Engineered Cementitious Composites (ECC) exhibit exceptional strain-hardening behavior, yet predicting their tensile properties remains challenging due to complex nonlinear interactions among mix constituents and fiber characteristics. This study presents a comprehensive machine learning framework for predicting both tensile stress and tensile strain capacity of ECC. A diverse multi-source database comprising 655 stress and 746 strain measurements spanning eight fiber types, water-to-binder ratios of 0.16–0.47, and fiber volume fractions of 0.5–3.0% was compiled from three published datasets. A physics-informed Fiber Reinforcement Index (FRI = V_f × L_f/d_f) was included among 18 input features. Fifteen individual machine learning models and three stacking ensembles were trained with Optuna Bayesian optimization (200 trials, ten-fold cross-validation), and overfitting was monitored via the train–test R² gap. CatBoost achieved the highest accuracy for tensile stress (R² = 0.923, RMSE = 0.975 MPa), while XGBoost performed best for strain capacity (R² = 0.853, RMSE = 0.921%). Per-source test R² ranged from 0.76 to 0.93, confirming cross-source generalization. SHapley Additive Explanations (SHAP) analysis revealed a dual mechanistic pattern: water content and supplementary cementitious materials dominate stress (matrix-controlled), whereas fiber type, aspect ratio, and FRI govern strain capacity (fiber-controlled), providing quantitative confirmation of ECC micromechanics theory.