xxcy12138/TSBench

TSDiag: A Decomposable Synthetic Benchmark for Diagnosing Time-Series Forecasting Models

Dataset Description

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.

Dataset Structure

The dataset contains 15 generators across 8 process families:

Family	Generators	Key Property
F1: Linear Autoregressive	AR1, ARFIMA	Stationary linear dependence
F2: Stochastic Differential Equation	OU, GBM	Continuous-time dynamics
F3: Long Memory	fBm	Self-similar, long-range dependence
F4: Deterministic Chaos	LogisticMap, MackeyGlass	Sensitivity to initial conditions
F5: Gaussian Process	GP_RBF, GP_Matern, GP_Periodic	Kernel-defined covariance
F6: Non-stationary	ChangePoint, RegimeSwitching	Regime changes
F7: Trend + Seasonality	Fourier, Polynomial	Deterministic patterns
F8: Multivariate Causal	VAR	Cross-variable dependence

Data Format

Each configuration (generator) contains:

• train.parquet: 10,000 samples in wide format
• test.parquet: 500 samples in wide format

Column schema (univariate):

sample_id, t_0, t_1, ..., t_287

Column schema (multivariate, e.g., F8_VAR with 3 channels):

sample_id, t_0_c0, t_0_c1, t_0_c2, t_1_c0, ..., t_287_c2

Temporal structure:

• Context (input): first 192 time steps (t_0 to t_191)
• Horizon (target): last 96 time steps (t_192 to t_287)

Configuration

Parameter	Value
Context length	192
Forecast horizon	96
Total length	288
Train samples per generator	10,000
Test samples per generator	500
Train seed	42
Test seed	9999

Usage

from datasets import load_dataset

# Load a specific generator
ds = load_dataset("YOUR_USERNAME/tsdiag", "F1_AR1")

# Access train/test splits
train_df = ds["train"].to_pandas()
test_df = ds["test"].to_pandas()

# Extract context and horizon
context_cols = [f"t_{i}" for i in range(192)]
horizon_cols = [f"t_{i}" for i in range(192, 288)]

X_context = train_df[context_cols].values  # (10000, 192)
Y_horizon = train_df[horizon_cols].values  # (10000, 96)

Citation

@inproceedings{tsdiag2026,
  title={TSDiag: A Decomposable Synthetic Benchmark for Diagnosing Time-Series Forecasting Models},
  author={Anonymous},
  booktitle={NeurIPS 2026 Datasets and Benchmarks Track},
  year={2026}
}

License

MIT