I want to find out which factors are used without using the predictor?

General
1

I ran a screenbacktest from the validation - models with this formula:"""
(AIFactorValidation("aa+aa+63D+rank+d+us micro", "lightgbm d310724 - Copy")

Is it possible to find which stock factors were used and how they are weighted in this formula for the backtest similar to how Coefficient shows it for linear models in 'importance'?

image
image1105×1248 105 KB

2

Feature Importance for LigthGBM is available. You can train a predictor using the same model and you will have access to those stats.

Note that your validation data is (probably) made up of multiple trained models, one for each fold, with each version having it's own set of feature importances. But a single predictor using the entire dataset should be representative of the importances.

Make sense?

1 Like
3

FYI for anyone.

LightGBM has 2 types of feature importances accessible in the full program. Not sure what P123 uses as a default or if that can be controlled with the hyperparameters using P123's implemenation.

You may want to use gain. Split is the default for LighGBM look like.

LightGBM: documentation:

  • importance_type (str , optional ( default='split' )) – The type of feature importance to be filled into feature_importances_. If ‘split’, result contains numbers of times the feature is used in a model. If ‘gain’, result contains total gains of splits which use the feature.

I think this is an accurate statement: Gain importance is likely to highlight features that make significant changes to stock predictions when they are used.

Jim