Deep time-series forecasting plays an integral role in numerous practical applications. However, existing research fall short by focusing narrowly on either neural architecture designs for long-term point forecasts or probabilistic models for short-term scenarios. By proposing a comprehensive framework, facilitated by a novel tool, ProbTS, that integrates diverse data scenarios, evaluation metrics, and methodological focuses, we aim to transcend the limitations of current forecasting practices. Rigorous experimentation uncovers pivotal insights, including the supreme importance of aligning forecasting methodologies with the unique characteristics of the data; the necessity of a broad spectrum of metrics for accurately assessing both point and distributional forecasts; and the challenges inherent in adapting existing forecasting methods to a wider range of scenarios. These findings not only challenge conventional approaches but also illuminate promising avenues for future research, suggesting a more nuanced and effective strategy for advancing the field of deep time-series forecasting.

翻译：暂无翻译