Question about Dataset Separation in Pretraining for Time Series Forecasting

[mirmuss]

Hi there! 👋
First off, I want to say I'm really excited about your work on time series forecasting - the results look promising! While exploring the forecasting examples, I noticed something interesting that might be worth discussing.
I came across these forecasting visualizations that show some impressive predictions, particularly how well the model captures the periodic patterns. Specifically:

test_data, metadata = get_gluonts_test_dataset(
    "ETTh1", prediction_length=192, regenerate=False
)

Salesforce/moirai-moe-1.0-R-small model gives:

However, this made me wonder about the dataset separation between pretraining and evaluation phases.
Looking through the pretraining configuration, I couldn't find any explicit mention of holding out datasets from LOTSA (or similar benchmark datasets) during the pretraining phase. This raises an interesting question about how we're evaluating the model's true generalization capabilities.
Why this matters:

1. If benchmark datasets are included in pretraining, the model's performance on downstream tasks might not reflect its true zero/few-shot learning abilities
2. The periodic patterns we're seeing in the forecasts might be partially attributed to the model having seen similar patterns during pretraining
3. This could affect how we interpret the model's transfer learning capabilities

Possible Enhancement:
Would it make sense to:

1. Create a clear list of held-out datasets that are never seen during pretraining?
2. Document which benchmark datasets are excluded from pretraining?
3. Maybe add a configuration option to explicitly specify which datasets should be reserved for evaluation?

I'd love to hear your thoughts on this! I might be totally missing something here too!