A Study on Prediction Intervals Produced Using Quantile Regression Forest With and Without Variable Selection

Abstract

Quantile Regression Forest (QRF) is a method that utilizes the random forest algorithm to estimate the conditional distribution of response variables and form quantile prediction intervals. However, when there is a high correlation between covariates, QRF performance may decrease due to the multicollinearity effect, thereby reducing the accuracy of the prediction interval for the target variable. In linear models, multicollinearity must be addressed because it can cause large variances. This study contributes to enhancing the reliability of prediction intervals in correlated data through the integration of adaptive-LASSO with QRF. Specifically, it examines the role of variable selection by the adaptive LASSO method on the performance of the QRF prediction interval in the simulated data, and the best model obtained in the study is then applied to predict the interval in the productivity data of oil palm fresh fruit bunches. The results of the study show that variable selection is proven to produce coverage close to the target prediction interval. In addition, the QRF model with variable selection applied to the productivity data of oil palm fresh fruit bunches produces a good prediction interval.