Adaptive Latent-Risk Synthesis: A Graph-Temporal Framework for Real-Time Portfolio Drawdown Forecasting and Risk Attribution

Portfolio risk management systems rely heavily on static covariance models and lagging indicators such as Value-at-Risk (VaR) and Expected Shortfall, which fail to capture rapidly evolving correlation structures during regime transitions. This gap leaves institutional portfolios exposed to sudden drawdowns that conventional models detect only after losses materialize. This paper proposes Adaptive Latent-Risk Synthesis (ALRS), a graph-temporal framework that fuses dynamic asset-correlation graphs with temporal convolutional forecasting to predict portfolio drawdown probability ahead of realization. ALRS introduces four original components: a Dynamic Asset Dependency Graph (DADG) that re-estimates correlation topology at sub-minute intervals, a Temporal Risk Propagation Network (TRPN) modeling contagion across graph edges, a Drawdown Probability Forecaster (DPF) built on a temporal convolutional network with attention pooling, and a Causal Attribution Layer (CAL) decomposing forecasted risk into asset-level contribution scores via Shapley-based attribution. The framework was validated on twelve years of daily multi-asset data spanning 1,200 instruments across equities, fixed income, and commodities, totaling 3.4 million observation-asset pairs. ALRS achieved drawdown-onset prediction accuracy of 91.4%, compared to 73.2% for GARCH-VaR baselines, with a mean forecast lead time of 4.6 trading days before drawdown realization and a 62% reduction in false-positive risk alerts relative to rolling-correlation benchmarks. These results demonstrate that graph-temporal fusion materially improves early risk detection and attribution clarity over conventional volatility-based approaches, offering a practical foundation for next-generation risk infrastructure.