Macroelement models characterise the macroscopic behaviour of foundations in terms of resultant forces and displacements at a reference point within the foundation. Conventional macroelement models are governed by constitutive equations whose parameters are calibrated against experimental or numerical results; however, the models’ response is not necessarily accurate for varying conditions when adopting a unique set of parameters. Data-driven approaches can be more flexible as they are not bounded by any mathematical formulation and so can fit the underlying foundation behaviour with greater accuracy. A new data-driven macroelement (DdM) model for suction buckets is developed in this paper. Firstly, a database of 3D finite element (FE) analyses was created to train the DdM model. The 3D FE analyses considered suction bucket foundations installed in sand. A state-parameter dependent constitutive model incorporating a nonlinear elastic overlay model was adopted with parameters calibrated for the Dunkirk PISA site. The database includes variations in the foundations embedment-to-diameter ratio, initial relative density of the sand and the applied loading direction. An artificial neural network (ANN) was then employed as the underlying DdM model. The ANN was trained to predict the foundations force components (vertical, horizontal and bending) given the current deformation state of the foundation. The performance of the ANN was extensively validated, demonstrating an excellent generalisation ability to unseen data. Overall, the proposed DdM model demonstrates the feasibility and potential of this class of models for future design of suction buckets.