Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Temporal-Aligned Meta-Learning for Risk Management: A Stacking Approach for Multi-Source Credit Scoring

O. Didkovskyi, A. Vidali, N. Jean, G. Le Pera

TL;DR

This work tackles temporal misalignment in multi-source credit scoring for Italian SMEs by decomposing risk into static annual anchoring and dynamic monthly evolution. It introduces a point-in-time consistency framework and a modular stacking architecture that meta-learns from base models without retraining them when new indicators are added. The static anchor model and dynamic fusion demonstrate superior AUC and stability across time horizons compared with baseline models, with effective translation of PDs into calibrated ratings via size-dependent delta shifts. The approach offers practical advantages for origination and monitoring in environments with asynchronous data, reducing operational complexity while preserving interpretability.

Abstract

This paper presents a meta-learning framework for credit risk assessment of Italian Small and Medium Enterprises (SMEs) that explicitly addresses the temporal misalignment of credit scoring models. The approach aligns financial statement reference dates with evaluation dates, mitigating bias arising from publication delays and asynchronous data sources. It is based on a two-step temporal decomposition that at first estimates annual probabilities of default (PDs) anchored to balance-sheet reference dates (December 31st) through a static model. Then it models the monthly evolution of PDs using higher-frequency behavioral data. Finally, we employ stacking-based architecture to aggregate multiple scoring systems, each capturing complementary aspects of default risk, into a unified predictive model. In this way, first level model outputs are treated as learned representations that encode non-linear relationships in financial and behavioral indicators, allowing integration of new expert-based features without retraining base models. This design provides a coherent and interpretable solution to challenges typical of low-default environments, including heterogeneous default definitions and reporting delays. Empirical validation shows that the framework effectively captures credit risk evolution over time, improving temporal consistency and predictive stability relative to standard ensemble methods.

Temporal-Aligned Meta-Learning for Risk Management: A Stacking Approach for Multi-Source Credit Scoring

TL;DR

This work tackles temporal misalignment in multi-source credit scoring for Italian SMEs by decomposing risk into static annual anchoring and dynamic monthly evolution. It introduces a point-in-time consistency framework and a modular stacking architecture that meta-learns from base models without retraining them when new indicators are added. The static anchor model and dynamic fusion demonstrate superior AUC and stability across time horizons compared with baseline models, with effective translation of PDs into calibrated ratings via size-dependent delta shifts. The approach offers practical advantages for origination and monitoring in environments with asynchronous data, reducing operational complexity while preserving interpretability.

Abstract

This paper presents a meta-learning framework for credit risk assessment of Italian Small and Medium Enterprises (SMEs) that explicitly addresses the temporal misalignment of credit scoring models. The approach aligns financial statement reference dates with evaluation dates, mitigating bias arising from publication delays and asynchronous data sources. It is based on a two-step temporal decomposition that at first estimates annual probabilities of default (PDs) anchored to balance-sheet reference dates (December 31st) through a static model. Then it models the monthly evolution of PDs using higher-frequency behavioral data. Finally, we employ stacking-based architecture to aggregate multiple scoring systems, each capturing complementary aspects of default risk, into a unified predictive model. In this way, first level model outputs are treated as learned representations that encode non-linear relationships in financial and behavioral indicators, allowing integration of new expert-based features without retraining base models. This design provides a coherent and interpretable solution to challenges typical of low-default environments, including heterogeneous default definitions and reporting delays. Empirical validation shows that the framework effectively captures credit risk evolution over time, improving temporal consistency and predictive stability relative to standard ensemble methods.
Paper Structure (21 sections, 2 equations, 4 figures, 7 tables)

This paper contains 21 sections, 2 equations, 4 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Dataset
  3. On data processing
  4. Point-in-Time Consistency
  5. Features
  6. Target
  7. Static Model Target
  8. Dynamic Model Target
  9. Modeling
  10. Static Model Architecture
  11. Dynamic Model Architecture
  12. Rating Assignment and Meta-Learning Integration
  13. Validation Strategies
  14. Results
  15. Discussion
  16. ...and 6 more sections

Figures (4)

  • Figure 1: Temporal misalignment: December 31st balance sheets published 6 months later create a prediction gap with the 12-month horizon
  • Figure 2: Exponential interpolation with k=3 selected through backtesting
  • Figure 3: Fusion architecture combining static balance sheet assessment with dynamic behavioral updates at evaluation point
  • Figure 4: Fusion rating frontier by company size