Non-crossing quantile regression via doubly penalized kernel machine (Q2430224): Difference between revisions

A nonparametric location-scale regression model is considered of the form \(y_i=\mu(x_i)+\sigma(x_i)\varepsilon_i\), where \(x_i\) is a covariate vector, \(y_i\) is a response, \(\mu\) and \(\sigma\) are unknown smooth functions, and \(\varepsilon_i\) are i.i.d. doubly exponential errors. The authors propose conditional quantile estimates for \(y_i\) given \(x_i\), based on nonparametric estimates for \(\mu\) and \(\sigma\). They are derived by minimization of a penalized log-likelihood corresponding to this model in the class of functions generated by a support vector machine. Performance of the estimates is investigated on real life and simulated data.