TSDecompose Benchmark Leaderboard
Interactive leaderboard for Time-Series Decomposition as a Standalone Task: A Mechanism-Identifiable Benchmark. The dataset repository is the source of record; this Space is the web presentation layer.
The classic paper benchmark uses 5 core component-recovery metrics. Each metric is computed per generated draw after output alignment, then averaged by scenario, tier, or regime.
Table 2 displays only 2 of those metrics: Trend R2 and Seasonal spectral correlation, split over stationary regimes (Tiers 1-2) and the non-stationary regime (Tier 3). Its stationary columns use a tier-balanced average of Tier 1 and Tier 2 means over valid metric values; non-stationary columns use Tier 3 means. Figure 3 is the five-metric profile view.
1 - sum_t (T_t - T_hat_t)^2 / sum_t (T_t - mean(T))^2
Scale-sensitive fit of the recovered trend against the ground-truth trend.
DTW(T, T_hat)
Shape-sensitive trend distance under monotone time alignment.
1 - sum_t (S_t - S_hat_t)^2 / sum_t (S_t - mean(S))^2
Scale-sensitive fit of the recovered seasonal component.
corr(P_S(f), P_S_hat(f))
Pearson correlation between normalized power spectra, measuring frequency-content recovery.
max_{lag in L} corr(S_aligned(lag), S_hat_aligned(lag))
Phase-robust seasonal similarity over a fixed lag search window.