An Entropy-Regularised AI Framework for Multi-Asset Volatility Spillover Forecasting and CVaR-Constrained Portfolio...

Area of Science:

• Quantitative Finance
• Artificial Intelligence
• Information Theory

Background:

• Accurate forecasting of multi-asset volatility spillovers is crucial for risk-aware portfolio management.
• Existing methods often struggle with directional information flow, state compression, calibrated uncertainty, and tail-risk limits.

Purpose of the Study:

• To propose TDV (Transfer-entropy, Dynamic-graph-attention, Variational-information-bottleneck), an AI framework for enhanced volatility spillover forecasting and portfolio construction.
• To demonstrate the practical value of calibrated predictive distributions in entropy-regulated, CVaR-constrained portfolios.

Main Methods:

• Coupling a time-varying transfer entropy network with a graph attention encoder.
• Regularizing the model with a variational information bottleneck.
• Establishing theoretical results on estimator consistency, generalization bounds, and CVaR feasibility.

Main Results:

• TDV reduced spillover forecasting errors by 24-42% compared to baseline models (LSTM, GAT, Transformer).
• Achieved an out-of-sample R2 of 0.331 and an annualized Sharpe ratio of 1.46 on a 32-asset global panel.
• Demonstrated significant reductions in Conditional Value-at-Risk (CVaR) by 28-36%.

Conclusions:

• The TDV framework offers a robust approach to multi-asset volatility spillover forecasting and risk-aware portfolio optimization.
• The theoretical underpinnings and empirical results validate the effectiveness of the proposed information-theoretic AI framework.