Short-horizon forecasting of antenna command-following errors is critical for maintaining the pointing accuracy of large-aperture antennas. However, existing multivariate forecasting architectures often either process variables independently or apply unrestricted mixing, leaving it unclear whether explicit information exchange between azimuth and elevation axes genuinely improves prediction. To address this problem, this studyevaluates whether explicit cross-axis interaction improves forecasting when only historical two-axis error sequences are used. We propose a compact End-DYN (end-positioned dynamics) architecture that combines a shared temporal encoder, a cross-axis refinement module for information exchange at aligned latent time steps, and aterminal direct history-to-future mapping. Operating on a 60-s historical error sequence sampled at 9Hz, the model forecastsfull 5-s or 10-s future trajectories. To isolate the effect of cross-axis refinement, we compare Coupled End-DYN with the parameter-matched Independent End-DYN variant over 15 eligible recording days. The results show that at 5s, Coupled End-DYNimproves the axis-balanced RMSE skill over the independent variant by 2.07±1.73 percentage points. At 10s, the mean gain decreases to 0.01±0.79 percentage points.  Thus, a benefit is observed at 5~s, whereas the balanced benefit becomes negligible at 10~s, suggesting that the effect of cross-axis refinement depends on the forecast horizon.

\begin{IEEEkeywords} large-aperture antenna, command-following error, multivariate time-series forecasting, cross-axis interaction, temporal mapping \end{IEEEkeywords}