/10
- Quadrant: Street Traders — empirical and applied, lower math complexity
Why this score: The paper introduces advanced calibration metrics (ECCE-MAD, ECCE-R) but the math is mostly applied from prior work, not dense derivation; empirical rigor is strong with real historical data validation and a detailed ML pipeline, though no explicit code or backtest metrics are provided.
Research Flowchart
flowchart TD
A["Research Goal:<br>Improve Debt Collection ROI<br>via ML"] --> B["Data Source:<br>Historical Debt Data<br>Europe 2020"]
B --> C["Methodology:<br>ML Pipeline with<br>Advanced Pre-processing"]
C --> D{"Model Selection<br>Optimized for Calibration"}
D --> E["Output:<br>Predicted Propensity to Pay & Cashflow"]
E --> F["Outcome:<br>Outperforms Current<br>Industry Strategies"]
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 proposes a new machine learning modelling pipeline for estimating propensity to pay and expected cashflow in debt collection.
Source sentence
Specifically, a new machine learning modelling pipeline is presented showing how outperforms current strategies employed in the sector.
Assumes 5
- The market for debt purchase is large and growing at double‑digit rates. market
Source sentence
The European debt purchase market as measured by the total book value of purchased debt approached 25bn euros in 2020 and it was growing at double-digit rates.
- Propensity to pay can be estimated for each debtor for the next seven days. model
Source sentence
In particular, the goal is to estimate the propensity to pay in the next seven days at the beginning of each week.
- Strictly proper scoring rules (Brier and Logarithmic scores) provide reliable assessment of forecast quality. statistical
Source sentence
HPO job of the model generator in Section 3.1 is done in function of a strictly proper scoring rule.
- The proprietary dataset is representative of the debt collection industry. data
Source sentence
The proposal in Section 3 is implemented on a proprietary dataset, generously provided by an important debt collection company.
Show 1 more
- Calibration metrics (ECCE) are more reliable than traditional ECE metrics for model selection. statistical
Source sentence
we showed practically the better reliability of the latest ones to perform model selection.
Data
- Proprietary debt collection dataset other proprietary
Source sentence
The data captures the time window from 2022-08 to 2023-04.
Costs
Not applicable
Finds 3
- The uncalibrated LightGBM model reduces error by 61% compared with Platt calibration and achieves a lower Logarithmic score. performance
Source sentence
the uncalibrated model has an error reduction in 61% (with respect to Platt calibration) and not only that, but also a lower Logarithmic score.
- ECE metrics can be misleading for calibration assessment, while ECCE metrics provide more stable model ranking across bin choices. statistical
Source sentence
the unreliable ECE metrics can lead not only to select wrongly the best calibrated model but also to have potentially big mistakes
- Using ECCE‑R metric and Logarithmic score in the pipeline improves propensity‑to‑pay model selection and reduces over‑confidence. predictive-accuracy
Source sentence
the pipeline workflow is designed to address the application specifications, taking into account several criteria considering strictly proper scoring rules (in particular, the Logarithmic score)
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 17 related papers). Coverage is partial, so this is neither replication nor refutation.
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (29)
- Metrics of Calibration for Probabilistic Predictions (2022)
- Verified Uncertainty Calibration (2019) — link
- Evaluating Probabilistic Classifiers: The Triptych (2023) — link
- Machine learning for predicting propensity-to-pay energy bills (2023) — link
- Model Comparison and Calibration Assessment: User Guide for Consistent Scoring Functions in Machine Learning and Actuarial Practice (2022) — link
- Predicting Good Probabilities with Supervised Learning (2005) — link
- Reliability of Subjective Probability Forecasts of Precipitation and Temperature (1977) — link
- Increasing the Reliability of Reliability Diagrams (2007) — link
and 21 more
- Stable reliability diagrams for probabilistic classifiers (2021) — link
- Mitigating Bias in Calibration Error Estimation (2022) — link
- An Empirical Distribution Function for Sampling with Incomplete Information (1955) — link
- Isotone Optimization in R: Pool-Adjacent-Violators Algorithm (PAVA) and Active Set Methods (2009) — link
- Characterizing the optimal solutions to the isotonic regression problem for identifiable functionals (2022) — link
- Obtaining Well Calibrated Probabilities Using Bayesian Binning (2015) — link
- On Calibration of Modern Neural Networks (2017) — link
- Estimating reliability and resolution of probability forecasts through decomposition of the empirical score (2012)
- Deep Credit Risk: Machine Learning with Python (2020) — link
- Calibration of Neural Networks using Splines (2021) — link
- Cumulative deviation of a subpopulation from the full population (2021) — link
- Calibration of P-values for calibration and for deviation of a subpopulation from the full population (2023) — link
- A bias-corrected decomposition of the Brier score (2012) — link
- Investigating the impact of calibration on the quality of explanations (2023) — link
- Get the best from your scikit-learn classifier: trusted probabilties and optimal binary decision (2023) — link
- VERIFICATION OF FORECASTS EXPRESSED IN TERMS OF PROBABILITY (1950) — link
- Strictly Proper Scoring Rules, Prediction, and Estimation (2007) — link
- Lightgbm: A highly efficient gradient boosting decision tree (2017)
- Spline-Based Probability Calibration (2018) — link
- XGBoost : A Scalable Tree Boosting System (2016) — link
- European Consumer Debt Collection and Debt Purchase Market Insight Report 2021 (2021)
0 of 29 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.