Reliable prediction of population displacement flows is essential for transport planning, demand-responsive service design, and evidence-based urban policy. This paper presents a novel, explainable machine learning framework for short- and medium-term displacement prediction applied to official Spanish geolocation data (MITMA dataset, January 2022–March 2024, ~750 GB/year). The framework addresses three nested spatial scales simultaneously—interprovincial, intermunicipal, and interdistrict—and introduces two original contributions beyond state-of-the-art benchmarking: (i) a scale-adaptive model selection criterion grounded in the Coefficient of Variation (CV) of demand and spatial entropy, providing a theoretically justified and operationally replicable decision rule; and (ii) a SHAP-based feature importance analysis that reveals how the dominant drivers of mobility shift qualitatively across spatial scales, yielding insights of direct value to transport planners.Four supervised learning models are rigorously evaluated: Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Sequential Deep Neural Network (SDNN), and Linear Time-Series Regression (LTSR). XGBoost consistently achieves the best predictive accuracy across all three scales: %EMAE = 15.6% (interprovincial, R² = 0.881), 13.1% (intermunicipal, R² = 0.914), and 10.3% (interdistrict, R² = 0.933), outperforming the linear baseline by 8–14 percentage points. Statistical superiority over all competing models is confirmed via Diebold–Mariano tests (p < 0.05). SHAP analysis reveals a systematic scale-dependency of feature importance: public holiday indicators dominate at the interprovincial scale (φ = 0.31), weekday encoding dominates at the intermunicipal scale (φ = 0.28), and same-week lag features dominate at the interdistrict scale (φ = 0.34). A complete model is validated.