Abstract
Abstract Frequent market instability and the lack of rigorously validated forecasting frameworks pose significant challenges for predicting market stress in the Dhaka Stock Exchange (DSE). This study proposes a leakage-safe machine learning framework for next-day market stress prediction using historical trading data from 2008 to 2022. Technical indicators representing trend, momentum, volatility, and volume are extracted, while market conditions are classified into three categories: Normal, High-Volatility, and Crash. Seven models, namely Logistic Regression, Random Forest, XGBoost, Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Hidden Markov Model (HMM), and GJR-GARCH, are evaluated using expanding-window walk-forward cross-validation with embargo periods. Random Forest achieves the best crash prediction performance with a crash PR-AUC of $$0.1697 \pm 0.0134$$ , followed by XGBoost and LSTM. At the same time, the Friedman test confirms statistically significant differences among the models ( $$\chi ^2 = 19.11$$ , $$p = 0.004$$ ). Permutation Importance and SHAP identify Bollinger Bandwidth and rolling volatility as the most influential predictors. The proposed framework demonstrates that leakage-safe and interpretable machine learning can effectively support market stress prediction and risk-aware decision-making in the DSE.