Multi-Horizon Financial Crisis Detection Through Adaptive Data Fusion

Abstract

Financial institutions need 12–18 months’ advance warning to implement effective crisis mitigation strategies. We develop a neural network framework integrating macroeconomic indicators (156 series), textual sentiment (2.8 million documents), and institutional networks (524 banks) through volatility-adaptive temporal alignment and cross-modal attention mechanisms. The system employs stratified classification across three horizons: immediate (1–3 months), medium (4–12 months), and long-term (12–36 months). Testing on 16 years of financial data (January 2008-December 2023) encompassing four crisis episodes spanning multiple countries demonstrates 89.7% accuracy (SD 2.1%; 95% CI: 87.1–92.3%) with median warning times of 16.3 months. Performance improvement reaches 7.6 percentage points over single-source baselines (82.1% to 89.7%, p<0.01). Strict temporal validation prevents data leakage while leave-one-crisis-out testing confirms cross-crisis generalization. Attention weight visualization provides interpretable insights for regulatory compliance, though full causal explanation remains limited. CIs are computed over model-level means across five random seeds (df=4), using identical train/validation/test splits; initialization is the only randomized factor.