Paper #3

A Residual Feature-Gated Variational Autoencoder with Transfer Learning for Economic Systemic Risk Exposure Prediction

Shiqiong Pan

Abstract: Accurate prediction of systemic economic risk is important for financial supervision and policy design, yet existing econometric and machine-learning models often suffer from limited data and unstable generalization. This paper proposes a two-step learning framework for predicting systemic risk exposure using a simulated economic resilience dataset. First, a Variational Autoencoder (VAE) is trained on real observations and used to generate 100,000 synthetic samples for data augmentation. Second, a residual neural network with a feature-gating mechanism is pretrained on the synthetic data and then fine-tuned on real samples through transfer learning. Experiments show that the proposed strategy leads to large performance gains. Compared with training from scratch, the two-stage model improves R² from 0.958 to 0.998, while reducing RMSE from 0.015 to 0.003 and MAE from 0.011 to 0.003. Classical models also benefit from augmentation: for SVR, R² increases from 0.018 to 0.090 after adding synthetic data. Feature screening based on mutual information highlights growth dynamics and risk-control indicators as the most influential drivers of systemic risk. Ablation studies further show that residual connections and feature gating are both essential, with the full model achieving the lowest error levels. These results demonstrate that generative data augmentation combined with transfer learning provides an effective and general solution for economic risk prediction under limited data conditions.

Keywords: Data Augmentation; Financial Stability; Systemic Risk Prediction; Transfer Learning; Variational Autoencoder.