In this paper, we are interested in the statistical modeling and forecasting of the daily maximum ozone concentration in three monitoring stations from Tunisia. A large number of explicative variables has been considered in our study. We have focused our attention on the problem of variable selection in order to improve the forecasting performance. To achieve our goal, we have used Support Vector Regression (SVR) and Random Forests (RF). The main novelties of this paper are: the variety and originality of the approaches for variable selection in regression, and the audaciousness to deal with a sticky situation characterized by a relatively big pannier of explicative variables compared to the number of observations. The experimental results demonstrate that Random Forests outperform Support Vector Regression in variable ranking and selection. Finally, it was shown that the forecasting accuracy is at least preserved, for the three stations, when using only the selected variables.