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

Battle Count: 146

The paper focuses on credit risk assessment for SME lending rather than trading strategies. However, the temporal alignment methodology and stacking architecture could inform credit portfolio risk monitoring, which indirectly affects fixed-income and credit-linked trading decisions. The PD estimation framework is relevant for credit spread modeling but not directly applicable to equity or derivatives trading.

Individual models are simple (logistic regression, OLS), but the overall architecture involves multiple components: temporal alignment logic, EWMA calibration, exponential interpolation between anchor points, size-specific delta shifts, and the stacking meta-learner. The operational complexity lies in managing asynchronous data pipelines (3-9 month balance sheet lags, 2-month CR lags) and maintaining point-in-time consistency. The modular design reduces retraining complexity when adding new features.

The methodology is well-described with explicit equations and coefficient tables, but the dataset is proprietary (illimity bank internal portfolio, Central Credit Register data restricted to internal clients). No code or data repository is provided. The CR data bottleneck limits external replication. Parameters (alpha, k, delta shifts) are market-specific and calibrated on internal data.