This study evaluated the short- and long-term prediction performances of a transformer-based trajectory-forecasting model for urban intersections. While a previous study focused on developing the basic structure of a transformer model for future trajectory prediction, the present study aimed to determine a practical prediction sequence length. To this end, multiple transformer models were trained with output sequence lengths ranging from 1 s to 10 s, and their performances were compared. The trajectory data used for training were generated through a microscopic traffic simulation, and the model accuracy was assessed using the metrics average displacement error (ADE) and final displacement error (FDE). The results demonstrate that the prediction accuracy decreases significantly when the output trajectory length exceeds 3 s. Specifically, straight-driving trajectories exhibit rapidly increasing errors, while turning trajectories maintained a relatively stable accuracy. In contrast, for turning-driving trajectories, prediction errors increased sharply during short-term forecasting, but the increase was more gradual in long-term forecasts. Additionally, the long-term prediction models produced higher errors even in the initial 1-second outputs, implying a tendency toward conservative inference under uncertain future scenarios. This conservative behavior is likely influenced by the model’s effort to minimize the overall loss across a broader prediction window, especially when trained with Smooth L1 loss function. This study provides practical insights into model design for edge-computing environments and contributes to the development of reliable short-term trajectory prediction systems for urban ITS applications.