Variable Bandwidth Kernel Estimators of the Spectral Density (Q4939808)

The authors consider a kernel spectral density estimator \[ \hat f(\lambda,h)=\sum_{i=1}^n K((\lambda_i-\lambda)/h)I(\lambda_i), \] where \(I\) is the periodogram, \(\lambda_k=k/T\), \(K\) is a kernel such that \(K\in\text{Lip}[-1,1]\) and for \(\nu_j=\int_{-1}^1 K(u)u^jdu\) hold \(\nu_0=1\), \(\nu_1=0\), \(\nu_2\not=0\). It is proved that under some assumptions the MSE of \(\hat f(\lambda)\) is \(\text{MSE}(f(\lambda,h))= E[(\hat f(\lambda,h)-f(\lambda))/f(\lambda)]^2=\) \[ =(nh)^{-1}(\mu_2+o(1))+(h^2/4)[\nu_2^2f''(\lambda)^2/f(\lambda)^2+o(1)], \] where \(\mu_2=\int K^2(u)du\). The locally optimal bandwidth obtained by minimizing the principal part of MSE is \[ h^*_\lambda=[\mu_2 f(\lambda)^2/4\nu_2^2f''(\lambda)^2]^{1/5} n^{-1/5}. \] The authors propose an adaptive algorithm for local bandwidth estimation at which the globally optimal bandwidth is previously estimated by quasi-likelihood maximization. It is proved that for this estimator \(\hat h_\lambda\) \[ \lim_{n\to\infty}\left(\int_0^1 \text{MSE}(f(\lambda,\hat h_\lambda))d\lambda\right)\left(\int_0^1 \text{MSE}(f(\lambda,h^*_\lambda))d\lambda\right)^{-1} =1. \] Simulation results and application to Spanish money multiplier analysis are discussed.