Accurately predicting pile load–displacement behavior is critical for optimizing deep foundation design and ensuring structural safety. Traditional empirical and numerical methods often fall short in capturing the complex, nonlinear interactions among pile geometry, soil resistance, overburden stress, and applied load under varied site conditions. To address these limitations, this study introduces a hybrid machine learning framework that integrates Moth–Flame Optimization and Tree-structured Parzen Estimator for hyperparameter tuning of the eXtreme Gradient Boosting algorithm. The model is trained on a comprehensive database of 2828 static load test results from real-world projects in Vietnam, including bored piles (1650 samples) and prestressed high-strength concrete driven piles (1178 samples). Pile type is explicitly considered to account for differences in load-bearing resistance mobilization. The proposed model demonstrates superior predictive performance (R² = 0.979, root mean squared error = 5.043 mm, mean absolute error = 2.860 mm). Sensitivity analysis reveals several geotechnical thresholds associated with variations in pile settlement. Partial dependence plots indicate that the standard penetration resistance below the pile toe consistently influences displacement, where values exceeding 35 blows per 0.3 m are associated with significantly reduced settlement. In contrast, piles embedded deeper than 60 m in weak soils—with penetration resistance below 35—may exhibit increased displacement due to limited side friction and inadequate toe mobilization. Overburden stress at greater depths also shows a more consistent contribution to settlement reduction than shallow confinement, underscoring the importance of deep-layer properties in load transfer. These findings provide actionable insights for serviceability-based design, enabling more efficient and reliable foundation solutions. The model’s robustness is further validated through pile-type-specific evaluation and Monte Carlo-based uncertainty propagation. Overall, the proposed framework offers a scalable, interpretable, and cost-effective solution for advancing data-driven pile foundation engineering.