Probabilistic Rule Models as Diagnostic Layers: Interpreting Structural Concept Drift in Post-Crisis Finance

TL;DR

This paper introduces a probabilistic rule-based correction layer for machine learning models that interprets and explains structural concept drift in high-stakes financial predictions, enhancing transparency and diagnostic capabilities.

Contribution

It proposes a novel correction layer using Probabilistic Rule Models with Markov Logic Networks for interpretable detection of concept drift in finance.

Findings

Successfully explained the impact of the 2008 crisis on borrower segments

Demonstrated interpretability and diagnostic utility of the PRM correction layer

Provided insights for risk management and regulatory compliance

Abstract

Machine learning models used for high-stakes predictions in domains like credit risk face critical degradation due to concept drift, requiring robust and transparent adaptation mechanisms. We propose an architecture, where a dedicated correction layer is employed to efficiently capture systematic shifts in predictive scores when a model becomes outdated. The key element of this architecture is the design of a correction layer using Probabilistic Rule Models (PRMs) based on Markov Logic Networks, which guarantees intrinsic interpretability through symbolic, auditable rules. This structure transforms the correction layer from a simple scoring mechanism into a powerful diagnostic tool capable of isolating and explaining the fundamental changes in borrower riskiness. We illustrate this diagnostic capability using Fannie Mae mortgage data, demonstrating how the interpretable rules extracted by the correction layer successfully explain the structural impact of the 2008 financial crisis on specific population segments, providing essential insights for portfolio risk management and regulatory compliance.