Nonparametric regression with warped wavelets and strong mixing processes (Q825064): Difference between revisions

The paper studies a nonparametric regression estimator built on (warped) wavelet functions, in a model of the form \(Y_i=f(X_i)+\varepsilon_i\), \(1\leq i\leq n\), where \(X_i\) is independent of \(\varepsilon_i\), the \(\varepsilon_i\) are homoskedastic, and either \((X_i)\) or \((\varepsilon_i)\) is \(\alpha-\)mixing with the other process being i.i.d. The exact setup is presented in Assumptions 1 and 2 in Section 3, after a description of wavelets, warped wavelets, and the important notion of Muckenhoupt weight in Sections~2.1 and~2.2. The estimator of wavelet coefficients is given in Formula~(8). The covariate values do not have to lie on a regular grid of the covariate space. Results in Section~4, on the risk and large deviations of the estimator, mainly focus on the case when the distribution function of the covariate (which intervenes in the warping scheme) is known, and then on the case when this function is estimated using a set of independent data and the resulting empirical distribution is plugged in Formula~(8). The simulation study in Section~5 assesses the performance of the estimator in a model where either \(X_i\) or \(\varepsilon_i\) follows an autoregressive process of order 1, and using regression functions having various degrees of regularity. The estimators are showcased on a set of real cryptocurrency data in Section~6.