This paper considers the variance estimation in nonlinear autoregressive time series models. The estimator, based on the residuals, is shown to be consistent (with View the MathML source rate) for the error variance. The asymptotic distribution of the estimator is obtained to be normal.