Ground motions observed at the ground surface are strongly influenced by near-surface geological conditions. Vertical array observations provide direct information on the dynamic relationship between motions at the ground surface and at the engineering bedrock. Recently, Shioi et al. [1] proposed a data-driven seismic response modeling framework based on Dynamic Mode Decomposition with control (DMDc) [2] combined with time-delay embedding. However, its applicability to large-scale real vertical array datasets and its robustness across different site conditions and ground motion intensities have not yet been systematically examined. This study evaluates the applicability of the DMDc framework using approximately 2,400 vertical array records obtained at four observation sites, where horizontal components at the ground surface and at the engineering bedrock were simultaneously recorded. The records span a wide range of earthquake events with predominantly low to intermediate amplitudes. The DMDc model is constructed following the methodology of Shioi et al., and its predictive performance is assessed through cross-validation. The dependence of prediction accuracy on site conditions and ground motion intensity measures, including peak ground acceleration, is examined. The results are compared with conventional one-dimensional linear transfer functions based on multiple reflection theory. Preliminary screening of the dataset indicates clear site-dependent differences in frequency characteristics between surface and bedrock motions. Pilot analyses suggest that the DMDc-based model provides stable waveform reconstruction for low to moderate ground motion levels, while deviations from linear transfer function predictions become more pronounced with increasing motion intensity. These observations imply potential advantages of data-driven operator representations in capturing complex site effects. This study provides a systematic assessment of DMDc applied to real vertical array ground motion records and contributes to the development of data-driven, uncertainty-aware approaches for site response analysis.