Datasets:
Training Data for Chronos Bolt
Hi @shchuro and autogluon team,
Thank you for the fantastic work with the hf datasets, chronos and chronos-bolt!
We are working on simulating supply chain risks, and tail risk is extremely relevant in our application. Hence, we'd like to train bolt models with custom quantiles (below/above p10/p90).
Are there any details on what data was used to train them? Specifically, I am wondering about the TSMixup and KernelSynth data. Any pointer helps!
We'd be willing to contribute back to autogluon-timeseries with base model training if this is a feature you'd be looking to support.
Hi @anchovy , we have recently released Chronos-2 that provides support for extreme quantiles (0.99, 0.95, 0.90) out of the box (example notebook). You can also request any quantiles below p99 and they will be automatically interpolated. Will this be sufficient for your use case?
Currently we are not planning to release the training data for Chronos-Bolt.