What Does Augmentation Actually Add When Adapting a Time-Series Foundation Model?

Time-series foundation models (TSFMs) such as Chronos, TimesFM, Moirai, and Granite TTM start adaptation from a strong pretrained forecasting prior rather than from scratch. If only a short target series is available, augmenting that series seems like an obvious way to create more adaptation data. But when performance improves, what did the synthetic windows actually contribute? A standard “real vs. augmented” comparison cannot answer this because it changes three things at once: the content of the target data, the weight of that target data relative to the pretrained prior, and, when epochs are fixed, the number of optimiser updates. We introduce two controls that are usually missing: COPY, which repeats the real windows to isolate target reweighting, and MATCH, which replaces half of the real block with synthetic windows while holding total target weight fixed. Thus, MATCH − REAL measures the content-only effect, while COPY − REAL measures the effect of giving the target data more weight. Across an 18-panel controlled adaptation testbed, COPY accounts for 93% of the apparent augmentation gain for the two largest random-feature heads, while the contribution of weight-matched synthetic content is only 0.06–0.07%, with confidence intervals that include zero. The same pattern appears in gradient-trained output heads and rank-4 low-rank updates, where COPY recovers 95–107% of the apparent gain simply by creating more optimiser steps. A sweep over anchor strength explains why: copying helps only when the model is being held too tightly to its pretrained state. Tuning that anchor using real data alone improves performance by 2.6–3.2%, more than the entire apparent benefit of augmentation. We find clear evidence for useful synthetic content only in the genuinely scarce-data regime.