Since the beginning of the 1990s, total column ozone is being forecasted on an operational basis at several institutes around the world. It usually serves as an input parameter for the UV index forecast. The ozone forecasting procedures described in the literature are ad hoc statistical regression models. They yield sufficiently accurate predictions. So far, no detailed description, including the validity of the underlying assumptions of regression analysis, has been communicated. We studied this subject and conclude that they are violated by the models currently in use. Taking into account the causes for this violation, we come up with a more sophisticated model, which combines a stepwise multilinear regression model with an autoregressive conditional heteroskedasticity error model. This model still has shortcomings, but it yields a reduction of variance with respect to the commonly used ordinary least squares stepwise regression model of the order of 35%. Expressed in terms of R2, the improvement is 6%. The root-mean-square error of the model we propose is 3.2 Dobson units lower than the usually applied regression model. ¿ 2000 American Geophysical Union