Post
72
In my ICML 2026 paper I treat time-series decomposition as its own task instead of a cleanup step: can a method recover the true trend, cycles and noise?
Core setup: 6 scenarios ร 50 deterministic draws ร 6 method families, five recovery metrics. I've put the leaderboard and the data on the Hub:
huggingface.co/spaces/Zipeng365/TSDecompose-Benchmark-Leaderboard
huggingface.co/datasets/Zipeng365/TSDecompose-Benchmark
Core setup: 6 scenarios ร 50 deterministic draws ร 6 method families, five recovery metrics. I've put the leaderboard and the data on the Hub:
huggingface.co/spaces/Zipeng365/TSDecompose-Benchmark-Leaderboard
huggingface.co/datasets/Zipeng365/TSDecompose-Benchmark