The rapid advancements in time-series forecasting have led to the emergence of datasets with increasingly diverse characteristics. Researchers typically focus on designing robust algorithms to handle these datasets. However, model performance can vary across different datasets. Existing studies commonly use Mean Absolute Error (MAE) and Mean Squared Error (MSE) to evaluate model performance. These metrics often need to fully account for the impact of dataset quality on forecast accuracy and reliability, leading to insufficient explanations of model forecasts and even dataset dependence. Furthermore, the lack of dataset evaluation makes it difficult to determine whether forecast results are influenced by data characteristics or model architecture, significantly hindering model improvement. To address this problem, this study proposes a new dataset evaluation paradigm-Multiscale Temporal Entropy Evaluation Paradigm. This paradigm aims to tackle the problem of data feature ambiguity affecting model forecast interpretability. By evaluating nine time-series datasets from domains such as weather, economics, transportation, and networks and comparing these evaluations with the forecast results from six different models, this study demonstrates the effectiveness of this method in explaining the relationship between dataset features and model performance.