Crowd management plays a vital role in urban planning and emergency response. Accurate crowd prediction is important for venue operators to respond effectively to adverse crowd dynamics during large gatherings. Although many studies have tried to predict crowd densities or movement dynamics with data-driven predictive models, their validation is often limited to data within the same scenario. As a result, the predictability of the data-driven model in unseen scenarios, such as evacuation scenarios, remains unknown due to the challenges of collecting out-of-distribution data regarding emergency conditions. To address this problem, we present an evaluation pipeline to evaluate different kinds of data-driven models. A method is proposed to generate realistic scenarios by simulation and collect synthetic data from these scenarios to acquire a comprehensive dataset. With these synthetic data, we evaluated different predictive models, from traditional machine learning methods to deep learning time-series prediction models, to explore their generalizability. Furthermore, we propose a weighted average metric, which is better suited to determine the performance of forecasting algorithms under adverse conditions. Through extensive experimentation, we showcase the heterogeneity and diversity of the simulation dataset. The evaluation results also revealed that all the data-driven models performed poorly in unseen scenarios, highlighting the urgent need to develop a robust and generalizable model for predicting crowd flow in indoor spaces.