Clarification on the Definition and Source of Prediction Intervals

[Claire-bx]

I would like to confirm my understanding of how the prediction interval (e.g., pred 0.9) is defined and implemented in the model.

When setting pred 0.9, does it indicate that the true value is expected to fall within this interval with a probability of 90%?
Specifically, are the lower and upper bounds corresponding to the 5% and 95% quantiles of the predictive distribution?

Regarding the source of uncertainty in interval prediction:

By model uncertainty, I refer to cases where, even with the same input and fixed model parameters, the model output may vary due to stochastic components (e.g., dropout, sampling, or ensemble randomness).
By data uncertainty, I refer to the uncertainty learned from the data itself, where the model structure or parameterization is designed to capture aleatoric uncertainty and produce probabilistic outputs.

Could you please clarify which type(s) of uncertainty are considered in our implementation when generating the prediction intervals?

Your insights would be greatly appreciated.

[lostella]

@Claire-bx taking the 0.9 prediction gives you the 0.9 quantile: the true value is expected to fall below this level 90% of the time.

This is an estimate of the distribution width learned from the data: Chronos models obtain this via sampling, Chronos-Bolt models instead were trained with a quantile regression objective, so they output the quantile estimates directly.

[lostella]

You can increase the number of sample paths drawn (default is 20) and that will give you more stable quantile estimates between runs.

Or you can use Chronos-Bolt models, that directly emit quantile estimates with no sampling involved: these will have no variation in the predicted quantiles.

[Claire-bx]

Does Chronos-Bolt support zero-shot and fine-tuning for the following tasks:
Univariate forecasting
Multivariate forecasting
Covariate-informed forecasting

I noticed that there are currently at least three models: Chronos, Chronos-Bolt, and Chronos-2. Could you please let me know if there is any comparison table or documentation summarizing their differences？
It would be very helpful for understanding which model to use in different scenarios

[lostella]

@Claire-bx see this notebook for such a table of features: https://github.com/amazon-science/chronos-forecasting/blob/main/notebooks/chronos-2-quickstart.ipynb

In short: chronos/chronos-bolt only support univariate forecasting, including fine-tuning. Chronos-2 also support additional features and variates, including fine-tuning: the notebook linked above contains an example of this (experimental code, things may change).

[Claire-bx]

I have a question regarding fine-tuning the Chronos-2 model via the API. Does the fine-tuning process run on a local GPU or on a cloud server by default? If it can be run locally, could you recommend an appropriate GPU configuration for efficient fine-tuning, and roughly what proportion of the model’s parameters would be updated during this process?
Thank you very much for your time and support.

[lostella]

@Claire-bx the fine-tuning will run on the machine where you run the code. A single GPU instance for fine-tuning. All model parameters will be fine-tuned.