Machine Learning Failures in Broker-Driven Emerging Markets: Concept Drift, Regime Switching, and Structural Breaks — Evidence from NEPSE Transaction-Level Data

Machine learning models have demonstrated strong predictive performance on developed market data, yet theirapplicability to broker-driven emerging markets remains poorly understood. We evaluate 12 models across five architecturalfamilies — including XGBoost, LightGBM, CatBoost, LSTM, GRU, and Transformer — on 57,285,204 transactions from theNepal Stock Exchange (NEPSE) spanning 902 trading days (2022–2025), using 4-fold expanding walk-forward validationwith 20 features comprising 13 technical indicators and seven novel broker microstructure features, including Order FlowImbalance and broker concentration indices. No model consistently outperforms the Naive Random Walk baseline, whichachieves the highest 4-fold average Sharpe ratio (0.511) and outperforms XGBoost (0.351) and all deep learning architectures,with LSTM producing negative average Sharpe (−0.003). Models show meaningful performance in structurally stable regimes —XGBoost achieves Sharpe 0.917 in the zero-break first fold — but universally fail when test windows contain structuralbreaks, producing negative Sharpe in the three-break thirdfold. Directional accuracy remains deceptively stable at 57–59%in structurally stable folds, confirming that accuracy alone is insufficient in fat-tailed markets, and no model survives NEPSE’sactual transaction costs (∼0.415% per side), with break-even costs not exceeding 0.032% for any model. Using a 3-stateGaussian Hidden Markov Model and PELT structural break detection, we identify five break dates and demonstrate that breakcount within the test window directly predicts failure severity, providing a measurable mechanism of concept drift, while brokermicrostructure features offer partial robustness in bear regimes. We propose a regime-gated deployment framework for machine learning practitioners in broker-driven emerging markets.