Temos o prazer de convidar a todos para mais um seminário do Programa de Pós-Graduação em Estatística da UFRGS (PPGEst-UFRGS).
Abstract: Quantifying uncertainty is essential for reliable predictive modeling and statistical inference. Traditional conformal methods ensure marginally valid prediction intervals but often lack adaptability and fail to achieve conditional coverage. We introduce a model-agnostic approach leveraging regression trees and Random Forests to construct data-adaptive partitions that approximate local coverage guarantees, enhancing scalability and accuracy of existing approaches. Extending this idea to statistical inference, we propose novel techniques for calibrating confidence sets with finite-sample and asymptotic conditional coverage. These methods rely on simulated data for calibration, making them effective even in small-sample regimes or when nuisance parameters are present. Experiments show these techniques outperform existing approaches, offering practical tools for uncertainty quantification in complex settings.