Quantitative empirical models combined with historical data archives serve as one of the main tools to reconstruct the geospace magnetic field dynamics during space weather events. The essence of the approach is to develop a flexible mathematical framework, aimed to derive the magnetosphere's response to varying interplanetary conditions using multi-year pools of spacecraft observations. There exist two different ways to formulate an empirical model: (i) the so-called "modular" approach, in which the total field is associated with a few current systems having a priori defined geometry, implied by previous experimental and theoretical knowledge. Alternatively, (ii) the magnetic field can be expanded into series of formal basis functions with a minimum of predetermined assumptions. Both techniques have their own pluses and minuses. In the present paper, an efficient method is described to maximize the advantages and minimize the drawbacks by combining the modular part of the model field, on the one hand, and the basis function expansions, on the other, into a synthetic "hybrid" approximation. In such a setting, the modular part renders the main features of the magnetospheric structure, while the basis function expansions take on the role of a second-order correction that removes the remaining inaccuracies. Our primary goal here is to reconstruct the magnetosphere dynamics during individual events rather than develop a universal magnetic field description. For that purpose, the hybrid model is combined with the "nearest-neighbor" data mining in the space of magnetospheric state variables, which allows one to create sequential data subsets and reconstruct the magnetic field for different time moments during a specific storm or substorm. The importance of proper choice of the state variables is discussed, including the ground activity indices and the solar wind/IMF parameters.