task_categories:
- time-series-forecasting
TSCOMP Corpus
The TSCOMP (Time-Series Component-level Benchmarking) Corpus is a curated collection of evaluation results from systematic component-level experiments in deep multivariate time-series forecasting.
Overview
The corpus contains metrics.npy files from over 20,000 experimental runs, each recording the performance of a specific model component configuration (e.g., normalization layers, attention mechanisms, patching strategies) across multiple forecasting benchmarks.
Purpose
This corpus enables researchers to train custom meta predictors — models that learn to predict the best component configuration for a given time-series dataset, without needing to run expensive ablation studies themselves.
Dataset Structure
Each subdirectory in the archive corresponds to a downstream dataset (e.g., ECL, ETTh1, Exchange, weather). Within each dataset folder, individual experiment directories encode the full component configuration in their names, and metrics.npy contains the evaluation metrics for that run.
Sample Usage
The corpus is designed to be used for meta-learning. You can use the provided code in the GitHub repository to run experiments or extract features:
# Run meta learning experiments
python meta/run.py --mode simple --test_dataset ETTh2 --meta_model_type mlp
# Extract meta-features for datasets
python meta/meta_features/get_meta_features_LTF.py --meta_feature_type tabpfn
# Apply meta selection to new datasets
python meta/run_custom.py --new_dataset my_dataset --checkpoint_path <path> --new_dataset_path <csv_path> --scripts_root <scripts_dir>
📝 Citation
If you find this work useful, please consider citing:
@inproceedings{
liang2026beyond,
title={Beyond Holistic Models: Systematic Component-level Benchmarking of Deep Multivariate Time-Series Forecasting},
author={Shuang Liang and Chaochuan Hou and Xu Yao and Shiping wang and Hailiang Huang and Songqiao Han and Minqi Jiang},
booktitle={KDD 2026 Datasets and Benchmarks Track (Cycle 2)},
year={2026}
}