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

Battle Count: 59

The paper is primarily focused on credit risk and mortgage default prediction rather than quantitative trading. However, the concept drift detection and interpretable model adaptation framework is relevant to any ML-based trading system that faces regime changes. The transfer learning correction layer approach could be adapted for trading signal degradation monitoring. The scenario analysis and stress testing capabilities are relevant to portfolio risk management in trading contexts.

The methodology involves multiple steps: building base models (XGBoost), data discretization, rule mining (association rules), PRM calibration via L-BFGS, and correction layer construction. The PRM/Markov Logic Network framework requires understanding of probabilistic graphical models. However, the case study uses only 15+1 single-factor rules, keeping the correction layer relatively simple. The main complexity lies in the rule mining and calibration pipeline rather than the model architecture itself.

The Fannie Mae data is publicly available. The PRM methodology is described in detail including rule mining, calibration (L-BFGS), and the correction layer construction steps. However, no code repository is provided, and specific hyperparameter choices (discretization granularity, number of rules, regularization parameters) are only partially specified. The rule set is provided in the appendix table.