Forecasting foundation models are increasingly adopted as frozen feature extractors for time series classification, yet it remains unclear whether their pretrained representations convey information beyond what strong classifiers without pretraining can capture, especially when labeled data are scarce. We conduct a controlled benchmark on six large MONSTER datasets using two forecasters, Reverso and Chronos‑Bolt, with 1%, 10% and 100% of the training labels. Their frozen representations are compared against four established time series classification methods and random‑weight versions of the same forecasters. Pretraining improves accuracy over random weights on five datasets, with gains up to 23 percentage points, but the frozen features never reliably outperform the strongest non‑pretrained baseline: they are worse in 15 of 17 settings and tie in the remaining two. Additional experiments show that restoring amplitude information recovers part of the gap, while fine‑tuning and adding forecast‑error features do not close it. These findings highlight the necessity of strong task‑specific baselines and random‑weight controls when evaluating forecasting foundation models for classification. Code is available at https://github.com/Rachet2006/frozen-forecasters-monster
Review: While frozen forecasters can leverage pretraining to beat random‑weight counterparts, they still fall short of dedicated non‑pretrained classifiers in realistic settings, underscoring the importance of robust baselines and the potential benefit of preserving amplitude information.