Beyond Isolation: Neighbor-Consistent Data Pruning for Multivariate Time Series Forecasting

Data pruning for multivariate time-series forecasting (MTSF) is an emerging topic that poses unique challenges due to the intricate temporal dependencies and high-dimensional correlations between variables. The dominant data pruning scores have been designed for i.i.d. image classification tasks and rate each data sample using its own training trajectory in isolation. However, this misses a structural property of MTSF training: samples (obtained by windowing time series) can easily overlap, especially if strides down to 1 timestep are used, thus breaking the standard training i.i.d. assumption. Indeed, our key observation is that this overlap drives neighbors' loss trajectories into lockstep across training; the few samples whose trajectories deviate from this lockstep carry gradient information their neighbors do not share—exactly the samples that are worth pruning away. Hence, we propose TWIN (Trajectory-Wise Inconsistency with Neighbors), a multi-scale Wasserstein-2 score and pruning algorithm that, for the first time, compares each sample's sorted training-loss trajectory against those of its sliding-window neighbors. Specifically, our design is motivated by two complementary bounds: the pairwise discrepancy between neighboring loss trajectories is at most $\mathcal{O}(\delta)$ when both samples are clean (Theorem 1), but at least $\Omega(\Delta^2)$ once one carries a $\Delta$-spike contamination (Theorem 2). Whenever $\Delta \gg \delta$, these scaling orders separate, exposing an asymptotic scaling-order gap that TWIN's rank-based rule reads without estimating any bound constant. A 16,000-cell synthetic audit empirically confirms this gap, with a median 44.23× separation between clean and contaminated populations. We further evaluate TWIN's effectiveness on six MTSF benchmarks, four horizons, and five seeds against five baseline methods, showing that it shortens the average retraining time to approximately 48% of the unpruned baseline and reduces the average Relative MSE by up to 1.56% under IQR self-calibration, thus setting a new state of the art.