Prediction intervals with random forests.

Area of Science:

• Machine Learning
• Statistical Modeling

Background:

• Parametric prediction intervals rely heavily on assumptions about the relationship between variables, limiting their real-world applicability.
• Random forests offer a non-parametric alternative, but their effectiveness in constructing accurate prediction intervals requires further optimization.

Purpose of the Study:

• To develop and evaluate novel methods for constructing more robust and accurate prediction intervals using random forests.
• To investigate the impact of different forest-building strategies and prediction interval construction techniques on performance.

Main Methods:

• Explored four forest-building methods, including Classification and Regression Trees (CART) with varied splitting criteria and transformation forests.
• Developed and assessed five flexible methods for prediction interval construction.
• Implemented a calibration procedure using out-of-bag data to ensure reliable confidence levels.
• Compared 20 distinct method variations against five established alternatives via simulations and real-world data.

Main Results:

• The proposed random forest-based methods demonstrated highly competitive performance.
• New methods consistently outperformed commonly used prediction interval techniques.
• Improvements were observed in both simulated environments and analyses of real datasets.

Conclusions:

• The investigated random forest methods provide a powerful and flexible alternative for building accurate prediction intervals.
• These enhanced methods overcome limitations of traditional parametric approaches, offering improved reliability and performance.