SynTSBench
收藏资源简介:
SynTSBench是一个用于时间序列预测模型评估的合成数据驱动评估框架。该框架旨在通过可编程的特征配置,系统地评估时间序列预测模型的基本建模能力。它通过构建包含趋势、季节性、短期/长期依赖关系、多变量相关性等四个关键时间特征的合成数据集,以及基于传统随机时间序列和信号处理模型的复杂数据集,来模拟现实世界场景。这使得模型能够系统地评估学习特定模式类型的能力,同时揭示不同架构方法之间的比较优势。此外,该框架还通过在干净的合成信号中注入渐进噪声和不同类型的异常来实现全面的弹性评估,并利用合成数据生成来为每种模式类型建立理论性能边界(最优解),从而能够将模型预测与数学最优解直接进行比较,为分析性能差距和改进潜力提供参考点。
SynTSBench is a synthetic data-driven evaluation framework for assessing time series forecasting models. This framework aims to systematically evaluate the core modeling capabilities of such models via programmable feature configuration. It constructs synthetic datasets integrating four key temporal features—trend, seasonality, short-term/long-term dependencies, and multivariate correlations—alongside complex datasets derived from traditional stochastic time series and signal processing models, to emulate real-world scenarios. This approach enables systematic assessment of models' capacity to learn specific pattern categories, while revealing the comparative strengths of different architectural methodologies. Furthermore, the framework enables comprehensive robustness evaluation by injecting progressive noise and diverse anomalies into clean synthetic signals. It also leverages synthetic data generation to establish theoretical performance bounds (optimal solutions) for each pattern category, allowing direct comparison between model predictions and the mathematically optimal solutions, thus providing reference points for analyzing performance gaps and potential improvement areas.

- 1通过清华大学深圳国际研究生院 · 2025年



