Mobility as a Service (MaaS) transforms urban transportation from car ownership to subscription-based models. A key factor for the success of MaaS is accurately predicting users’ Willingness to Pay (WTP) for various subscription packages, enhancing their adoption and satisfaction. This paper employs a “smart predict-then-optimize” framework, where the weekly, annual, and monthly MaaS subscription models are formulated as online, offline, and hybrid online-offline mobility resource allocation problems, respectively. We develop a group method of data handling (GMDH)-driven self-adaptive harmony search (SAHS) algorithm to solve the proposed mobility resource allocation problems effectively. Initially, GMDH-type neural networks predict users’ WTP using their historical travel data, such as travel distance and service time, and socio-demographic characteristics, including inconvenience tolerance and travel delay budget; then these predicted WTP values are fed into the weekly, annual, and monthly mobility resource allocation problems, respectively. Comprehensive numerical experiments based on a simulated dataset demonstrate the robust prediction performance of the GMDH neural network across weekly, monthly, and annual subscription models, as well as the effectiveness of the GMDH-driven SAHS algorithm in managing resource allocation for these models. Our numerical findings highlight that the monthly subscription model strikes an optimal balance, combining the flexibility of the weekly model with the strategic depth of the annual model. This study proposes three distinct MaaS subscription models and a data-driven metaheuristic algorithm to customize MaaS offerings to user needs.