Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Crisis-Resilient Portfolio Management via Graph-based Spatio-Temporal Learning

Zan Li, Rui Fan

TL;DR

The paper addresses how crisis-induced regime shifts alter asset correlations, breaking traditional fixed-topology graph methods. It introduces CRISP, which learns regime-specific, sparse graphs by integrating temporal encoding via BiLSTM+self-attention, spatial encoding with a prior GCN, and a learnable graph via multi-head Graph Attention Networks to forecast both the next-period graph $A_{t+1}$ and allocations $\mathbf{w}_{t+1}$ over a 5-day horizon. A multi-objective loss optimizes risk-adjusted returns, tail risk, diversification, and turnover, enabling cross-regime generalization from training crises (2008, 2020) to unseen inflation-driven crises (2022–2024). Empirical results show CRISP significantly outperforms equal-weight and static-graph baselines, with Sharpe $3.76$, ~12 edges retained, and pronounced interpretable attention shifts toward defensive connections during crises. This approach offers a principled framework for adaptive, regime-aware forecasting with transparent decision-making, applicable beyond finance to any domain with evolving latent networks.

Abstract

Financial time series forecasting faces a fundamental challenge: predicting optimal asset allocations requires understanding regime-dependent correlation structures that transform during crisis periods. Existing graph-based spatio-temporal learning approaches rely on predetermined graph topologies--correlation thresholds, sector classifications--that fail to adapt when market dynamics shift across different crisis mechanisms: credit contagion, pandemic shocks, or inflation-driven selloffs. We present CRISP (Crisis-Resilient Investment through Spatio-temporal Patterns), a graph-based spatio-temporal learning framework that encodes spatial relationships via Graph Convolutional Networks and temporal dynamics via BiLSTM with self-attention, then learns sparse structures through multi-head Graph Attention Networks. Unlike fixed-topology methods, CRISP discovers which asset relationships matter through attention mechanisms, filtering 92.5% of connections as noise while preserving crisis-relevant dependencies for accurate regime-specific predictions. Trained on 2005--2021 data encompassing credit and pandemic crises, CRISP demonstrates robust generalization to 2022--2024 inflation-driven markets--a fundamentally different regime--by accurately forecasting regime-appropriate correlation structures. This enables adaptive portfolio allocation that maintains profitability during downturns, achieving Sharpe ratio 3.76: 707% improvement over equal-weight baselines and 94% improvement over static graph methods. Learned attention weights provide interpretable regime detection, with defensive cluster attention strengthening 49% during crises versus 31% market-wide--emergent behavior from learning to forecast rather than imposing assumptions.

Crisis-Resilient Portfolio Management via Graph-based Spatio-Temporal Learning

TL;DR

The paper addresses how crisis-induced regime shifts alter asset correlations, breaking traditional fixed-topology graph methods. It introduces CRISP, which learns regime-specific, sparse graphs by integrating temporal encoding via BiLSTM+self-attention, spatial encoding with a prior GCN, and a learnable graph via multi-head Graph Attention Networks to forecast both the next-period graph and allocations over a 5-day horizon. A multi-objective loss optimizes risk-adjusted returns, tail risk, diversification, and turnover, enabling cross-regime generalization from training crises (2008, 2020) to unseen inflation-driven crises (2022–2024). Empirical results show CRISP significantly outperforms equal-weight and static-graph baselines, with Sharpe , ~12 edges retained, and pronounced interpretable attention shifts toward defensive connections during crises. This approach offers a principled framework for adaptive, regime-aware forecasting with transparent decision-making, applicable beyond finance to any domain with evolving latent networks.

Abstract

Financial time series forecasting faces a fundamental challenge: predicting optimal asset allocations requires understanding regime-dependent correlation structures that transform during crisis periods. Existing graph-based spatio-temporal learning approaches rely on predetermined graph topologies--correlation thresholds, sector classifications--that fail to adapt when market dynamics shift across different crisis mechanisms: credit contagion, pandemic shocks, or inflation-driven selloffs. We present CRISP (Crisis-Resilient Investment through Spatio-temporal Patterns), a graph-based spatio-temporal learning framework that encodes spatial relationships via Graph Convolutional Networks and temporal dynamics via BiLSTM with self-attention, then learns sparse structures through multi-head Graph Attention Networks. Unlike fixed-topology methods, CRISP discovers which asset relationships matter through attention mechanisms, filtering 92.5% of connections as noise while preserving crisis-relevant dependencies for accurate regime-specific predictions. Trained on 2005--2021 data encompassing credit and pandemic crises, CRISP demonstrates robust generalization to 2022--2024 inflation-driven markets--a fundamentally different regime--by accurately forecasting regime-appropriate correlation structures. This enables adaptive portfolio allocation that maintains profitability during downturns, achieving Sharpe ratio 3.76: 707% improvement over equal-weight baselines and 94% improvement over static graph methods. Learned attention weights provide interpretable regime detection, with defensive cluster attention strengthening 49% during crises versus 31% market-wide--emergent behavior from learning to forecast rather than imposing assumptions.
Paper Structure (14 sections, 8 equations, 3 figures, 2 tables)

This paper contains 14 sections, 8 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Our Contributions
  3. Methodology
  4. Problem Formulation: Regime-Aware Forecasting
  5. Module 1: Temporal Encoding
  6. Module 2: Spatial Encoding
  7. Module 3: Learnable Graph via GAT
  8. Module 4: Portfolio Optimization and Training
  9. Experimental Evaluation
  10. Dataset and Experimental Setup
  11. Main Results: Superior Crisis-Period Performance
  12. Interpretability and Attention Analysis
  13. Ablation Study
  14. Conclusion

Figures (3)

  • Figure 1: CRISP architecture. Input features ($N=13$ stocks, $T=20$ days, $F=31$ features) are processed through parallel temporal (BiLSTM+Self-Attention) and spatial (GCN with prior knowledge) encoders, producing embeddings $\mathbf{h}_{\text{spat}}$ and $\mathbf{h}_{\text{temp}}$. Core innovation: Multi-head GAT learns sparse graph structure from fully connected topology (156 edges $\rightarrow$$\sim$12 edges, 92.5% filtering) through differentiable attention. LSTM processes graph sequence representations, and MLP generates portfolio weights for 5-day rebalancing periods.
  • Figure 2: Cumulative returns: CRISP vs Equal Weight. CRISP (blue solid, Sharpe 3.76) achieves 62.8% substantially outperforming Equal Weight (orange dashed, Sharpe 0.47, 19.5%) over 710 days with 5-day rebalancing. Performance gap widens during 2022 Q1--Q2 crisis period, demonstrating superior downside protection through accurate regime forecasting.
  • Figure 3: Attention weight evolution during 2022 crisis. Defensive cluster attention (green) rises from 25% pre-crisis average to 38% crisis peak (+49% relative increase) as model concentrates on protective asset relationships. Market average (blue) increases from 13% to 17% (+31%). Selective strengthening—defensive +49% vs market +31%—demonstrates the model learns to concentrate on protective assets without explicit supervision on crisis periods or defensive stock labels. Post-crisis (July 2022+), defensive attention returns to 25% baseline, showing reversible adaptation rather than overfitting.