<p>Accurate forecasting of agricultural commodity prices is essential for production planning, market regulation, and supply chain decision-making. However, most existing forecasting models formulate agricultural price prediction as a point-wise regression task and mainly focus on reducing numerical error at each future step, while paying limited attention to the global evolution pattern of the predicted sequence. To address this limitation, this study proposes a Trajectory Shape Modeling Network (TSMNet), a shape-aware forecasting framework that jointly models future numerical values and trajectory evolution patterns. The proposed method introduces a dual-branch architecture consisting of a numerical forecasting branch and a trajectory shape learning branch, together with a shape-guided refinement mechanism that injects global structural information into value prediction. In addition, a joint supervision strategy is designed to enforce consistency between the predicted sequence and the target trajectory in terms of trend evolution, turning behavior, and fluctuation pattern. Experiments on two real-world potato price datasets from the World Food Programme, covering Myanmar and the Philippines in Southeast Asia, demonstrate that TSMNet consistently outperforms representative forecasting baselines, including CNN, LSTM, N-BEATS, Autoformer, and Informer, in terms of MAE and RMSE. These results suggest that incorporating trajectory shape as an additional learning target provides a more informative and structurally faithful forecasting paradigm for agricultural price prediction.</p>