Forecasting financial cascades: A heterogeneous agent-based approach with Tsallis entropy

Financial cascades represent a critical source of systemic risk in complex markets, yet their early detection remains a challenging problem due to nonlinear interactions and heterogeneous agent behavior. In this paper, we propose a unified early-warning framework that integrates heterogeneous agent-based modeling with Tsallis entropy to capture and monitor emerging systemic fragility. The proposed model incorporates three complementary classes of agents — Particle Swarm Optimization (PSO), Genetic Algorithm (GA), and Ant Colony Optimization (ACO) — to represent social learning, adaptive evolution, and path-dependent reinforcement, respectively. These heterogeneous interactions generate rich market dynamics, which are summarized at the system level through a non-extensive entropy measure applied to the distribution of agent sentiment. A data-driven entropy threshold is introduced to identify transitions into high-risk regimes. Extensive Monte Carlo simulations demonstrate that the proposed framework achieves strong predictive performance. Tsallis entropy significantly outperforms classical indicators, including Shannon entropy, volatility, approximate entropy, and return variance, in terms of both AUROC and AUPRC. Ablation analysis further shows that heterogeneity is essential for robust early-warning capability, while lead-time analysis highlights the importance of temporal stability in practical applications. Overall, the results suggest that combining heterogeneous swarm intelligence with entropy-based monitoring provides an effective and interpretable approach for early detection of financial cascades in complex adaptive systems.