The rapid decline of groundwater levels (GWL) due to pervasive groundwater abstraction in the densely populated (~1 billion) Indus-Ganges-Brahmaputra-Meghna (IGBM) transboundary river basins of South Asia, necessitates a robust framework of prediction and understanding. While few localized studies exist, three-dimensional regional-scale characterization of GWL prediction is yet to be implemented. Here, `support vector machine', a machine-learning-based method, is applied to data from the Gravity Recovery and Climate Experiment (GRACE) and data on land-surface-model-based groundwater storage and meteorological variables, to predict the GWL anomaly (GWLA) in the IGBM. The study has three main objectives, (1) to understand the spatial (observation well locations) and subsurface (shallow vs. deep observation wells) variability in prediction results for in-situ GWLA data for a large number of observation wells (n = 4,791); (2) to determine its relationship with groundwater abstraction, and; (3) to outline the advantages and limitations of using GRACE data for predicting GWLAs. The findings, based on individual observation well results, suggest significant prediction efficiency (median statistics: r > 0.71, NSE > 0.70; p < 0.05) in most of the IGBM; however, the study identifies hotspots, mostly in the agriculture-intensive regions, having relatively poor model performance. Further analysis of the subsurface depth-wise prediction statistics reveals that the significant dominance of pumping in the deeper depths of the aquifer is linked to the relatively poor model performance for the deep observation wells (screen depth > 35 m) compared with the shallow observation wells (screen depth < 35 m), thus, highlighting the limitation of GRACE in representing spatial and depth-dependent local-scale pumping.