Hi there, I want to conformalize quantile regression (Multi-quantile NHITS/LSTM, Multi-quantile catboostRegressor). Does the current implementation of CP on Nixtla deals with such cases?

I don't think we can do this at the moment

@Christian Ngnie I am working on the support of conformal prediction in Neuralforecast. Please see the comment in https://github.com/Nixtla/neuralforecast/pull/1171/files#r1788935653 I am relatively new to conformal predictions, would like to understand better is that what you are looking for and is it appropriate to conformalize quantiles outputs in this manner. cc: @José Morales @Valeriy

As far as I understand it's okay to conformalize quantile outputs; similarly as it's okay to conformalize over models trained using RMSE. The conformal intervals just provide a (quantile) interval over the point forecast based on a validation set.

@Olivier Sounds good to me. I will provide an optional argument such that we can disable this by default, but users can choose to do so if needed

@Jing Qiang Goh awesome. It's totally ok and appropriate to conformalize quantiles in that way.

Usually people do two things 1) create prediction intervals from point predictions 2) if the model predicts quantiles can conformalize quantiles. A good example is NeuralProphet it has both options.

@Olivier It is interesting to learn that Darts library is introducing conformal predictions https://github.com/unit8co/darts/pull/2552 I have not checked the details yet but like to share this, in case there are lessons or ideas we can learn from too @Valeriy FYI