Regime Dependent Machine Learning and the Limits of Algorithmic Trading in the Dhaka Stock Exchange

Abstract This study tests whether Hidden Markov Model (HMM) regime-conditioning improves short-horizon return forecasting and trading profitability for equities listed on the Dhaka Stock Exchange (DSE), using three independently constructed stock panels spanning 42 securities across 17 sectors: a deliberately selected twelve-stock Fixed panel, an indepen-dently sampled six-stock Random panel, and a 24-stock Scope-Expansion panel added specifically to test these findings at a substantially broader sectoral and cross-sectional scale. Five forecasting models are evaluated under leakage-safe walk-forward validation: a naive random-walk benchmark, AR(1)/Ridge, Linear SVR, LightGBM, and an LSTM, in pooled and regime-as-feature specifications. Diebold-Mariano tests find no model signif-icantly outperforms the naive benchmark on any panel, and find LightGBM significantly worse than naive on all three panels independently; a nested ablation study and SHAP anal-ysis corroborate this. A regime-conditioned trading strategy, backtested under an itemized DSE cost structure and a liquidity constraint, loses the majority of capital on all three pan-els; a Deflated Sharpe Ratio, an out-of-sample validation split, and a stationary block boot-strap confirm no strategy demonstrates statistically defensible skill. The Hidden Markov Model also systematically misclassifies extended floor-price trading freezes as low-risk rather than as market stress. These results, replicated across three panels spanning 17 sec-tors and an alternative non-HMM regime detector, are consistent with weak-form market efficiency and caution against assuming regime-aware machine learning transfers cleanly to a frontier market with an active price-floor mechanism JEL Classification: C22; C53; C58; G11; G14; G17