Crowd flow prediction plays a vital role in supporting public services, such as traffic management, urban planning, and emergency response. However, achieving reliable predictions requires integrating multiple data sources, including historical crowd flow, regional traffic, weather conditions, and points of interest, which presents significant challenges. Crowd data inherently exhibits strong spatiotemporal dependencies, as adjacent crowd points are highly correlated and influenced by diverse external factors. Additionally, the distributed storage of crowd flow and external factors data, coupled with privacy protection requirements, makes centralized training prone to data sharing barriers and privacy leakage risks. To address these challenges, we propose FedVCP, a crowd flow prediction framework based on vertical federated learning (VFL), which enables the construction of efficient prediction models while preserving data privacy. Specifically, FedVCP captures the spatiotemporal dependencies of crowd flow data and integrates external factors simultaneously. By leveraging VFL, the model is jointly trained on distributed data sources, effectively addressing data heterogeneity and autocorrelation while ensuring privacy preservation. Experimental results demonstrate that FedVCP maintains data privacy while significantly improving model performance, reducing RMSE by up to 35.98% compared to the baseline, and exhibits strong stability and generalization ability.