Open Beerstabr opened 1 year ago
Hi @Beerstabr, this sounds like a great idea to me!
I like your proposed approach leveraging LinearRegressionModel
. If for some reason it turns out to be not a great fit, we can also consider making new classes (eg. BayesianRidgeModel
) inheriting from RegressionModel
. But I think your proposed approach should be tried first.
I'd be very keen on seeing a PR along those lines!
I really like the option of setting the likelihood to 'quantile' in the LinearRegressionModel.
It's great because it allows you to create probabilistic forecasts using Sklearn's QuantileRegressor which also includes regularization and thus makes for a very powerful forecasting tool.
Sklearn offers a wide range of models and I think some of them could be included into Darts in a similar fashion as QuantileRegressor.
Two models that might be of value are BayesianRidge and ARDRegression. They are great at feature selection and offer probabilistic predictions. And they could be included in the LinearRegressionModel and called in a similar way as QuantileRegressor by simply setting the likelihood to 'bayesian' or 'ard-bayesian'.
If this sounds interesting to you, I'd gladly help out and come up with a pull-request.