A physically constrained wavelet-aided statistical model (PCWASM) is presented to analyse and predict monthly groundwater dynamics on multi-decadal or longer time scales. The approach retains the simplicity of regression modelling but is constrained by temporal scales of processes responsible for groundwater level variation, including aquifer recharge and pumping. The methodology integrates statistical correlations enhanced with wavelet analysis into established principles of groundwater hydraulics including convolution, superposition and the Cooper–Jacob solution. The systematic approach includes (1) identification of hydrologic trends and correlations using cross-correlation and multi-time scale wavelet analyses; (2) integrating temperature-based evapotranspiration and groundwater pumping stresses and (3) assessing model prediction performances using fixed-block k-fold cross-validation and split calibration-validation methods. The approach is applied at three hydrogeologicaly distinct sites in North Florida in the United States using over 40 years of monthly groundwater levels. The systematic approach identifies two patterns of cross-correlations between groundwater levels and historical rainfall, indicating low-frequency variabilities are critical for long-term predictions. The models performed well for predicting monthly groundwater levels from 7 to 22 years with less than 2.1 ft (0.7 m) errors. Further evaluation by the moving-block bootstrap regression indicates the PCWASM can be a reliable tool for long-term groundwater level predictions. This study provides a parsimonious approach to predict multi-decadal groundwater dynamics with the ability to discern impacts of pumping and climate change on aquifer levels. The PCWASM is computationally efficient and can be implemented using publicly available datasets. Thus, it should provide a versatile tool for managers and researchers for predicting multi-decadal monthly groundwater levels under changing climatic and pumping impacts over a long time period.