Ensemble methods combine predictions from multiple models to improve forecasting accuracy. This paper investigates the effectiveness of multi-output ensembles for multi-step time series forecasting problems. While dynamic ensembles have been extensively studied for one-step ahead forecasting, their application to multi-step forecasting remains largely unexplored, particularly regarding how combination rules should be applied across different forecasting horizons. We conducted comprehensive experiments using 3568 time series from diverse domains and an ensemble of 30 multi-output models to address this research gap. Our findings reveal that dynamic ensembles based on arbitrating and windowing techniques achieve the best performance according to average rank. Interestingly, we observed that most dynamic approaches struggle to outperform a simple static ensemble that assigns equal weights to all constituent models, especially as the forecasting horizon increases. The performance advantage of dynamic methods is more pronounced in short-term forecasting scenarios. The experiments are publicly available in a repository.