TIMERECIPE
收藏资源简介:
TIMERECIPE是一个统一的时间序列预测基准框架,旨在评估时间序列预测方法在模块级别的有效性。该框架在广泛的预测场景和数据集上进行了超过10,000次实验,以评估个体组件的有效性。TIMERECIPE揭示了模块设计空间的详尽探索可以产生优于现有方法的预测模型,并揭示了模块有效性与特定时间序列数据和预测任务特征之间的有意义的相关性。此外,TIMERECIPE还提供了一个实用的工具包,根据这些实证见解推荐合适的模型架构。
TIMERECIPE is a unified benchmark framework for time series forecasting, which aims to evaluate the effectiveness of time series forecasting methods at the module level. This framework conducts over 10,000 experiments across a wide range of forecasting scenarios and datasets to assess the effectiveness of individual components. TIMERECIPE reveals that exhaustive exploration of the module design space can yield forecasting models that outperform existing methods, and uncovers meaningful correlations between module effectiveness and the characteristics of specific time series data and forecasting tasks. Furthermore, TIMERECIPE also provides a practical toolkit that recommends suitable model architectures based on these empirical insights.

- 1TimeRecipe: A Time-Series Forecasting Recipe via Benchmarking Module Level Effectiveness佐治亚理工学院 · 2025年



