State-AttentNet: A Dynamic Volatility-Adaptive Hybrid Framework for Market State Classification in Frontier Economies

The proper identification of latent market states is important for effective risk management. However, existing frameworks are often not good at distinguishing the difference between stable situations and systemic crashes in the high entropy environments of frontier economies. To address this challenge, this study presents State-AttentNet, an application-driven market-state classification framework for the Dhaka Stock Exchange. The framework combines a volatility-adaptive labeling scheme with a bidirectional LSTM encoder and temporal attention to classify stable and crash-like market states in the Dhaka Stock Exchange. The market sensitive technical features are synthesized with Bidirectional LSTM encoder of the temporal context into a hierarchical pipeline proposed methodology. This model is also supplemented by Adaptive Temporal Attention model to bring focus to high impact volatility events in a rolling window. Here, ‘crash’ means an adaptive volatility-stress state, not a return direction. Empirical evaluation on a longitudinal 26-year dataset shows that the model achieves 93% classification accuracy and an AUC of 0.97. It outperforms traditional baseline models in discriminating between crash and stable states. The results have significant practical implications for institutional investors as it is a trusted, automated detection instrument. This system is conducive for state contingent risk reduction measures and minimizing false alarms in instances of correction of incidental markets. Lastly, SHAP used to check the operational integrity of the model. This suggests that structurally informative lag signals and not stochastic noise determine classification decisions.