xxcy12138/TSBench

TSDiag是一个合成时间序列基准数据集，专门设计用于诊断时间序列预测模型，而不仅仅是简单排名。每个生成器都有已知的贝叶斯最优预测器，使得预测误差可以分解为四个可识别组件：不可约噪声、架构瓶颈、数据不足和优化方差。数据集包含15个生成器，分为8个过程家族：F1线性自回归（AR1、ARFIMA，具有平稳线性依赖性）、F2随机微分方程（OU、GBM，具有连续时间动态）、F3长记忆（fBm，具有自相似和长程依赖性）、F4确定性混沌（LogisticMap、MackeyGlass，具有对初始条件的敏感性）、F5高斯过程（GP_RBF、GP_Matern、GP_Periodic，具有核定义的协方差）、F6非平稳（ChangePoint、RegimeSwitching，具有制度变化）、F7趋势+季节性（Fourier、Polynomial，具有确定性模式）和F8多变量因果（VAR，具有跨变量依赖性）。数据格式方面，每个配置包含训练集（10,000个样本，宽格式）和测试集（500个样本，宽格式），时间结构包括上下文（前192个时间步）和预测范围（后96个时间步），总长度为288。参数设置包括上下文长度192、预测范围96、训练样本数10,000、测试样本数500等。TSDiag is a synthetic time-series benchmark designed for diagnosing forecasting models beyond simple ranking. Each generator has a known Bayes-optimal predictor, enabling decomposition of forecasting error into four identifiable components: irreducible noise, architecture bottleneck, data insufficiency, and optimization variance. The dataset contains 15 generators across 8 process families: F1: Linear Autoregressive (AR1, ARFIMA, with stationary linear dependence), F2: Stochastic Differential Equation (OU, GBM, with continuous-time dynamics), F3: Long Memory (fBm, with self-similar, long-range dependence), F4: Deterministic Chaos (LogisticMap, MackeyGlass, with sensitivity to initial conditions), F5: Gaussian Process (GP_RBF, GP_Matern, GP_Periodic, with kernel-defined covariance), F6: Non-stationary (ChangePoint, RegimeSwitching, with regime changes), F7: Trend + Seasonality (Fourier, Polynomial, with deterministic patterns), and F8: Multivariate Causal (VAR, with cross-variable dependence). In terms of data format, each configuration includes a train set (10,000 samples in wide format) and a test set (500 samples in wide format), with a temporal structure of context (first 192 time steps) and horizon (last 96 time steps), totaling 288 steps. Key parameters include context length 192, forecast horizon 96, train samples per generator 10,000, test samples per generator 500, etc.