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.
