There are three primary ways to compute global feature importance values:

• Gain: This classic approach introduced by Leo Breiman in 1984 uses the total reduction of loss or impurity contributed by all splits for a given feature. The motivation is largely heuristic, but it is a commonly used method to select features.
• Split count: This is an alternative approach that counts how often a feature is used to make a split decision, based on the selection of features for this purpose based on the resultant information gain.
• Permutation: This approach randomly permutes the feature values in a test set and measures how much the model's error changes, assuming that an important feature should create a large increase in the prediction error. Different permutation choices lead to alternative implementations of this basic approach.

Individualized feature importance values that compute the relevance of features for a single prediction are less common because available model-agnostic explanation methods are much slower than tree-specific methods.

All gradient boosting implementations provide feature-importance scores after training as a model attribute. The XGBoost library provides five versions, as shown in the following list:

• total_gain and gain as its average per split
• total_cover as the number of samples per split when a feature was used
• weight as the split count from preceding values

These values are available using the trained model's .get_score() method with the corresponding importance_type parameter. For the best performing XGBoost model, the results are as follows (the total measures have a correlation of 0.8, as do cover and total_cover):

While the indicators for different months and years dominate, the most recent 1 month return is the second-most important feature from a total_gain perspective, and is used frequently according to the weight measure, but produces low average gains as it is applied to relatively few instances on average (see the notebook for implementation details).