Forward variable selection for random forest models.

Area of Science:

• Statistics
• Machine Learning
• Environmental Science

Background:

• Random forests are effective for high-dimensional data but lack interpretability.
• Interpretable predictive models are crucial for understanding complex relationships.

Purpose of the Study:

• To develop a forward variable selection method for interpretable predictive modeling.
• To minimize the continuous ranked probability score (CRPS) for optimal variable selection.
• To provide a statistically rigorous method for selecting relevant covariates.

Main Methods:

• A stepwise forward selection procedure minimizing CRPS.
• A stopping criterion based on CRPS risk difference estimation.
• Mathematical proof of optimality in a population sense.
• Simulation studies comparing performance with existing methods.

Main Results:

• The proposed method achieves a lower false positive rate compared to existing techniques.
• Demonstrated effectiveness in statistical post-processing of temperature forecasts.
• Selected approximately 10% of covariates while maintaining predictive power.

Conclusions:

• The developed method offers an interpretable alternative for high-dimensional prediction.
• It provides a robust approach for variable selection in statistical modeling.
• The method is applicable to real-world forecasting problems, enhancing model transparency.