Using techniques of nonparametric regression, we develop a nonparametric approach in the context of kernel estimation to realize short-term forecastings of time series. This procedure is applied to an OZONE (O\sb3) daily maximum series, whose values were filtered according to the Tukey (biweight) kernel function: K(x)=1516(1−x\sp2)\sp2I\sb(−1,1)(x). Some parametric approaches such as multivariate regression and autoregressive integrated moving average (ARIMA) models (under assumptions of normality, stationarity, invertibility, etc.) are also shown and compared with the nonparametric approach, which is an attractive alternative. Moreover a procedure for the estimation of missing observations in time series, and a method to improve the optimal "bandwidth" selection for the nonparametric regression kernel estimator are proposed.