When trying to fit a population into a model (e.g. classification or regression) outliers tend to decrease the performance of the model. Thus, in most of the research outliers are detected and removed from the population to overcome this decrease. This common judgment about outliers may increase the modeling accuracy of inliers; however knowledge about outliers is discarded by this way. If outliers occur from natural phenomena (i.e. when outliers are not noise) their identification may accomplish even more important knowledge than the inliers. In other words, outliers should also be considered as a member of the population while trying to fit the members of a population into a model. This paper attacks to this idea in spatio-temporal domain. Firstly, spatio-temporal outliers are identified, secondly this information is used explicitly while fitting the population into a regression model via support vector regression to investigate whether there will be increase in the prediction accuracy of the model or not. Experiments are carried out through a transactional data set on forest fires. It is shown that regression performance increases by considering spatio-temporal outliers.