Offline data-driven evolutionary algorithms (DDEAs) can utilize historical data to build surrogate models that guide population optimization. It overcomes the limitations of traditional evolutionary algorithms, when faced with expensive and computationally intensive problems or problems with complex mechanisms that are difficult to model mathematically. Therefore, it has attracted the attention of many scholars. However, DDEAs face two challenges: first, constructing high-quality surrogate models requires sophisticated model management strategies. While these strategies enhance algorithmic performance, they increase computational time. Second, while radial basis function networks (RBFNs) are widely used in DDEAs, selecting appropriate hyperparameters for different problems remains challenging. To address these issues, this paper first proposes a preselection strategy that uses low-complexity coarse models to quickly iterate the population towards the vicinity of the optimal solution. Then, a model ranking confidence indicator based on the Kendall's rank correlation coefficient is introduced, along with a selection strategy designed to identify the most suitable hyperparameters for the current problem from several RBFN hyperparameter configurations. Based on these two components and combined with the ensemble method of stacked generalization, an offline data-driven evolutionary algorithm based on presearch and model selection (DDEA-PMS) is proposed. Experimental results on 5 benchmark problems, compared with six state-of-the-art DDEAs, demonstrate that the proposed DDEA-PMS can achieve significantly better results with extremely low computational overhead.