This paper develops a predictive framework for customer churn in the Moroccan telecommunications sector using a hybrid ensemble of XGBoost, feedforward neural networks, and LSTMs. Unlike conventional churn models, which often rely on static features and uniform sampling, our approach integrates domain expertise at each stage of the pipeline. The dataset spans 22, 238 subscribers (2021–2023) and includes 43 engineered predictors distilled from 127 raw variables, enriched by expert validation. The model achieves 89.7% accuracy and 0.943 ROC-AUC, with network quality (34% contribution) emerging as the primary driver of churn, far ahead of pricing factors (11.3%). Contributions include (i) explicit labeling and leakage control, (ii) extended evaluation metrics (PR-AUC, Brier, calibration, lift), (iii) interpretability via SHAP (global, local, and segment-level), (iv) fairness diagnostics (TPR/FPR gaps across sub-populations), (v) robustness analysis of the 75% advertised-speed threshold, and (vi) validation on a held-out Q4-2023 split. Results demonstrate that in emerging markets, operational quality improvements outweigh pricing adjustments.