Supplementary Material for Conformalized Quantile Regression

Yaniv Romano, Evan Patterson, Emmanuel J Candès

Abstract:Yi∼ Pois (sin2 (Xi)+ 0.1)+ 0.03 Xi ϵ1, i+ 25 1 {Ui< 0.01} ϵ2, i,(1) where Pois (λ) is the Poisson distribution with mean λ, both ϵ1, i and ϵ2, i are iid standard Gaussian noise, and the Ui are uniform on the interval [0, 1]. We generate a test set of 5000 samples in the same way. The last term in equation (1) creates few but large outliers. This is illustrated in Figure 1, which plots the synthetic data across its full range.In the synthetic experiment of Section 4, we construct a 90% prediction interval for the test data using split conformal prediction. Specifically, we split the training data into two subsets, train a random forest regressor on the first set, and calibrate the intervals on the second set. We do the same for locally adaptive split conformal prediction. The scale estimator is another random forest. The experiment is insensitive to the value of the hyper-parameter γ; we set it to zero. Finally, we instantiate our method, conformal quantile regression, with quantile random forests [1] as the underlying quantile regression algorithm.