Time series foundation models can be few-shot learners (research.google)

Google Research announced TimesFM-ICF, a simple yet powerful way to turn a time-series foundation model into a few-shot learner by continued pre‑training with in‑context examples. Starting from TimesFM (a patched decoder that tokenizes 32-point patches, runs them through a stacked transformer and an MLP to emit 128-point outputs), the team inserts a learnable “common separator” token between each example and continues decoder‑only next‑token prediction training. The separator tokens prevent the model from conflating multiple example histories, and causal self‑attention (CSA) + feed‑forward layers learn to attend to, and generalize from, a handful of relevant series at inference time. On a 23‑dataset benchmark of unseen time series, TimesFM-ICF improves accuracy by 6.8% over the base TimesFM and matches the performance of supervised fine‑tuning (TimesFM‑FT) without dataset‑specific training. The method scales predictably—more in‑context examples improve forecasts at the cost of longer inference—and better leverages curated examples than a plain long‑context baseline. Practically, this enables businesses to deploy one adaptable forecasting model and specialize it on the fly by supplying a few relevant examples, cutting costs and lowering the barrier to state‑of‑the‑art forecasting. Future work targets automated selection of the most useful in‑context examples.