Mental health challenges among college students continue to rise, motivating scalable and proactive support. Passive smartphone sensing provides an unobtrusive way to capture behavioral signals related to mental well-being, enabling machine learning models that predict future symptom changes. However, most existing work targets mental health detection, while leaving mental health forecasting underexplored. Forecasting remains technically challenging, and we still lack a clear understanding of which modeling choices most effectively improve long-term forecasting accuracy in realistic deployments. In this paper, we present the first large-scale empirical study of college student mental health forecasting using the College Experience Study (CES), a multi-year longitudinal dataset with passive sensing and weekly surveys. We systematically evaluate three practical design dimensions: (1) single-user forecasting under privacy-restricted, data-scarce settings; (2) model granularity, comparing population-level generic models, similarity-based models, and personalized fine-tuned models; and (3) architecture choice, contrasting one-stage end-to-end forecasting with a two-stage decoupled pipeline. Across controlled comparisons, population-level generalization consistently delivers the highest forecasting performance, achieving up to 0.777 accuracy, while similarity-based transfer and fine-tuning provide limited gains. Two-stage pipelines often reduce accuracy due to objective mismatch between stages. These findings provide actionable baselines and guidance for deployable mental health forecasting.