This study develops a model to predict the weekly sales of an underwear shop in New Taipei City, Taiwan. Most commonly employed prediction methods failed when applied to the sales patterns of the case company and could not create an appropriate ordering strategy. This study is based on the assumption that the choice of a predictive model may need to be matched to the fluctuation structure of the data sequence. Not all data sequences are suitable for a single prediction model. An adaptive generalized Nonlinear Grey Bernoulli Model (NGBM(1,1)) is proposed for the case company. The training data sequences are fed into an Adaptive Resonance Theory 2 (ART2) model to generate typical templates. Simultaneously, the generalized NGBM(1,1) model is associated with the Real-valued Genetic Algorithm (RGA) to identify the best hyper-parameters for each training data sequence. The relationships between the typical templates and the 24 model types are mapped. In the testing stage, each data sequence is allocated to the most similar typical template using ART2 and mapped to the corresponding suitable model type, which makes a one-step-ahead prediction. The four-year weekly sales data of an underwear shop were empirically analyzed. The proposed method was compared with traditional prediction models. The experimental results show that the proposed method can provide more precise predictions.