Forecasting intermittent and lumpy demand is a significant supply chain challenge, as traditional metrics like MAPE and RMSE are unreliable for such sparse data. This study provides a comprehensive comparative evaluation of forecasting methods, benchmarking statistical, machine learning, and deep learning models across 3,671 intermittent time series. We utilized robust, cost-sensitive metrics like SPEC, alongside MAAPE and MWQL, for a more accurate assessment. Our results demonstrate that deep learning models significantly outperform traditional and machine learning approaches. DeepAR achieved the best probabilistic forecasting, making it ideal for managing uncertainty. In contrast, the Deep Renewal Hybrid model excelled in point accuracy. Statistical methods remained competitive in cost-sensitive scenarios, while machine learning models underperformed. This research establishes a practical decision framework to guide practitioners in selecting the optimal model based on specific operational priorities, whether they are focused on uncertainty, accuracy, or cost. This provides actionable, evidence-based recommendations for real-world supply chain implementation.