Datasets:
Formats:
json
Languages:
English
Size:
< 1K
Tags:
time-series
time-series-decomposition
benchmark
component-recovery
symbolic-regression
icml-2026
License:
| # Semi-Synthetic Transfer Track | |
| This directory contains benchmark-only 22-method semi-synthetic transfer checks under real low-frequency backgrounds. These files support the living benchmark and are not camera-ready paper tables. | |
| Large downloaded source files are excluded from this Hugging Face release. The included files are: | |
| - canonical small background CSV files under `data/track_c/`; | |
| - method defaults and benchmark manifest under `configs/`; | |
| - merged and summarized metric outputs under `results/`. | |
| The extension ranking source is `results/summary/ranking_paper_5metric_overall.csv`, which ranks methods by mean within-setting rank over the five paper recovery metrics. | |
| The headline interpretation is that the mechanism-level picture transfers, while method-level conclusions become more conservative. In the aligned 22-method track, STL ranks first, followed by AMD block, MA baseline, Times2D block, and a tie among Autoformer, DLinear, and the moving-average decomposition block; CEEMDAN ranks sixth. CEEMDAN preserves strong seasonal agreement under real backgrounds, but its trend recovery remains sensitive to mode mixing against low-frequency real structure. | |