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

[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:
Chronos Notebook.

Thank you for your assistance!

[lostella]

The models do scale the input data (each series independently of the others):

• scaling in chronos-t5

  chronos-forecasting/src/chronos/chronos.py

  Line 187 in ad410c9

  scaled_context = context / scale.unsqueeze(dim=-1)

• scaling in chronos-bolt:

  chronos-forecasting/src/chronos/chronos_bolt.py

  Line 95 in ad410c9

  return (x - loc) / scale, (loc, scale)

[mrlucasrib]

Thank you for your reply!

Just to clarify, I see now that the model performs scaling for context length, but does this also help manage datasets with high variance? Could leaving the data unscaled negatively affect the model's performance in such cases?

Is there any situation where scaling the entire dataset at once is necessary, or is this approach optional depending on the model's design? My master's advisor strongly insists on scaling the entire time series as a preprocessing step, and I want to confirm if this is critical for this specific foundation model.

[lostella]

What do you mean by “scaling the entire dataset at once”?

[mrlucasrib]

What do you mean by “scaling the entire dataset at once”?

When I mention "scaling the entire dataset at once," I'm referring to the process of applying a scaler, such as MinMaxScaler or StandardScaler from scikit-learn, to the entire time series dataset. For instance, if I have a time series representing the number of disease cases over several months, and one month has a significantly higher number of cases (e.g., 2880 cases) while other months have much smaller numbers (e.g., 5 or 7 cases), I can scale the data so that all the values are between 0 and 1 (with MinMaxScaler) or have a mean of 0 and a standard deviation of 1 (with StandardScaler).

[lostella]

if I have a time series representing the number of disease cases over several months, and one month has a significantly higher number of cases (e.g., 2880 cases) while other months have much smaller numbers (e.g., 5 or 7 cases), I can scale the data so that all the values are between 0 and 1 (with MinMaxScaler) or have a mean of 0 and a standard deviation of 1 (with StandardScaler).

Something like this is done already by the models:

• chronos-t5 applies a scaling that makes the data have mean absolute value equal to 1
• chronos-bolt applies an affine transformation that makes the data have mean 0 and stddev 1

Hence, any affine transformation you apply to the data prior to invoking the model (of the form lambda data: a*data + b with a > 0) will effectively be undone in the case of chronos-bolt (and also in the case of chronos-t5 if b == 0)