A hybrid deep learning framework for early detection and forecasting of financial crises in local government economies

Abstract

Financial crises severely threaten local government finances, risking fiscal instability, economic slowdown, and reduced social welfare. Effective risk management requires sophisticated early warning systems. Conventional econometric models struggle with nonlinearity and dynamics, limiting real-world predictive accuracy. This paper introduces BiTransAttnNet, a hybrid deep learning model for early detection and forecasting of local government financial crises. The framework integrates Bidirectional LSTM, Transformer encoders, and feature-level attention to capture complex feature interactions and nonlinear risk patterns. Using a publicly available benchmark dataset comprising 2000 city-year observations across fiscal, financial, and ecological indicators, the study employs a rigorous preprocessing and validation framework for methodological evaluation. BiTransAttnNet outperforms baselines (TabNet, BiLSTM, Random Forest, SVM), achieving 97.33% accuracy, 98.06% precision, 94.39% recall, and 96.19% F1-score. The model demonstrates strong generalization, robustness, and SHAP-based interpretability analysis of the original fiscal indicators, revealing that expenditure, fiscal revenue, and debt-related variables are the primary contributors to predicted financial risk. The interpretability analysis is limited to the original fiscal indicators rather than the complete engineered feature space.