ADDITIVE NONPARAMETRIC REGRESSION WITH AUTOCORRELATED ERRORS

A Bayesian approach is presented for nonparametric estimation of an ad ditive regression model with autocorrelated errors. Each of the potent ially non-linear components is modelled as a regression spline using m any knots, while the errors are modelled by a high order stationary au toregressive process parameterized in terms of its autocorrelations. T he distribution of significant knots and partial autocorrelations is a ccounted for using subset selection. Our approach also allows the sele ction of a suitable transformation of the dependent variable. All aspe cts of the model are estimated simultaneously by using the Markov chai n Monte Carlo method. It is shown empirically that the approach propos ed works well on several simulated and real examples.