In credit risk modeling, method-of-moment approaches are popular for estimating latent asset return correlations within and between rating buckets. However, the autocorrelation often present in time series of default rates leads to estimations that are systematically too low. We propose a new estimator that adjusts to the problems of this autocorrelation and the shortness of the time series, thus eliminating a significant portion of the bias observed with classical estimators. The adjustment is based on convergence and approximation results for general autocorrelated time series, and it is easily implementable and nonparametric.