Clarification on Data Scaling for Zero-Shot Time Series Predictions Using TimesFM

[mrlucasrib]

Hello,

I am working on time series predictions in a zero-shot setting and noticed that neither this model nor other time series foundation models seem to perform any scaling (e.g., standardization or normalization) on the dataset before running it through the model.

This raises a question:

Is scaling omitted because the models assume that users will handle it beforehand?
Or is scaling unnecessary for these models?
I would greatly appreciate any clarification or guidance regarding the appropriate preprocessing steps, particularly in the context of zero-shot predictions.

Additional Information:
I am using the GluonTS library with a class wrapper to run the model. For reference, here is the implementation I am working with:
TimesFM Notebook.

Thank you for your assistance!

[0trm]

Scaling is unnecessary: TimesFM normalises every input series inside the model and maps the forecast back to the original scale, so standardising beforehand changes nothing.

• TimesFM 2.5 normalises each series with running per-patch mean and std (RevIN), and the config documents the guarantee TimesFM(aX + b) = a * TimesFM(X) + b for a >= 0 (configs.py).
• TimesFM 1.0/2.0, which the gift-eval notebook wraps, normalises each series by the mean and std of its first patch in _forward_transform (pytorch_patched_decoder.py).

Checked on 2.5 (timesfm 3.0.2 from PyPI): forecasting x and 1000 * x + 5e4, then undoing the transform, gives the same forecast to within 3e-6.

p1, _ = model.forecast(horizon=24, inputs=[x])
p2, _ = model.forecast(horizon=24, inputs=[1000 * x + 5e4])
np.abs((p2[0] - 5e4) / 1000 - p1[0]).max()  # ~3e-6

Both versions also have an opt-in flag that normalises the whole series first: ForecastConfig(normalize_inputs=True) in 2.5, documented for inputs of extremely large or small magnitude, and forecast(..., normalize=True) in 1.0/2.0.