| dc.description.abstract | This study evaluated and compared jackknife and jackknife+ conformal prediction intervals for three machine learning regression models, namely random forest, gradient boosting, and support vector regression. Two settings were used: a simulation that varied the predictor-to-sample-size ratio from 0.2 to 2.0 across ten scenarios with 100 replications each, using synthetic data drawn from a semiparametric Gaussian copula, and an application to Susenas 2023 household expenditure for Bogor City. Six method-model combinations were assessed at 90% nominal coverage using coverage rate, interval width, Winkler score, computation time, and peak memory. In the simulation, median coverage reached 91-92%, but the first quartile fell to 86-88%, and interval width was governed mainly by the base model, not the interval method. On 815 households, coverage ranged from 90.80% to 92.64%, with random forest and jackknife+ closest to nominal. Coverage failures concentrated in the right tail, so these dataset-specific findings require cautious generalization. | |
