Can Chronos 2 be used for probabilistic scenario forecasts (outputting trajectories)?

[birdMarcus]

I am interested in doing probabilistic forecasts where I would need different realistic scenarios/trajectories that follow the marginal distributions (for instance for stochastic optimization or simulations).

Is this possible with Chronos2 out of the box?
With Chronos 1 it reads like there is an intermediate step where time series are sampled and then the quantiles are calculated from that (this is how I understand Figure 1 in the paper for Chronos).
For Chronos 2 it reads like it is more direct or is there still this intermediate step and is it possible to access those intermediate samples?

I could maybe use Chronos 1 for my problem, however, I would likely need the covariate capabilities of Chronos 2.

[abdulfatir]

@birdMarcus Chronos-2 directly generates quantiles while Chronos generates forecast samples. There's no out of the box way to draw samples from Chronos-2. That said, maybe the following could work

• Generate the full quantile-grid from Chronos-2 using the .predict method.
• Construct an empirical inverse CDF using the quantiles.
• For each future time-step, draw u ~ np.random.rand().
• Interpolate values for u from the inverse CDF.

You can do this multiple times to draw multiple trajectories.

[birdMarcus]

Dear @abdulfatir thanks for the reply.

To maintain the autoregressive dependencies between the steps in the horizon I would then need copulas or something like that though, right (as in this paper for wind power https://onlinelibrary.wiley.com/doi/epdf/10.1002/we.284)?

But I will check if that works for my use case. Thanks!

PS: great work with Chronos 2 in any case - am excited to have a time series problem at hand to finally get hands on with time series foundation models. Excited to see where those efforts go!

[abdulfatir]

Yes, copulas may be one way to go, although I am not super sure how to incorporate them here. My advice would be to give the inverse CDF idea a go. It may not be statistically sound but could give you good enough trajectories.

Thank you for your wishes!

[abdulfatir]

Of course, if you find a clever idea to sample better trajectories from Chronos-2, don't forget to share it here with the community.