Perpetual GBM model

[dean-sh]

Description

PerpetualBooster is a new gradient boosting machine (GBM) algorithm which doesn't need hyperparameter optimization unlike other GBM algorithms.
It includes a budget parameter which can be tweaked to optimise the search.

From their results, it seems to be faster for training/inference against a tuned LightGBM with similar performance, and mostly better than AutoGluon across 10 datasets.

https://github.com/perpetual-ml/perpetual

Use case

I'd be keen to test this as a drop-in replacement to any other GBM model and integrate it to nixtla MLforecast natively.

[jmoralez]

Hey. If it implements fit and predict you can just provide it directly to the models argument.

[dean-sh]

Would it support quantiles (confidence intervals)?

[jmoralez]

Yes, with conformal prediction.