Numerical meteorological models are used to assist in the prediction of weather. Each run of a numerical model produces forecasts of meteorological variables which are used as preliminary predictions of the future values of these variables. These initial predictions are referred to as first-guess values. Estimation of the mean-square first-guess error is required in the optimal interpolation process in the numerical prediction of atmospheric variables. Several predictors for the mean-square error of the first-guess wind speeds are studied. The results suggest that prediction using observed covariates tend to be better than those using first-guess covariates. However, observed covariates are not always available. Predictions using first-guess covariates are better at the 250 mb level than the 850 or 500 mb levels. Of those first-guess covariates studied, first-guess wind speed appears to be the best. Gaussian model with log-linear scale parameter, Nonparametric models, Prediction of mean square errors, First-guess errors in meteorological models, Generalized linear regression.