Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Incremental Hybrid Ensemble with Graph Attention and Frequency-Domain Features for Stable Long-Term Credit Risk Modeling

Jiajing Wang

TL;DR

HYDRA-EI tackles long-term credit risk prediction under concept drift by integrating graph-based relational features, automated cross-feature construction, and frequency-domain temporal signals within an incremental ensemble. It combines LightGBM, CatBoost, and DenseLight+ with a Bayesian performance-aware gating mechanism and weekly updates to adapt to drift without full retraining. The framework's multistage features, graph synthesizer, AutoCross, and spectrotemporal encoder improve stability, generalization, and probabilistic calibration, as demonstrated on a private leaderboard where HYDRA-EI outperforms baselines. This approach offers a practical, scalable solution for stable long-horizon credit risk in dynamic financial environments.

Abstract

Predicting long-term loan defaults is hard because borrower behavior often changes and data distributions shift over time. This paper presents HYDRA-EI, a hybrid ensemble incremental learning framework. It uses several stages of feature processing and combines multiple models. The framework builds relational, cross, and frequency-based features. It uses graph attention, automatic cross-feature creation, and transformations from the frequency domain. HYDRA-EI updates weekly using new data and adjusts the model weights with a simple performance-based method. It works without frequent manual changes or fixed retraining. HYDRA-EI improves model stability and generalization, which makes it useful for long-term credit risk tasks.

Incremental Hybrid Ensemble with Graph Attention and Frequency-Domain Features for Stable Long-Term Credit Risk Modeling

TL;DR

HYDRA-EI tackles long-term credit risk prediction under concept drift by integrating graph-based relational features, automated cross-feature construction, and frequency-domain temporal signals within an incremental ensemble. It combines LightGBM, CatBoost, and DenseLight+ with a Bayesian performance-aware gating mechanism and weekly updates to adapt to drift without full retraining. The framework's multistage features, graph synthesizer, AutoCross, and spectrotemporal encoder improve stability, generalization, and probabilistic calibration, as demonstrated on a private leaderboard where HYDRA-EI outperforms baselines. This approach offers a practical, scalable solution for stable long-horizon credit risk in dynamic financial environments.

Abstract

Predicting long-term loan defaults is hard because borrower behavior often changes and data distributions shift over time. This paper presents HYDRA-EI, a hybrid ensemble incremental learning framework. It uses several stages of feature processing and combines multiple models. The framework builds relational, cross, and frequency-based features. It uses graph attention, automatic cross-feature creation, and transformations from the frequency domain. HYDRA-EI updates weekly using new data and adjusts the model weights with a simple performance-based method. It works without frequent manual changes or fixed retraining. HYDRA-EI improves model stability and generalization, which makes it useful for long-term credit risk tasks.

Paper Structure

This paper contains 28 sections, 18 equations, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Multistage Feature Engineering Module
  5. Graph Feature Synthesizer
  6. AutoCross Feature Evolution
  7. SpectroTemporal Encoder
  8. Base Models
  9. LightGBM with GOSS
  10. CatBoost with Incremental Ordered Boosting
  11. DenseLight+ Neural Network
  12. Incremental Training Strategy
  13. Bayesian Performance-aware Ensemble Gating
  14. Global Optimization Objective
  15. Feature Engineering
  16. ...and 13 more sections

Figures (5)

  • Figure 1: Overview of the HYDRA‐EI framework. The architecture consists of three models.
  • Figure 2: The architecture consists of Graph Feature Synthesizer.
  • Figure 3: Wavelet transform analysis of client financial behavior over 104 weeks.
  • Figure 4: Data preprocessing analysis. (a) Population Stability Index (PSI) heatmap showing feature drift across time periods. Features exceeding the 0.2 threshold (marked with red borders) were candidates for removal or re-binning. (b) Visualization of robust normalization effects on heavy-tailed distributions, demonstrating effective outlier handling while preserving data structure..
  • Figure 5: Model indicator change chart.