There is no “best”. A common thing to do is to fit a ton of models and then choose the best based on out-of-sample/CV performance, or a mix of top performing models.