Random Initialization as a Baseline for Time Series Foundation Models

Time series foundation models rely on self-supervised objectives, such as masked reconstruction or contrastive learning, to build reusable representations for downstream tasks such as multivariate time series classification (MTSC). In this work, we test how much pretrained weights improve over a baseline that requires no pretraining: the same architecture with randomly initialized weights. We evaluate four published time series foundation models across all publicly released scales, yielding seven model instances, on the UEA MTSC archive, with per-model exclusions where evaluation cannot be completed due to computational constraints. For each instance, we compare the pretrained weights against five randomly initialized variants using four downstream classifiers on frozen representations: two linear classifiers (ridge classifier and logistic regression) and two nonlinear classifiers (RBF SVM and random forest), yielding 183 paired dataset-level comparisons. Pretrained variants achieve higher mean accuracy across all seven instances, but the mean advantage is modest, ranging from 1.43 to 5.19 percentage points, with an overall average gain of 2.58 percentage points. Under per-instance Wilcoxon signed-rank tests, three of seven raw $p$-values fall below 0.05, but after Holm correction for multiple comparisons, only NuTime remains statistically significant. A secondary pooled analysis across all 183 pairs yields $p < 0.001$, while the corrected per-instance tests remain the primary statistical evidence. We further compare against three established time-series-specific classifiers namely, ROCKET, Hydra, and Quant which remain competitive with the pretrained foundation models without requiring large-scale pretraining. Together, these results motivate a more careful assessment of the computational cost of large-scale pretraining relative to its downstream benefit for MTSC and support the systematic inclusion of architecture-matched random initialization and non-pretrained baselines in future evaluations. Code, experiment configurations, and complete results are available in a repository. https://anonymous.4open.science/r/tsfm-random-init-baselines-65A7