Dual-Level Credit Assessment

• Dual-level credit assessment is a structured framework integrating borrower-centric metrics and contextual features to improve risk evaluation.
• It leverages hierarchical models, feature fusion, and network analysis to capture both individual and systemic risk factors.
• Empirical results demonstrate enhanced ROC-AUC and F1 scores compared to single-level approaches, confirming its practical benefit in credit risk prediction.

Dual-level credit assessment is a methodological paradigm in credit risk modeling that integrates, in a structured or hierarchical fashion, two distinct analytical strata or “levels.” These typically correspond to: (1) a primary or individual-level assessment—often anchored in traditional credit attributes, behavioral or application data, or legal/economic fundamentals—and (2) a secondary or contextual-level analysis that incorporates auxiliary, indirect, or network-derived features, often leveraging advanced machine learning, representation learning, or network science. The dual-level approach aims to improve both the predictive performance and the interpretability of credit decisions by exploiting complementary information sources and modeling structures.

1. Motivations and Conceptual Foundations

Credit assessment in its classical form relies primarily on borrower-centric metrics such as credit scores, repayment histories, and demographic variables. However, such primary-level signals are frequently incomplete or non-discriminative, especially for underbanked populations, novel business models, or rapidly evolving market contexts. Dual-level credit assessment structures introduce a secondary analytical layer to capture additional nonstandard factors—such as peer influence, geographic mobility, social reputation, or commercial network exposure—which augment and refine the baseline assessment.

Distinct instantiations in the literature include:

• The combination of traditional features (e.g., FICO, grade) and “secondary” machine-learned attributes in P2P lending (Bhuvaneswari et al., 2020).
• The fusion of commercial creditworthiness with social/behavioral reputation under federated architectures (Hoang et al., 2021).
• Modeling of borrower-specific risk in multilayer networks reflecting both local (individual) and systemic (network) exposures (Óskarsdóttir et al., 2020).
• Hierarchical deep learning on spatiotemporal footprints (region-level + trajectory/user-level) (Han et al., 2019).
• Architectures for aggregate interpretability using subscale-driven two-layer models (Chen et al., 2018).
• Multi-agent LLM systems decomposing risk and reward assessments between specialist teams (Jajoo et al., 30 Jul 2025).
• Functional risk mapping as a function of loan size and sale-side covariates (Zhang et al., 18 Jun 2025).

2. Dual-Level Model Structures and Information Flows

Dual-level architectures are typically instantiated as hierarchies, multi-input ensembles, or multi-agent systems. The two principal forms are:

1. Primary-Secondary Aggregation (“Stacked”):
   • Separate models produce primary ($S_{\mathrm{primary}}$) and secondary ($S_{\mathrm{secondary}}$) risk scores using, for example, credit history and engineered ML features. These are linearly combined into a composite score:

   $$S_{\mathrm{composite}} = w_1 S_{\mathrm{primary}} + w_2 S_{\mathrm{secondary}}, \quad w_1 + w_2 = 1$$

   Optimal weights are determined via validation to maximize AUC or F1 (Bhuvaneswari et al., 2020).

2. Hierarchical Feature Decomposition:

   • Layer 1 maps input features to interpretable subscales (piecewise constant, monotonic functions grouped into conceptually meaningful blocks), with Layer 2 aggregating subscale scores to predict global risk:

   $$R(x)=\sigma(\alpha_0 + \sum_{k=1}^K \alpha_k s_k(x))$$

   where each $s_k(x)$ is a sigmoid-transformed subscale, and $\alpha_k$ its weight (Chen et al., 2018).

3. Representation Fusion or Federated Modalities:

   • Distinct neural encoders map financial, social, contextual, and technological modalities to a unified embedding, with individual scores computed per level (e.g., commercial, social), then combined convexly:

   $$S_i^{\text{unified}} = \alpha s_i^{\mathrm{ccw}} + (1-\alpha) s_i^{\mathrm{sr}}$$

   with $\alpha$ tuned for segment-specific optimality (Hoang et al., 2021).

4. Network-Augmented Assessment:

   • Borrower-centric models are augmented with network-derived exposure features based on multilayer PageRank or degree in product/geographic networks. Features are concatenated and input to a joint regression or boosting model (Óskarsdóttir et al., 2020).
5. Multi-Agent and Multi-Task Frameworks:
   • Credit decisions are modularized as a series of LLM-based agent tasks (e.g., pre-processing, feature engineering, specialized risk/reward modeling) with explicit inter-agent communication protocols and decision orchestration (Jajoo et al., 30 Jul 2025).
6. Generative and Functional Risk Modeling:
   • Conditional generative models map covariates to full outcome distributions (e.g., sales), from which risk measures (VaR, CVaR, PD as function of loan size $l$) are estimated:

   $r_1(l) = \Pr\{L(l) > 0\}, \quad L(l) = (l - rY)^+$

   and the optimal loan amount $l^*$ is computed subject to functional risk constraints (Zhang et al., 18 Jun 2025).

3. Feature Engineering and Selection at Multiple Levels

A defining aspect of dual-level frameworks is the systematic selection and validation of features at both primary and secondary levels.

• Primary Features: Canonically include current and historical credit scores (e.g., FICO), lender grades, basic demographics, and loan characteristics. Normalization and banding convert ratings to numeric risk indices for aggregation (Bhuvaneswari et al., 2020).
• Secondary Features: Identified through correlation filtering, regularization, and machine learning importances (Elastic Net, RF Gini, χ²). Typical examples: debt-to-income ratio (DTI), home ownership, loan purpose, revolving utilizations, inquiry counts, mobility entropies, or social graph features (Bhuvaneswari et al., 2020, Han et al., 2019, Hoang et al., 2021).
• Contextual and Network Features: Engineered from borrower connectivity in multilayer or bipartite networks, including degree counts, the number of defaulted neighbors, multilayer personalized PageRank scores, and intersectional exposures (Óskarsdóttir et al., 2020).
• Unstructured/Modal Data: For federated and deep learning systems, feature sets may reach $2\cdot 10^4$–$3\cdot 10^4$ dimensions prior to dimensionality reduction by L1-regularization or autoencoder bottlenecks (Hoang et al., 2021).

4. Modeling Methodologies and Training Paradigms

Prevalent modeling techniques in dual-level frameworks include:

• Classical ML classifiers: Multiclass logistic regression (one-vs-rest, Elastic Net), random forests (Gini impurity, with cross-validated hyperparameters), linear SVMs (SGD, hinge loss), and boosted tree ensembles (XGBoost) (Bhuvaneswari et al., 2020, Óskarsdóttir et al., 2020).
• Hierarchical Deep Architectures: Graph convolutional networks (GCN) on region graphs with learned attention over multiple adjacency types, followed by temporal sequence models (GRU + attention) to aggregate user trajectory embeddings (Han et al., 2019).
• Federated Learning: Distributed multi-modal neural networks trained via FedAvg, supporting privacy-preserving collaborative optimization across data silos, with gradient clipping and additive Gaussian noise for differential privacy (Hoang et al., 2021).
• Generative Quantile Networks: Quantile-Regression-based Generative Metamodeling (QRGMM) or its Deep Factorization Machine (DeepFM) variant to learn inverse CDFs of sales, enabling Monte Carlo estimation of functional risk curves with uniform consistency guarantees (Zhang et al., 18 Jun 2025).
• Multi-Agent Systems: Layered orchestration via JSON-communicating LLM agents specializing in data processing, contextualization, sub-modeling, risk-reward optimization, and fairness monitoring (Jajoo et al., 30 Jul 2025).

5. Composite Scoring and Decision Synthesis

Integration of dual-level evidence for actionable credit decisions generally follows convex aggregation or meta-learning schemes. Notable strategies:

• Linear or convex score integration:

$$S_{\mathrm{final}} = w_1 S_{\mathrm{primary}} + w_2 S_{\mathrm{secondary}},\quad w_1 + w_2 = 1$$

(Bhuvaneswari et al., 2020, Hoang et al., 2021).

• Nonlinear aggregation or decision-rule stacking, e.g., two-layer additive risk models or deep meta-classifiers (Chen et al., 2018, Jajoo et al., 30 Jul 2025).
• Threshold-based approvals, with $\tau$ tuned to specified trade-offs between recall and precision, or risk budget constraints (e.g., PD or CVaR thresholds for loan approval or size setting) (Bhuvaneswari et al., 2020, Zhang et al., 18 Jun 2025).
• Adaptive weighting ($\alpha$) for segment-specific optimization (e.g., higher emphasis on commercial attributes for banked, social reputation for unbanked) (Hoang et al., 2021).

6. Evaluation Protocols and Empirical Findings

Dual-level architectures consistently demonstrate empirical uplift over single-level baselines, as measured by micro-averaged F1, ROC-AUC, precision, and recall.

Source	Baseline (AUC)	Dual-Level (AUC)	ΔAUC	F1 Gain
LendingClub (Bhuvaneswari et al., 2020)	0.68	0.73	+0.05	+0.07
Federated AI (Hoang et al., 2021)	0.81 (CCW) / 0.68 (SR)	0.83 (unified)	+0.02 (banked)	+0.03 (recall)
CreditPrint (Han et al., 2019)	0.707	0.784	+0.077	--
Networks (Óskarsdóttir et al., 2020)	0.639 / 0.660	0.703 / 0.737	~+0.06–0.08	--
MASCA (Jajoo et al., 30 Jul 2025)	58.5 (F1, single-level)	66.9 (F1, dual-level)	--	+8.4 pp

A key result is that the addition of well-validated secondary or contextual features contributes significant discriminative power, even after extensive tuning of traditional models (Bhuvaneswari et al., 2020, Óskarsdóttir et al., 2020). Dual-level systems further support more rigorous fairness, model transparency, and robust scenario simulation (Jajoo et al., 30 Jul 2025, Hoang et al., 2021).

7. Interpretability, Deployment, and Future Directions

Interpretability is addressed at both layers via mechanisms such as grouped subscales (coarse: e.g., “Delinquency,” “TradeOpenTime”) and within-subscale feature thresholds/rules (fine) (Chen et al., 2018). Set-cover explanations and SHAP are recommended for transparent auditing of feature contributions.

Deployment best practices include:

• Periodic retraining of secondary models to accommodate drift (Bhuvaneswari et al., 2020).
• Privacy-preserving federated learning for regulated environments (Hoang et al., 2021).
• Real-time score and threshold monitoring to maintain calibrated approval rates (Bhuvaneswari et al., 2020).
• Fairness auditing and group-specific threshold calibration (Jajoo et al., 30 Jul 2025).

Research frontiers comprise dynamic meta-learning of aggregation weights, extension to multi-task outputs (credit line, delinquency prediction), integration of spiking networks or cognitive-inspired reasoning, and continuous risk/function estimation across loan sizes. The inclusion of advanced generative modeling, network science, and agent-based hierarchical design is progressively shaping dual-level credit assessment as the new default paradigm for heterogeneous, high-stakes lending environments.

---

References:

• (Bhuvaneswari et al., 2020) Determining Secondary Attributes for Credit Evaluation in P2P Lending
• (Hoang et al., 2021) Federated Artificial Intelligence for Unified Credit Assessment
• (Han et al., 2019) CreditPrint: Credit Investigation via Geographic Footprints by Deep Learning
• (Chen et al., 2018) An Interpretable Model with Globally Consistent Explanations for Credit Risk
• (Zhang et al., 18 Jun 2025) Conditional Generative Modeling for Enhanced Credit Risk Management in Supply Chain Finance
• (Jajoo et al., 30 Jul 2025) MASCA: LLM based-Multi Agents System for Credit Assessment
• (Óskarsdóttir et al., 2020) Multilayer Network Analysis for Improved Credit Risk Prediction