Custom loss function?

[lejarx]

I’d like to enforce non negative forecasts (poisson, negative binomial), do you have examples on how we can do this with chronos-2?

[abdulfatir]

This is not a bug, please use discussions for such questions.

[abdulfatir]

@lejarx Chronos-2 generates quantile forecasts, so you cannot enforce specific distributions like Poisson. That said, if you're interested in non-negative integral forecasts, you can clip the predictions to zero and round the floating point values.

[lejarx]

@lejarx Chronos-2 generates quantile forecasts, so you cannot enforce specific distributions like Poisson. That said, if you're interested in non-negative integral forecasts, you can clip the predictions to zero and round the floating point values.

Thanks @abdulfatir , on slightly similar note, do you have studies on how Chronos-2 performs on intermittent time series data?

[abdulfatir]

We evaluated on general purpose benchmarks which comprises tasks from multiple domains, possibly also intermittent series. However, we did not conduct case studies specifically on this kind of data. If you happen to do this, don't forget to share it here with the community. I would be curious to know your findings! :)